Built-in models catalog

Generated from the manifests of every model plugin installed through the mapi.models entry point. Read Built-in models for how to use them.

Model	Version	Inference	Representation	Score modes
causal.did	1.0.0	mcmc	draws	counterfactual, predict
causal.impact	1.0.0	mcmc	draws	forecast, predict
causal.synthetic_control	1.0.0	mcmc	draws	predict
choice.multinomial_logit	1.0.0	bootstrap	draws	counterfactual, evaluate, forecast, predict
dl.cnn_classifier	1.0.0	custom	draws	evaluate, predict
dl.finetune_classifier	1.0.0	custom	draws	evaluate, predict
dl.mlp	1.0.0	custom	draws	evaluate, predict
mmm.bayesian	1.1.0	mcmc	draws	counterfactual, evaluate, forecast, predict
mmm.frequentist	1.0.0	bootstrap	draws	counterfactual, evaluate, forecast, predict
mmm.multivariate	1.0.0	mcmc	draws	counterfactual, evaluate, forecast, predict
mmm.nested	1.0.0	mcmc	draws	counterfactual, evaluate, forecast, predict
sci.areal_bym2	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.circular_regression	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.cmb_field	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.colour_naming	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.fisher_kpp	1.0.0	mcmc	draws	evaluate, forecast, predict
sci.graphical_model	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.heat_plate	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.ibp_features	1.0.0	mcmc	draws	evaluate, predict
sci.lgcp_grid	1.0.0	mcmc	draws	evaluate, forecast, predict
sci.network_latent_space	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.phylo_regression	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.storm_localisation	1.0.0	mcmc	draws	evaluate, predict
sci.strong_lens	1.0.0	mcmc	draws	evaluate, predict, simulate
sci.traction_force	1.0.0	mcmc	draws	evaluate, predict
sci.turing_pattern	1.0.0	mcmc	draws	evaluate, predict, simulate
sklearn.multiclass_classifier	1.0.0	point	point	evaluate, predict
sklearn.tabular_bootstrap	1.0.0	bootstrap	draws	counterfactual, evaluate, forecast, predict
sklearn.tabular_classifier	1.0.0	point	point	counterfactual, evaluate, forecast, predict
sklearn.tabular_conformal	1.0.0	custom	interval	counterfactual, evaluate, forecast, predict
sklearn.tabular_quantiles	1.0.0	custom	quantiles	counterfactual, evaluate, forecast, predict
sklearn.tabular_regressor	1.0.0	point	point	counterfactual, evaluate, forecast, predict
tabular.hierarchical_glm	1.1.0	mcmc	draws	counterfactual, evaluate, predict
tabular.image_logistic	1.1.0	mcmc	draws	evaluate, predict
tabular.regressor	1.0.0	custom	quantiles	counterfactual, evaluate, forecast, predict
ts.ardl	1.0.0	point	quantiles	evaluate, forecast, predict
ts.arima	1.0.0	kalman	quantiles	evaluate, forecast, predict
ts.bsts	1.0.0	mcmc	draws	evaluate, filter, forecast, predict
ts.dynamic_factor	1.0.0	kalman	quantiles	evaluate, forecast, predict
ts.ecm	1.0.0	point	quantiles	evaluate, forecast, predict
ts.ets	1.0.0	custom	quantiles	evaluate, forecast, predict
ts.gaussian_hmm	1.0.0	mcmc	draws	evaluate, filter, forecast, predict
ts.hsgp_trend	1.0.0	mcmc	draws	evaluate, forecast, predict
ts.markov_switching	1.0.0	custom	quantiles	evaluate, forecast, predict
ts.sarimax	1.0.0	kalman	quantiles	evaluate, forecast, predict
ts.theta	1.0.0	custom	quantiles	evaluate, forecast, predict
ts.unobserved_components	1.0.0	kalman	quantiles	evaluate, forecast, predict
ts.var	1.0.0	point	quantiles	evaluate, forecast, predict
ts.varmax	1.0.0	kalman	quantiles	evaluate, forecast, predict
ts.vecm	1.0.0	point	quantiles	evaluate, forecast, predict

causal.did

Bayesian difference-in-differences and event study on a panel of units over time: unit and period effects plus one effect per period relative to the adoption of the treatment, compared with units that never got it.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	counterfactual, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	scaling
LOO	valid
Default plots	causal.event_study, causal.impact, causal.pre_fit, core.convergence, core.posterior_forest
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(unit, time)	of y	Expected value or predictive draw of y; with the intervention treatment = 0, the outcome without the treatment

Settings (model.settings)

Field	Type	Default	Level	Description
design	event_study or did	event_study	–	event_study: one effect per period relative to the adoption; did: one effect for all treated cells (x-ui-level: basic)
leads	integer	4	–	Periods before the adoption with an effect (≥ 0, ≤ 52)
lags	integer	8	–	Periods after the adoption with an effect (≥ 0, ≤ 104)
bin_endpoints	boolean	true	–	The first lead and the last lag also take the periods further out
time_effects	iid or random_walk	iid	–	iid: independent period effects; random_walk: smooth period effects
covariates	list of string	–	–	Control variables over (unit, time) (max items 16)
likelihood	normal or student_t	normal	–
effect_sigma	number	1.0	–	Prior sd of the event effects (> 0, ≤ 10)
noise_sigma	number	0.5	–	Half-normal scale of sigma (> 0, ≤ 10)

Input data contract (schema template)

{
  "params": {
    "covariates": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "unit": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "time": {
        "kind": "time",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "unit",
          "time"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "treatment": {
        "dims": [
          "unit",
          "time"
        ],
        "dtype": "float64",
        "role": "covariate",
        "bounds": [
          0.0,
          1.0
        ]
      },
      "{*covariates as x}": {
        "dims": [
          "unit",
          "time"
        ],
        "dtype": "float64",
        "role": "covariate"
      }
    }
  }
}

causal.impact

CausalImpact / interrupted time series: a Bayesian structural time series fitted on the periods before an intervention, with control series as regressors, predicts the series without the intervention; the effect is actual minus that counterfactual.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	scaling
LOO	invalid: the Kalman likelihood is a sum over the whole series; steps depend on each other, so pointwise leave-one-out is not defined
Default plots	causal.impact, causal.pre_fit, core.decomposition, core.convergence, core.posterior_forest
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(time)	of y	What y would be without the intervention
effect	(time)	of y	The effect of the intervention: y minus what y would be without it

Settings (model.settings)

Field	Type	Default	Level	Description
intervention_start	string	required	–	First period of the intervention: an ISO date (YYYY-MM-DD) or a period number
intervention_end	string (optional)	null	–	Last period of the intervention (inclusive); empty: until the end of the data
trend	level or local_linear or smooth	local_linear	basic	level: a level that wanders; local_linear: a level and a slope that both wander; smooth: a slope that wanders slowly (a smooth trend)
seasonality	list of SeasonSettings	–	basic	Seasonal patterns (at most three) (max items 3)
level_sigma	number	0.1	–	Half-normal scale of level changes (> 0, ≤ 5)
slope_sigma	number	0.02	–	Half-normal scale of slope changes (> 0, ≤ 5)
obs_sigma	number	0.5	–	Half-normal scale of the noise (> 0, ≤ 10)
initial_sd	number	1.0	–	Prior sd of the initial states and seasonal effects (> 0, ≤ 10)
regression_prior	normal or horseshoe	normal	–	normal: every driver has an effect; horseshoe: most drivers have none
regression_sigma	number	1.0	–	Prior scale of the driver effects (standardised) (> 0, ≤ 10)
controls	list of string	–	basic	Control series that the intervention did not affect (other regions, competitors, search interest); their values are needed for every period (max items 32)

Nested settings types

SeasonSettings

One seasonal pattern.

Field	Type	Default	Description
period	number	required	Length of the cycle in periods of the time dim (12 for monthly data with a yearly cycle, 52.18 for weekly data, 7 for daily data) (> 1, ≤ 1000)
harmonics	integer (optional)	null	Number of sine/cosine pairs. Empty: one effect per season (needs a whole-number period of at most 24)
stochastic	boolean	false	Let the seasonal pattern change over time
sigma	number	0.05	Half-normal scale of the seasonal changes (standardised) (> 0, ≤ 5)

Input data contract (schema template)

{
  "params": {
    "controls": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "time": {
        "kind": "time",
        "policy": "extend"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "time"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "{*controls as x}": {
        "dims": [
          "time"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  }
}

causal.synthetic_control

Bayesian synthetic control for geo-lift tests: each treated market is a weighted mix of untreated markets fitted before the test; the mix after the start is the counterfactual, and the difference to the actual sales is the effect.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	scaling
LOO	invalid: the likelihood covers only the pre-period of the treated units, so pointwise leave-one-out does not describe the scored data
Default plots	causal.impact, causal.pre_fit, causal.donor_weights, core.convergence
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(unit, time)	of y	The synthetic control: what y of a treated unit would be without the intervention (NaN for donor units)

Settings (model.settings)

Field	Type	Default	Level	Description
intervention_start	string	required	–	First period of the intervention: an ISO date (YYYY-MM-DD) or a period number
intervention_end	string (optional)	null	–	Last period of the intervention (inclusive); empty: until the end of the data
treated	list of string	required	–	Labels of the units (markets) that got the intervention; a text 'a, b' is split at the commas (x-ui-level: basic) (min items 1, max items 20)
donors	list of string (optional)	null	–	Units to build the synthetic copy from; default: every unit that is not treated
weights	simplex or horseshoe	simplex	–	simplex: positive weights that sum to 1 (classic synthetic control); horseshoe: sparse weights of any sign (can extrapolate)
concentration	number	1.0	–	Dirichlet concentration of simplex weights (< 1: sparse) (> 0, ≤ 100)
intercept	boolean	true	–	Allow a level shift between a unit and its synthetic copy
normalize	mean or none	mean	–	mean: divide each unit by its pre-period mean before the fit
noise_sigma	number	0.5	–	Half-normal scale of sigma (> 0, ≤ 10)
likelihood	normal or student_t	normal	–

Input data contract (schema template)

{
  "template": {
    "dims": {
      "unit": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "time": {
        "kind": "time",
        "policy": "extend"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "unit",
          "time"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  }
}

choice.multinomial_logit

Which product customers choose on each shopping trip: a conditional (multinomial) logit over alternative-specific covariates such as price and promotion, with household-specific product preferences. Uncertainty comes from the Bayesian bootstrap over households; new households draw their preferences from the fitted spread.

Version	1.0.0
Inference	bootstrap
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: bootstrap draws have no pointwise log-likelihood
Default plots	core.data_overview, core.posterior_forest
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(hh, trip, product)	probability	Choice probability of each product per household and trip (predictive: one sampled choice, one-hot)

Settings (model.settings)

Field	Type	Default	Level	Description
covariates	list of string (optional)	null	advanced	Alternative-specific variables along product (default: every numeric variable with the product dim)
random_effects	list of hh	["hh"]	advanced	Household-specific product intercepts ([hh]) or none ([]) (max items 1)
effect_sd	number	1.0	advanced	Prior sd of the household effects (smaller pools households more) (> 0, ≤ 10)
coef_sd	number	10.0	advanced	Prior sd of the product intercepts and coefficients (weak) (> 0, ≤ 1000)
reference	string (optional)	null	advanced	Product whose intercept is fixed at 0 (default: the first label)
max_iter	integer	500	advanced	≥ 10, ≤ 10000

Input data contract (schema template)

{
  "params": {
    "covariates": {
      "type": "string_list",
      "path": "covariates",
      "default": []
    }
  },
  "template": {
    "dims": {
      "hh": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "hyperprior"
      },
      "trip": {
        "kind": "index",
        "policy": "extend"
      },
      "product": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "hh",
          "trip"
        ],
        "dtype": "str",
        "role": "target",
        "missing": "allow"
      },
      "{*covariates as v}": {
        "dims": [
          "hh",
          "trip",
          "product"
        ],
        "dtype": "float64",
        "role": "covariate",
        "missing": "allow"
      }
    }
  }
}

dl.cnn_classifier

A small residual convolutional network trained on labelled images. It streams large image collections in batches, logs training and validation curves, and gives probability draws by a last-layer Laplace approximation or MC dropout.

Version	1.0.0
Inference	custom
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: a network trained by gradient descent has no pointwise log likelihood
Default plots	ml.training_curves, ml.confusion_epochs, ml.misclassified, core.roc_pr, core.confusion
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
y_hat	(class)	1	Probability of each class per image (sums to 1 over class)

Settings (model.settings)

Field	Type	Default	Level	Description
pixels	string	pixels	–	The image variable (image, y, x, channel)
target	string	label	–	The class label of each image
image_dim	string	image	–	The dim that lists the images
image_size	list of any (optional)	null	–	Height and width the images are resized to (default: their size)
mode	auto or gray or rgb	auto	–	Channels of the input
normalize	unit or standardize	unit	–	unit: pixels / 255; standardize: per-channel training mean and sd
augment	Augment	–	–	Training augmentation
train	TrainConfig	–	–
uncertainty	UncertaintyConfig	–	–
class_order	list of string (optional)	null	–	Classes in order (default: sorted)
widths	list of integer	[16, 32, 64]	–	Channels of each stage (min items 1, max items 8)
blocks	list of integer	[1, 1, 1]	–	Residual blocks per stage (min items 1, max items 8)
groups	integer	8	–	GroupNorm groups (at most) (≥ 1, ≤ 64)
stem_stride	integer	1	–	Stride of the first convolution (≥ 1, ≤ 2)
dropout	number	0.1	–	Dropout before the head (≥ 0, < 1)

Nested settings types

Augment

Random per-image transforms a loader applies to training batches. Flips and 90-degree turns (only half turns for non-square images) keep the pixels; crop_pad pads by n pixels (edge values) and crops back to the size at a random offset; brightness/contrast add or scale by a random amount up to the value (in units of the 0..255 range); noise adds Gaussian noise with that standard deviation.

Field	Type	Default	Description
flip_x	boolean	false	Mirror left-right with probability 0.5
flip_y	boolean	false	Mirror top-bottom with probability 0.5
rotate90	boolean	false	Turn by a random multiple of 90 degrees
crop_pad	integer	0	Pad by n pixels, then crop at random (≥ 0, ≤ 64)
brightness	number	0	Largest brightness change (≥ 0, ≤ 1)
contrast	number	0	Largest relative contrast change (≥ 0, ≤ 1)
noise	number	0	Gaussian noise sd (0..1 units) (≥ 0, ≤ 0.5)

EarlyStopping

Stop when the monitored validation metric has not improved for patience evaluations.

Field	Type	Default	Description
monitor	val_loss or val_accuracy or val_rmse	val_loss	The validation metric to watch
patience	integer	5	Evaluations without improvement (≥ 1, ≤ 1000)
min_delta	number	0.0	The smallest change that counts as better (≥ 0)
restore_best	boolean	true	Keep the weights of the best evaluation

TrainConfig

How a network is trained (optimizer, schedule, validation, logging).

Field	Type	Default	Description
epochs	integer	20	Passes over the training rows (≥ 1, ≤ 10000)
batch_size	integer	64	Rows per gradient step (≥ 1, ≤ 8192)
optimizer	adamw or adam or sgd	adamw
learning_rate	number	0.001	Peak learning rate (> 0, ≤ 10)
weight_decay	number	0.0001	Decoupled weight decay (adamw) (≥ 0, ≤ 1)
momentum	number	0.9	Momentum (sgd) (≥ 0, < 1)
schedule	constant or cosine or warmup_cosine	warmup_cosine
warmup_epochs	number	1.0	Linear warm-up (warmup_cosine) (≥ 0, ≤ 100)
grad_clip	number (optional)	1.0	Clip the global gradient norm
label_smoothing	number	0.0	Classification only (≥ 0, < 1)
val_fraction	number	0.1	Stratified share of the training rows for validation (≥ 0, < 0.9)
val_split_value	string (optional)	val	Rows whose split coord equals this value are the validation set
early_stopping	EarlyStopping (optional)	–
log_every_steps	integer (optional)	null	Steps between logged training metrics (default: <= 1000 points)
eval_every_epochs	integer	1	Epochs between validation passes (≥ 1)
media_every_epochs	integer (optional)	null	Epochs between logged images (default: at most 20 times)
max_steps	integer (optional)	null	Stop after this many gradient steps

UncertaintyConfig

How a trained network gives draws (chain=1, draw).

Field	Type	Default	Description
method	laplace or mc_dropout or none	laplace	laplace: last-layer Laplace; mc_dropout: dropout at score; none: one draw
draws	integer (optional)	null	Draws per prediction (default: fit draws, else 100)
hessian	auto or full or diag	auto	auto: full when (features + 1) x outputs <= 2048
max_rows	integer	20000	Training rows used for the Laplace Hessian (≥ 10)
prior_precision	number (optional)	null	None: maximise the Laplace evidence on a grid 1e-4..1e4

Input data contract (schema template)

{
  "params": {
    "pixels": {
      "type": "string",
      "default": "pixels"
    },
    "target": {
      "type": "string",
      "default": "label"
    },
    "image_dim": {
      "type": "string",
      "default": "image"
    }
  },
  "template": {
    "dims": {
      "{image_dim}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "pooled"
      },
      "y": {
        "kind": "integer",
        "policy": "fixed"
      },
      "x": {
        "kind": "integer",
        "policy": "fixed"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "{pixels}": {
        "dims": [
          "{image_dim}",
          "y",
          "x",
          "channel"
        ],
        "dtype": "uint8",
        "role": "covariate"
      },
      "{target}": {
        "dims": [
          "{image_dim}"
        ],
        "dtype": "str",
        "role": "target",
        "missing": "allow"
      }
    }
  }
}

dl.finetune_classifier

A pretrained residual network (weights from a dataset version) with a new classifier head, trained on labelled images. The backbone can stay frozen. No download: the weights are a pinned side table.

Version	1.0.0
Inference	custom
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: a network trained by gradient descent has no pointwise log likelihood
Default plots	ml.training_curves, ml.confusion_epochs, ml.misclassified, core.roc_pr, core.confusion
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
y_hat	(class)	1	Probability of each class per image (sums to 1 over class)

Settings (model.settings)

Field	Type	Default	Level	Description
pixels	string	pixels	–	The image variable (image, y, x, channel)
target	string	label	–	The class label of each image
image_dim	string	image	–	The dim that lists the images
image_size	list of any (optional)	null	–	Height and width the images are resized to (default: their size)
mode	auto or gray or rgb	auto	–	Channels of the input
normalize	unit or standardize	unit	–	unit: pixels / 255; standardize: per-channel training mean and sd
augment	Augment	–	–	Training augmentation
train	TrainConfig	–	–
uncertainty	UncertaintyConfig	–	–
class_order	list of string (optional)	null	–	Classes in order (default: sorted)
pretrained	string	pretrained	–	The data.side_tables entry with the pretrained weights
freeze	none or backbone	none	–	backbone: train the head only
head_epochs	integer	1	–	Epochs that train the head only before the rest (≥ 0, ≤ 10000)
backbone_lr_scale	number	0.1	–	Learning rate of the backbone relative to the head (> 0, ≤ 1)
dropout	number (optional)	null	–	Dropout before the head (default: the weights' value)

Nested settings types

Augment

Random per-image transforms a loader applies to training batches. Flips and 90-degree turns (only half turns for non-square images) keep the pixels; crop_pad pads by n pixels (edge values) and crops back to the size at a random offset; brightness/contrast add or scale by a random amount up to the value (in units of the 0..255 range); noise adds Gaussian noise with that standard deviation.

Field	Type	Default	Description
flip_x	boolean	false	Mirror left-right with probability 0.5
flip_y	boolean	false	Mirror top-bottom with probability 0.5
rotate90	boolean	false	Turn by a random multiple of 90 degrees
crop_pad	integer	0	Pad by n pixels, then crop at random (≥ 0, ≤ 64)
brightness	number	0	Largest brightness change (≥ 0, ≤ 1)
contrast	number	0	Largest relative contrast change (≥ 0, ≤ 1)
noise	number	0	Gaussian noise sd (0..1 units) (≥ 0, ≤ 0.5)

EarlyStopping

Stop when the monitored validation metric has not improved for patience evaluations.

Field	Type	Default	Description
monitor	val_loss or val_accuracy or val_rmse	val_loss	The validation metric to watch
patience	integer	5	Evaluations without improvement (≥ 1, ≤ 1000)
min_delta	number	0.0	The smallest change that counts as better (≥ 0)
restore_best	boolean	true	Keep the weights of the best evaluation

TrainConfig

How a network is trained (optimizer, schedule, validation, logging).

Field	Type	Default	Description
epochs	integer	20	Passes over the training rows (≥ 1, ≤ 10000)
batch_size	integer	64	Rows per gradient step (≥ 1, ≤ 8192)
optimizer	adamw or adam or sgd	adamw
learning_rate	number	0.001	Peak learning rate (> 0, ≤ 10)
weight_decay	number	0.0001	Decoupled weight decay (adamw) (≥ 0, ≤ 1)
momentum	number	0.9	Momentum (sgd) (≥ 0, < 1)
schedule	constant or cosine or warmup_cosine	warmup_cosine
warmup_epochs	number	1.0	Linear warm-up (warmup_cosine) (≥ 0, ≤ 100)
grad_clip	number (optional)	1.0	Clip the global gradient norm
label_smoothing	number	0.0	Classification only (≥ 0, < 1)
val_fraction	number	0.1	Stratified share of the training rows for validation (≥ 0, < 0.9)
val_split_value	string (optional)	val	Rows whose split coord equals this value are the validation set
early_stopping	EarlyStopping (optional)	–
log_every_steps	integer (optional)	null	Steps between logged training metrics (default: <= 1000 points)
eval_every_epochs	integer	1	Epochs between validation passes (≥ 1)
media_every_epochs	integer (optional)	null	Epochs between logged images (default: at most 20 times)
max_steps	integer (optional)	null	Stop after this many gradient steps

UncertaintyConfig

How a trained network gives draws (chain=1, draw).

Field	Type	Default	Description
method	laplace or mc_dropout or none	laplace	laplace: last-layer Laplace; mc_dropout: dropout at score; none: one draw
draws	integer (optional)	null	Draws per prediction (default: fit draws, else 100)
hessian	auto or full or diag	auto	auto: full when (features + 1) x outputs <= 2048
max_rows	integer	20000	Training rows used for the Laplace Hessian (≥ 10)
prior_precision	number (optional)	null	None: maximise the Laplace evidence on a grid 1e-4..1e4

Input data contract (schema template)

{
  "params": {
    "pixels": {
      "type": "string",
      "default": "pixels"
    },
    "target": {
      "type": "string",
      "default": "label"
    },
    "image_dim": {
      "type": "string",
      "default": "image"
    }
  },
  "template": {
    "dims": {
      "{image_dim}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "pooled"
      },
      "y": {
        "kind": "integer",
        "policy": "fixed"
      },
      "x": {
        "kind": "integer",
        "policy": "fixed"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "{pixels}": {
        "dims": [
          "{image_dim}",
          "y",
          "x",
          "channel"
        ],
        "dtype": "uint8",
        "role": "covariate"
      },
      "{target}": {
        "dims": [
          "{image_dim}"
        ],
        "dtype": "str",
        "role": "target",
        "missing": "allow"
      }
    }
  }
}

dl.mlp

A multilayer perceptron for a table: numeric features predict a number or a class. It logs training and validation curves and gives draws by a last-layer Laplace approximation or MC dropout.

Version	1.0.0
Inference	custom
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: a network trained by gradient descent has no pointwise log likelihood
Default plots	ml.training_curves
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	()	–	Prediction per row; a classification adds a class dim (the probability of each class, summing to 1 over class)

Settings (model.settings)

Field	Type	Default	Level	Description
target	string	y	–	The target variable
rows	list of string	required	–	The dims of a row (min items 1, max items 8)
target_dtype	float64 or str or int64	float64	–	dtype of the target: numbers, text labels or integer labels
features	list of string (optional)	null	–	Feature variables (default: every numeric one)
task	auto or regression or classification	auto	–	auto: classification for text or integer labels
new_levels	pooled or zero or error	pooled	–	Unseen row levels at score
hidden	list of integer	[64, 64]	–	Hidden layer widths (max items 16)
dropout	number	0.1	–	Dropout between and after the layers (≥ 0, < 1)
train	TrainConfig	–	–
uncertainty	UncertaintyConfig	–	–
class_order	list of string (optional)	null	–

Nested settings types

EarlyStopping

Stop when the monitored validation metric has not improved for patience evaluations.

Field	Type	Default	Description
monitor	val_loss or val_accuracy or val_rmse	val_loss	The validation metric to watch
patience	integer	5	Evaluations without improvement (≥ 1, ≤ 1000)
min_delta	number	0.0	The smallest change that counts as better (≥ 0)
restore_best	boolean	true	Keep the weights of the best evaluation

TrainConfig

How a network is trained (optimizer, schedule, validation, logging).

Field	Type	Default	Description
epochs	integer	20	Passes over the training rows (≥ 1, ≤ 10000)
batch_size	integer	64	Rows per gradient step (≥ 1, ≤ 8192)
optimizer	adamw or adam or sgd	adamw
learning_rate	number	0.001	Peak learning rate (> 0, ≤ 10)
weight_decay	number	0.0001	Decoupled weight decay (adamw) (≥ 0, ≤ 1)
momentum	number	0.9	Momentum (sgd) (≥ 0, < 1)
schedule	constant or cosine or warmup_cosine	warmup_cosine
warmup_epochs	number	1.0	Linear warm-up (warmup_cosine) (≥ 0, ≤ 100)
grad_clip	number (optional)	1.0	Clip the global gradient norm
label_smoothing	number	0.0	Classification only (≥ 0, < 1)
val_fraction	number	0.1	Stratified share of the training rows for validation (≥ 0, < 0.9)
val_split_value	string (optional)	val	Rows whose split coord equals this value are the validation set
early_stopping	EarlyStopping (optional)	–
log_every_steps	integer (optional)	null	Steps between logged training metrics (default: <= 1000 points)
eval_every_epochs	integer	1	Epochs between validation passes (≥ 1)
media_every_epochs	integer (optional)	null	Epochs between logged images (default: at most 20 times)
max_steps	integer (optional)	null	Stop after this many gradient steps

UncertaintyConfig

How a trained network gives draws (chain=1, draw).

Field	Type	Default	Description
method	laplace or mc_dropout or none	laplace	laplace: last-layer Laplace; mc_dropout: dropout at score; none: one draw
draws	integer (optional)	null	Draws per prediction (default: fit draws, else 100)
hessian	auto or full or diag	auto	auto: full when (features + 1) x outputs <= 2048
max_rows	integer	20000	Training rows used for the Laplace Hessian (≥ 10)
prior_precision	number (optional)	null	None: maximise the Laplace evidence on a grid 1e-4..1e4

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "target_dtype": {
      "type": "string",
      "enum": [
        "float64",
        "str",
        "int64"
      ],
      "default": "float64"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "{target_dtype}",
        "role": "target",
        "missing": "allow"
      }
    }
  }
}

mmm.bayesian

mmm-framework's Bayesian marketing-mix model: adstock and saturation per channel, trend and seasonality, national or hierarchical over geos and products, channel effects stated as returns on spend, business priors and lift tests as calibration, budget optimisation with mmm-framework planning.

Version	1.1.0
Inference	mcmc
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	additive
Decision objective	yes
Fast scoring	–
Owns	carryover, inverse_units, scaling
LOO	valid
Default plots	core.ppc_fan, core.convergence, core.energy, core.posterior_forest, core.prior_posterior, mmm.decomposition, mmm.waterfall, mmm.response_curves, mmm.marginal_cost, mmm.roi, mmm.spend_effect, mmm.adstock, mmm.lift_tests, mmm.effectiveness
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
y_hat	(time)	of y	Expected (or predictive) KPI over the panel dims and time
contributions	(component, time)	of y	Every channel, the baseline, trend, seasonality and each control; they sum to y_hat

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
contributions	(sample, channel, time)	target	fit	Channel, baseline, trend, seasonality and control contributions over the fit window per thinned draw (panel dims included)
response_curves	(sample, channel, multiplier)	target	fit	Total contribution per spend multiplier and channel (and geo) per thinned draw (Guided-mode slider grids, budget optimisation)
marginal_cost	(sample, channel)	spend per target	fit	Extra spend per extra unit of the KPI at current spend
adstock	(sample, channel, lag)	share	fit	Share of a unit of spend's total effect that lands each number of periods after the spend, per thinned draw (carry-over shape)
effectiveness	(sample, channel, time)	target per spend	fit	KPI per unit of spend at the average spend level with the channel effect of each period (changes over time only for time-varying channels)

Settings (model.settings)

Field	Type	Default	Level	Description
adstock	AdstockSettings	{"kind": "geometric", "l_max": 8, "normalize": true}	basic
saturation	SaturationSettings	{"kind": "logistic"}	basic
channels	map of ChannelSettings	–	advanced	Per-channel adstock/saturation overrides (channel label -> settings)
trend	TrendSettings	{"kind": "linear", "n_changepoints": 10, "changepoint_range…	basic
seasonality	SeasonalitySettings	{"yearly_order": 2, "monthly_order": null, "weekly_order": …	basic
hierarchy	HierarchySettings	{"pool_over": [], "media_by_geo": false, "media_geo_sigma":…	basic
controls	list of string (optional)	null	advanced	Control variables (default: every other numeric variable over time)
frequency	W or D or M (optional)	null	advanced	Data frequency: W (weekly), D (daily) or M (monthly); default: read from the time labels
new_levels	pooled or hyperprior or zero or error	pooled	advanced	How a geo or product unseen at fit is scored: pooled (the overall level and average channel effects), hyperprior (a level drawn from the fitted spread, so its range is wider), zero (no channel effects) or error (refuse)
grid	GridSettings	{"max_multiplier": 2.5, "points": 26, "marginal_step": 0.01}	advanced
media_prior	roi or coefficient	roi	advanced	Parameterise channel effects by their return on spend (priors are stated in ROI, comparable across channels) or by a standardised coefficient
priors	PriorSettings	{"roi": {}, "roi_default": null, "intercept": null}	advanced	Prior overrides (written by business-prior translation)
likelihood	normal or student_t or poisson or negative_binomial or beta	normal	advanced	Observation model of the KPI: normal, student_t (heavy-tailed outliers), poisson or negative_binomial (counts, log link) or beta (shares between 0 and 1, logit link)
media_effect_sigma	number	0.5	advanced	Count and share likelihoods only: prior scale (half-normal) of a channel's largest effect on the link scale (> 0, ≤ 5)
time_varying	TimeVaryingSettings	{"channels": [], "innovation_sigma": 0.15, "forecast": "ran…	advanced	Channels whose effect changes over time

Nested settings types

AdstockSettings

Carry-over of spend into later periods (mmm-framework AdstockConfig).

Field	Type	Default	Description
kind	geometric or delayed or weibull or none	geometric	geometric: decays from the spend period; delayed: peaks a few periods later; weibull: a flexible shape; none: no carry-over
l_max	integer	8	How many periods an effect can last (the kernel length) (≥ 1, ≤ 52)
normalize	boolean	true	Normalise the lag weights to sum to one (the effect size stays in the channel coefficient)

ChannelSettings

Per-channel overrides of the shared adstock and saturation settings.

Field	Type	Default	Description
adstock	AdstockSettings (optional)	null
saturation	SaturationSettings (optional)	null

GridSettings

Response-curve grids cached per thinned draw (Guided-mode sliders, optimisation).

Field	Type	Default	Description
max_multiplier	number	2.5	Largest spend multiplier on the grid (1 = spend over the fit window, which the grid always holds) (≥ 1, ≤ 20)
points	integer	26	Grid points from 0 to max_multiplier (plus 1, today's spend, when the spacing skips it) (≥ 3, ≤ 201)
marginal_step	number	0.01	Relative spend change used for marginal cost (central difference) (> 0, ≤ 0.25)

HierarchySettings

Panel structure: the dims the KPI varies over besides time. pool_over lists the panel dims of the data (geo and/or product). Every level gets its own baseline, partially pooled towards the overall level (HierarchicalConfig.pool_across_geo/product); with media_by_geo each geo also gets its own channel effects, partially pooled across geos.

Field	Type	Default	Description
pool_over	list of geo or product	–	Panel dims of the data (geo, product); empty for one national series (max items 2)
media_by_geo	boolean	false	Let channel effects differ by geo (partially pooled)
media_geo_sigma	number	0.3	Prior sd of the geo deviations of log channel effects (> 0, ≤ 5)

NormalPrior

Normal(mu, sigma) on the standardised KPI scale (KPI standard deviations).

Field	Type	Default	Description
mu	number	required	≥ -50, ≤ 50
sigma	number	required	> 0, ≤ 50

PriorSettings

Prior overrides; business-prior translation (translate_priors) writes them. * roi - per channel, a log-normal prior on the return on spend (mmm-framework's ROI-parameterised channel effect, roi_prior_mu/roi_prior_sigma); * roi_default - the log-normal ROI prior of channels not listed (default LogNormal(0, 1): median break-even); * intercept - the baseline level in KPI standard deviations from the mean.

Field	Type	Default	Description
roi	map of RoiPrior	–	Per-channel log-normal ROI priors (channel label -> {mu, sigma})
roi_default	RoiPrior (optional)	null
intercept	NormalPrior (optional)	null

RoiPrior

A log-normal prior on a channel's return on spend (KPI per unit of spend over the fit window): roi ~ LogNormal(mu, sigma), so exp(mu) is the median ROI.

Field	Type	Default	Description
mu	number	required	Log of the median ROI (≥ -20, ≤ 20)
sigma	number	required	Spread on the log scale (> 0, ≤ 10)

SaturationSettings

Diminishing returns of (carried-over) spend (mmm-framework SaturationConfig).

Field	Type	Default	Description
kind	logistic or hill or michaelis_menten or tanh or root or none	logistic	logistic (default, most stable), hill (S-shaped), michaelis_menten, tanh, root (power curve) or none (a straight line)

SeasonalitySettings

Fourier seasonality (mmm-framework SeasonalityConfig).

Field	Type	Default	Description
yearly_order	integer (optional)	2	Fourier pairs of the yearly cycle (0 or null: none)
monthly_order	integer (optional)	null
weekly_order	integer (optional)	null	Daily data only
prior_sigma	number	0.3	Prior sd of each Fourier coefficient, in KPI standard deviations (raise it for strongly seasonal categories) (> 0, ≤ 5)

TimeVaryingSettings

Channels whose effect drifts over time (mmm-framework's time-varying coefficient: a random walk on the log of the channel coefficient, one step per period).

Field	Type	Default	Description
channels	list of string	–	Channels whose effect may change over time (creative wear-out, changes in the market) (max items 64)
innovation_sigma	number	0.15	Prior scale of the change in the log effect from one period to the next (smaller: closer to a constant effect) (> 0, ≤ 2)
forecast	random_walk or hold_last or mean	random_walk	The effect after the fit window: random_walk (continue the random walk, so the range widens with the horizon), hold_last (the last fitted value) or mean (the average over the fit window)

TrendSettings

The slow-moving baseline (mmm-framework TrendConfig).

Field	Type	Default	Description
kind	none or linear or piecewise or spline or gp	linear	none, linear (default), piecewise (Prophet-style changepoints), spline or gp (a smooth Gaussian-process curve). Forecasts hold flexible trends at their last fitted level; a linear trend is extrapolated.
n_changepoints	integer	10	≥ 1, ≤ 50
changepoint_range	number	0.8	> 0, ≤ 1
n_knots	integer	10	≥ 3, ≤ 50
growth_prior_sigma	number	0.5	Prior sd of the linear slope, in KPI standard deviations over the whole history (> 0, ≤ 10)

Input data contract (schema template)

{
  "params": {
    "panel": {
      "type": "string_list",
      "path": "hierarchy.pool_over",
      "default": []
    },
    "strategy": {
      "type": "string",
      "path": "new_levels",
      "enum": [
        "pooled",
        "hyperprior",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "time": {
        "kind": "time",
        "policy": "extend"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "{*panel as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{strategy}"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*panel}",
          "time"
        ],
        "dtype": "float64",
        "role": "target"
      },
      "spend": {
        "dims": [
          "{*panel}",
          "channel",
          "time"
        ],
        "dtype": "float64",
        "role": "spend",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "strategy": "error"
      },
      "merge": {
        "dims": {
          "{*panel as d}": {
            "policy": "fixed"
          }
        }
      }
    }
  ]
}

mmm.frequentist

mmm-framework's fast frequentist marketing-mix model: adstock and saturation chosen by an out-of-sample search, a ridge solve with non-negative channel effects and bootstrap intervals; seconds instead of minutes.

Version	1.0.0
Inference	bootstrap
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	additive
Decision objective	yes
Fast scoring	–
Owns	carryover, inverse_units, scaling
LOO	invalid: bootstrap replicates are not posterior draws
Default plots	core.ppc_fan, mmm.decomposition, mmm.response_curves, mmm.marginal_cost
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(time)	of y	Expected (or predictive) KPI over the panel dims and time
contributions	(component, time)	of y	Every channel, the baseline, trend, seasonality and each control; they sum to y_hat

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
contributions	(sample, channel, time)	target	fit	Channel, baseline, trend, seasonality and control contributions over the fit window per thinned draw (panel dims included)
response_curves	(sample, channel, multiplier)	target	fit	Total contribution per spend multiplier and channel (and geo) per thinned draw (Guided-mode slider grids, budget optimisation)
marginal_cost	(sample, channel)	spend per target	fit	Extra spend per extra unit of the KPI at current spend
adstock	(sample, channel, lag)	share	fit	Share of a unit of spend's total effect that lands each number of periods after the spend, per thinned draw (carry-over shape)
effectiveness	(sample, channel, time)	target per spend	fit	KPI per unit of spend at the average spend level with the channel effect of each period (changes over time only for time-varying channels)

Settings (model.settings)

Field	Type	Default	Level	Description
adstock	AdstockSettings	{"kind": "geometric", "l_max": 8, "normalize": true}	basic
saturation	SaturationSettings	{"kind": "logistic"}	basic
channels	map of ChannelSettings	–	advanced	Per-channel adstock/saturation overrides (channel label -> settings)
trend	TrendSettings	{"kind": "linear", "n_changepoints": 10, "changepoint_range…	basic
seasonality	SeasonalitySettings	{"yearly_order": 2, "monthly_order": null, "weekly_order": …	basic
hierarchy	HierarchySettings	{"pool_over": [], "media_by_geo": false, "media_geo_sigma":…	basic
controls	list of string (optional)	null	advanced	Control variables (default: every other numeric variable over time)
frequency	W or D or M (optional)	null	advanced	Data frequency: W (weekly), D (daily) or M (monthly); default: read from the time labels
new_levels	pooled or hyperprior or zero or error	pooled	advanced	How a geo or product unseen at fit is scored: pooled (the overall level and average channel effects), hyperprior (a level drawn from the fitted spread, so its range is wider), zero (no channel effects) or error (refuse)
grid	GridSettings	{"max_multiplier": 2.5, "points": 26, "marginal_step": 0.01}	advanced
estimator	ridge or constrained	ridge	advanced	ridge (closed form, non-negative channel effects) or constrained (a convex program; needs the optional cvxpy dependency)
bootstrap	integer	200	basic	Bootstrap replicates (the 'draws' of every output) (≥ 20, ≤ 5000)
search_budget	integer	200	advanced	Adstock/saturation candidates evaluated by the transform search (≥ 10, ≤ 5000)
ridge_alpha	number	1.0	advanced	Fallback L2 penalty on the standardised design (> 0, ≤ 1000000.0)

Nested settings types

AdstockSettings

Carry-over of spend into later periods (mmm-framework AdstockConfig).

Field	Type	Default	Description
kind	geometric or delayed or weibull or none	geometric	geometric: decays from the spend period; delayed: peaks a few periods later; weibull: a flexible shape; none: no carry-over
l_max	integer	8	How many periods an effect can last (the kernel length) (≥ 1, ≤ 52)
normalize	boolean	true	Normalise the lag weights to sum to one (the effect size stays in the channel coefficient)

ChannelSettings

Per-channel overrides of the shared adstock and saturation settings.

Field	Type	Default	Description
adstock	AdstockSettings (optional)	null
saturation	SaturationSettings (optional)	null

GridSettings

Response-curve grids cached per thinned draw (Guided-mode sliders, optimisation).

Field	Type	Default	Description
max_multiplier	number	2.5	Largest spend multiplier on the grid (1 = spend over the fit window, which the grid always holds) (≥ 1, ≤ 20)
points	integer	26	Grid points from 0 to max_multiplier (plus 1, today's spend, when the spacing skips it) (≥ 3, ≤ 201)
marginal_step	number	0.01	Relative spend change used for marginal cost (central difference) (> 0, ≤ 0.25)

HierarchySettings

Panel structure: the dims the KPI varies over besides time. pool_over lists the panel dims of the data (geo and/or product). Every level gets its own baseline, partially pooled towards the overall level (HierarchicalConfig.pool_across_geo/product); with media_by_geo each geo also gets its own channel effects, partially pooled across geos.

Field	Type	Default	Description
pool_over	list of geo or product	–	Panel dims of the data (geo, product); empty for one national series (max items 2)
media_by_geo	boolean	false	Let channel effects differ by geo (partially pooled)
media_geo_sigma	number	0.3	Prior sd of the geo deviations of log channel effects (> 0, ≤ 5)

SaturationSettings

Diminishing returns of (carried-over) spend (mmm-framework SaturationConfig).

Field	Type	Default	Description
kind	logistic or hill or michaelis_menten or tanh or root or none	logistic	logistic (default, most stable), hill (S-shaped), michaelis_menten, tanh, root (power curve) or none (a straight line)

SeasonalitySettings

Fourier seasonality (mmm-framework SeasonalityConfig).

Field	Type	Default	Description
yearly_order	integer (optional)	2	Fourier pairs of the yearly cycle (0 or null: none)
monthly_order	integer (optional)	null
weekly_order	integer (optional)	null	Daily data only
prior_sigma	number	0.3	Prior sd of each Fourier coefficient, in KPI standard deviations (raise it for strongly seasonal categories) (> 0, ≤ 5)

TrendSettings

The slow-moving baseline (mmm-framework TrendConfig).

Field	Type	Default	Description
kind	none or linear or piecewise or spline or gp	linear	none, linear (default), piecewise (Prophet-style changepoints), spline or gp (a smooth Gaussian-process curve). Forecasts hold flexible trends at their last fitted level; a linear trend is extrapolated.
n_changepoints	integer	10	≥ 1, ≤ 50
changepoint_range	number	0.8	> 0, ≤ 1
n_knots	integer	10	≥ 3, ≤ 50
growth_prior_sigma	number	0.5	Prior sd of the linear slope, in KPI standard deviations over the whole history (> 0, ≤ 10)

Input data contract (schema template)

{
  "params": {
    "panel": {
      "type": "string_list",
      "path": "hierarchy.pool_over",
      "default": []
    },
    "strategy": {
      "type": "string",
      "path": "new_levels",
      "enum": [
        "pooled",
        "hyperprior",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "time": {
        "kind": "time",
        "policy": "extend"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "{*panel as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{strategy}"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*panel}",
          "time"
        ],
        "dtype": "float64",
        "role": "target"
      },
      "spend": {
        "dims": [
          "{*panel}",
          "channel",
          "time"
        ],
        "dtype": "float64",
        "role": "spend",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "strategy": "error"
      },
      "merge": {
        "dims": {
          "{*panel as d}": {
            "policy": "fixed"
          }
        }
      }
    }
  ]
}

mmm.multivariate

mmm-framework's multivariate marketing-mix model: one national spend series per channel drives two or more outcomes, with correlated residuals and cross-effects between outcomes (cannibalization, halo); budgets are optimised on the weighted total of the outcomes.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	additive
Decision objective	yes
Fast scoring	–
Owns	carryover, inverse_units, scaling
LOO	valid
Default plots	core.ppc_fan, core.convergence, core.energy, core.prior_posterior, mmm.decomposition, mmm.waterfall, mmm.response_curves, mmm.roi, mmm.spend_effect, mmm.adstock, mmm.cross_effects
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
y_hat	(time)	of y	Expected (or predictive) value of every outcome over time
contributions	(component, time)	of y	Every channel, the baseline, trend and seasonality per outcome, with cross-effects passed through; they sum to y_hat

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
contributions	(sample, channel, outcome, time)	target	fit	Channel, baseline, trend and seasonality contributions per outcome over the fit window per thinned draw
response_curves	(sample, channel, multiplier)	target	fit	Weighted total of the outcomes per spend multiplier and channel per thinned draw, plus the curves per outcome
marginal_cost	(sample, channel)	spend per target	fit	Extra spend per extra unit of the weighted outcome total at current spend
adstock	(sample, channel, lag)	–	fit	Share of a unit of spend's effect that lands each period later
cross_effects	(sample, pair)	target per target	fit	Effect of one outcome on another (target units per unit of source)

Settings (model.settings)

Field	Type	Default	Level	Description
trend	TrendSettings	{"kind": "linear", "n_changepoints": 10, "changepoint_range…	basic
seasonality	SeasonalitySettings	{"yearly_order": 2, "monthly_order": null, "weekly_order": …	basic
frequency	W or D or M (optional)	null	advanced	Data frequency: W (weekly), D (daily) or M (monthly); default: read from the time labels
grid	GridSettings	{"max_multiplier": 2.5, "points": 26, "marginal_step": 0.01}	advanced
likelihood	normal or student_t	normal	advanced	Observation noise: normal, or student_t for heavy-tailed outliers
outcomes	list of OutcomeSettings	–	basic	Settings per outcome; empty: every outcome of the data, with defaults (max items 8)
primary_outcome	string (optional)	null	basic	The outcome reports show first (default: the first outcome)
cross_effects	list of CrossEffectSettings	–	basic	Effects of one outcome on another (max items 28)
lkj_eta	number	2.0	advanced	LKJ concentration of the residual correlation (larger: closer to none) (> 0, ≤ 50)
share_adstock	boolean	true	advanced	One carry-over rate for every channel
share_trend	boolean	false	advanced	One trend for every outcome
share_seasonality	boolean	true	advanced	One seasonal pattern for every outcome

Nested settings types

CrossEffectSettings

An effect of one outcome on another (mmm-framework CrossEffectConfig).

Field	Type	Default	Description
source	string	required
target	string	required
kind	unconstrained or cannibalization or halo	unconstrained	unconstrained (the data picks the sign), cannibalization (the source takes sales from the target) or halo (the source adds sales to the target)
prior_sigma	number	0.3	Prior sd of the effect (target units per unit of source) (> 0, ≤ 5)

GridSettings

Response-curve grids cached per thinned draw (Guided-mode sliders, optimisation).

Field	Type	Default	Description
max_multiplier	number	2.5	Largest spend multiplier on the grid (1 = spend over the fit window, which the grid always holds) (≥ 1, ≤ 20)
points	integer	26	Grid points from 0 to max_multiplier (plus 1, today's spend, when the spacing skips it) (≥ 3, ≤ 201)
marginal_step	number	0.01	Relative spend change used for marginal cost (central difference) (> 0, ≤ 0.25)

OutcomeSettings

One outcome (a label of the data's outcome dim) and its priors.

Field	Type	Default	Description
name	string	required	A label of the data's outcome dim
weight	number	1.0	Value of one unit of this outcome in the budget objective (the plans and response curves use the weighted total of the outcomes) (≥ 0, ≤ 1000000.0)
include_trend	boolean	true
include_seasonality	boolean	true
intercept_sigma	number	2.0	Prior sd of the baseline level, in outcome standard deviations (> 0, ≤ 20)
media_sigma	number	0.5	Prior sd of each channel effect, in outcome standard deviations (> 0, ≤ 10)

SeasonalitySettings

Fourier seasonality (mmm-framework SeasonalityConfig).

Field	Type	Default	Description
yearly_order	integer (optional)	2	Fourier pairs of the yearly cycle (0 or null: none)
monthly_order	integer (optional)	null
weekly_order	integer (optional)	null	Daily data only
prior_sigma	number	0.3	Prior sd of each Fourier coefficient, in KPI standard deviations (raise it for strongly seasonal categories) (> 0, ≤ 5)

TrendSettings

The slow-moving baseline (mmm-framework TrendConfig).

Field	Type	Default	Description
kind	none or linear or piecewise or spline or gp	linear	none, linear (default), piecewise (Prophet-style changepoints), spline or gp (a smooth Gaussian-process curve). Forecasts hold flexible trends at their last fitted level; a linear trend is extrapolated.
n_changepoints	integer	10	≥ 1, ≤ 50
changepoint_range	number	0.8	> 0, ≤ 1
n_knots	integer	10	≥ 3, ≤ 50
growth_prior_sigma	number	0.5	Prior sd of the linear slope, in KPI standard deviations over the whole history (> 0, ≤ 10)

Input data contract (schema template)

{
  "template": {
    "dims": {
      "time": {
        "kind": "time",
        "policy": "extend"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "outcome": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "outcome",
          "time"
        ],
        "dtype": "float64",
        "role": "target"
      },
      "spend": {
        "dims": [
          "channel",
          "time"
        ],
        "dtype": "float64",
        "role": "spend",
        "future_required": true
      }
    }
  }
}

mmm.nested

mmm-framework's nested marketing-mix model: channels move mediators (visits, awareness, measured in some periods or latent) that move the KPI, plus direct channel effects; one national series, trend and seasonality, lift tests as calibration, budget optimisation.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	additive
Decision objective	yes
Fast scoring	–
Owns	carryover, inverse_units, scaling
LOO	valid
Default plots	core.ppc_fan, core.convergence, core.energy, core.prior_posterior, mmm.decomposition, mmm.waterfall, mmm.response_curves, mmm.roi, mmm.spend_effect, mmm.adstock, mmm.mediation, mmm.lift_tests
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
y_hat	(time)	of y	Expected (or predictive) KPI over time
contributions	(component, time)	of y	Every channel's total effect (through the mediators and direct), the baseline, trend and seasonality; they sum to y_hat
mediator_hat	(mediator, time)	of mediators	Expected mediators (latent mediators in standard units)

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
contributions	(sample, channel, time)	target	fit	Channel, baseline, trend and seasonality contributions over the fit window per thinned draw
response_curves	(sample, channel, multiplier)	target	fit	Total contribution per spend multiplier and channel per thinned draw (Guided-mode slider grids, budget optimisation)
marginal_cost	(sample, channel)	spend per target	fit	Extra spend per extra unit of the KPI at current spend
adstock	(sample, channel, lag)	–	fit	Share of a unit of spend's effect that lands each period later
mediation	(sample, channel, mediator)	target	fit	Direct and indirect (through each mediator) effect of each channel over the fit window

Settings (model.settings)

Field	Type	Default	Level	Description
trend	TrendSettings	{"kind": "linear", "n_changepoints": 10, "changepoint_range…	basic
seasonality	SeasonalitySettings	{"yearly_order": 2, "monthly_order": null, "weekly_order": …	basic
frequency	W or D or M (optional)	null	advanced	Data frequency: W (weekly), D (daily) or M (monthly); default: read from the time labels
grid	GridSettings	{"max_multiplier": 2.5, "points": 26, "marginal_step": 0.01}	advanced
likelihood	normal or student_t	normal	advanced	Observation noise: normal, or student_t for heavy-tailed outliers
mediators	list of MediatorSettings	–	–	The mediators (default: every label of the mediator dim, measured, moved by every channel, with direct effects) (max items 8)

Nested settings types

GridSettings

Response-curve grids cached per thinned draw (Guided-mode sliders, optimisation).

Field	Type	Default	Description
max_multiplier	number	2.5	Largest spend multiplier on the grid (1 = spend over the fit window, which the grid always holds) (≥ 1, ≤ 20)
points	integer	26	Grid points from 0 to max_multiplier (plus 1, today's spend, when the spacing skips it) (≥ 3, ≤ 201)
marginal_step	number	0.01	Relative spend change used for marginal cost (central difference) (> 0, ≤ 0.25)

MediatorSettings

One mediator: a measured (or latent) result that channels move and that moves the KPI (mmm-framework MediatorConfig).

Field	Type	Default	Description
name	string	required	A mediator label
channels	list of string (optional)	null	Channels that move this mediator (default: every channel)
direct_effect	boolean	true	The channels of this mediator also act on the KPI directly
latent	boolean	false	The mediator is not measured: its series is ignored (it can be empty)
media_sign	positive or none	positive	Sign of the channel effects on the mediator
outcome_sign	none or positive or negative	none	Sign of the mediator's effect on the KPI
observation_sigma	number	0.1	Prior scale of the measurement noise, in mediator standard deviations (> 0, ≤ 5)

SeasonalitySettings

Fourier seasonality (mmm-framework SeasonalityConfig).

Field	Type	Default	Description
yearly_order	integer (optional)	2	Fourier pairs of the yearly cycle (0 or null: none)
monthly_order	integer (optional)	null
weekly_order	integer (optional)	null	Daily data only
prior_sigma	number	0.3	Prior sd of each Fourier coefficient, in KPI standard deviations (raise it for strongly seasonal categories) (> 0, ≤ 5)

TrendSettings

The slow-moving baseline (mmm-framework TrendConfig).

Field	Type	Default	Description
kind	none or linear or piecewise or spline or gp	linear	none, linear (default), piecewise (Prophet-style changepoints), spline or gp (a smooth Gaussian-process curve). Forecasts hold flexible trends at their last fitted level; a linear trend is extrapolated.
n_changepoints	integer	10	≥ 1, ≤ 50
changepoint_range	number	0.8	> 0, ≤ 1
n_knots	integer	10	≥ 3, ≤ 50
growth_prior_sigma	number	0.5	Prior sd of the linear slope, in KPI standard deviations over the whole history (> 0, ≤ 10)

Input data contract (schema template)

{
  "template": {
    "dims": {
      "time": {
        "kind": "time",
        "policy": "extend"
      },
      "channel": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "mediator": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "time"
        ],
        "dtype": "float64",
        "role": "target"
      },
      "spend": {
        "dims": [
          "channel",
          "time"
        ],
        "dtype": "float64",
        "role": "spend",
        "future_required": true
      },
      "mediators": {
        "dims": [
          "mediator",
          "time"
        ],
        "dtype": "float64",
        "role": "covariate",
        "missing": "allow"
      }
    }
  }
}

sci.areal_bym2

Counts of a disease in areas (districts, counties) against their expected counts, with a covariate and area effects that borrow strength from neighbouring areas: an ICAR, a proper CAR, BYM2 (a scaled ICAR mixed with independent noise) or iid effects. Outputs the smoothed relative risk of each area, the probability that it is above 1, and the expected counts. The polygons and the graph go into the bundle for the maps.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.area_map, net.graph, core.ridge
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
cases	(area)	of cases	Cases in each area: the expected count expected * rr, or a Poisson draw
rr	(area)	1	Relative risk of each area (1 = the reference rate)
p_exceed	(area)	probability	1 when rr is above exceed_threshold in a draw, else 0: the mean over draws is the probability P(rr > threshold)
effect	(area)	1	The area effect u on the log relative risk (what the covariate does not explain)

Settings (model.settings)

Field	Type	Default	Level	Description
kind	poisson or iid or icar or car or bym2	bym2	–	Prior of the area effects: poisson (none), iid, icar, car (proper CAR) or bym2 (ICAR plus iid with a mixing share rho) (x-ui-level: basic)
islands	link_nearest or isolate or error	link_nearest	–	Areas without neighbours: link_nearest links each to the area with the closest boundary; isolate keeps them alone (iid and icar only); error stops (x-ui-level: basic)
covariate_scale	number	10.0	–	Divide the centred covariate by this (b1 is the effect of one unit) (> 0)
center_covariate	boolean	true	–	Subtract the mean of the covariate
intercept_sd	number	1.0	–	Prior sd of the intercept b0 (> 0)
slope_sd	number	1.0	–	Prior sd of the covariate effect b1 (> 0)
sigma_scale	number	1.0	–	Half-normal scale of sigma, the sd of the area effects (> 0)
rho_alpha	number	1.0	–	Beta prior of rho (bym2): first shape (> 0)
rho_beta	number	1.0	–	Beta prior of rho (bym2): second shape (> 0)
exceed_threshold	number	1.0	–	p_exceed is the probability that rr is above this value (> 0)

Input data contract (schema template)

{
  "template": {
    "dims": {
      "area": {
        "kind": "categorical",
        "policy": "fixed",
        "description": "The areas (districts, counties): the nodes of the graph"
      },
      "area_": {
        "kind": "categorical",
        "policy": "fixed",
        "description": "The same areas, as the second axis of the adjacency matrix"
      },
      "vertex": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Vertices of the polygon outlines (NaN padded)"
      }
    },
    "vars": {
      "cases": {
        "dims": [
          "area"
        ],
        "dtype": "float64",
        "role": "target",
        "units": "cases",
        "missing": "allow",
        "bounds": [
          0,
          null
        ]
      },
      "expected": {
        "dims": [
          "area"
        ],
        "dtype": "float64",
        "role": "exposure",
        "units": "cases",
        "bounds": [
          0,
          null
        ]
      },
      "covariate": {
        "dims": [
          "area"
        ],
        "dtype": "float64",
        "role": "covariate"
      },
      "adjacency": {
        "dims": [
          "area",
          "area_"
        ],
        "dtype": "int64",
        "role": "constant",
        "bounds": [
          0,
          1
        ]
      },
      "ring_x": {
        "dims": [
          "area",
          "vertex"
        ],
        "dtype": "float64",
        "role": "constant",
        "missing": "allow"
      },
      "ring_y": {
        "dims": [
          "area",
          "vertex"
        ],
        "dtype": "float64",
        "role": "constant",
        "missing": "allow"
      }
    }
  }
}

sci.circular_regression

Angles (wind directions, times of day) on the circle. Regression kinds put the hour of day and the day of year into harmonic features and use a von Mises with a canonical link (or a projected normal, or a mixture of two regimes with a gate); kind activity fits a log-Fourier density per group (camera-trap times) with activity levels and overlaps. Outputs are mean directions, resultant lengths and predictive densities by block of the day or by group.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.rose, sci.profile_bands, sci.field_map
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
direction_hat	(obs)	of direction	Mean direction of the row's predictive distribution (expected), or a simulated angle (predictive), from 0 to one turn
resultant	(obs)	1	Mean resultant length of the row's predictive distribution (1: one direction, 0: no preferred direction)
density	(frame, angle)	1/unit	Predictive density per unit of the direction on a grid of angles, for each block of the day (regression) or group (activity)
mean_direction	(frame)	of direction	Direction of the mean resultant vector of each frame
resultant_length	(frame)	1	Length of the mean resultant vector of each frame
activity_level	(frame)	1	kind activity: fraction of the day the group is active, 1 / (2 pi max f)
overlap	(frame, frame_)	1	kind activity: coefficient of overlap of the activity densities of two groups

Settings (model.settings)

Field	Type	Default	Level	Description
kind	von_mises or projected_normal or regime_mixture or activity	von_mises	–	von_mises (canonical link), projected_normal, regime_mixture (two von Mises regimes with a gate), or activity (a log-Fourier density per group) (x-ui-level: basic)
units	degrees or radians or hours	degrees	–	Units of the direction: degrees, radians or hours (a time of day) (x-ui-level: basic)
group_dim	string	group	–	kind activity: the dim of the groups (species); the index variable is _idx
daily_harmonics	integer	3	–	Harmonics of the hour of day (≥ 0, ≤ 8)
annual_harmonics	integer	2	–	Harmonics of the day of year (≥ 0, ≤ 6)
interaction_harmonics	integer	2	–	Daily harmonics multiplied by the first annual harmonic (the daily cycle changes with the season) (≥ 0, ≤ 8)
intercept_sd	number	2.0	–	Prior sd of the intercepts (not used by regime_mixture) (> 0)
coef_sd	number	1.0	–	Prior sd of the other coefficients (> 0)
regime_directions_deg	list of number	–	–	regime_mixture: prior mean direction of each regime (degrees, in the units' own orientation) (min items 2, max items 2)
regime_names	list of string	–	–	regime_mixture: names of the two regimes (min items 2, max items 2)
regime_strength	number	2.0	–	regime_mixture: prior mean concentration of each regime (> 0)
regime_daily_harmonics	integer	2	–	≥ 0, ≤ 8
regime_annual_harmonics	integer	1	–	≥ 0, ≤ 6
gate_intercept_sd	number	1.5	–	Prior sd of the gate intercept (> 0)
harmonics	integer	16	–	activity: Fourier harmonics (≥ 1, ≤ 64)
tau_sd	number	2.0	–	activity: half-normal scale of tau (> 0)
grid_points	integer	256	–	activity: angles of the normalising sum (≥ 32, ≤ 4096)
frames	integer	8	–	Regression kinds: blocks of the day for density and mean_direction (8: three hours each) (≥ 1, ≤ 48)
angle_points	integer	72	–	Angles of the density output (≥ 8, ≤ 720)
rows_per_frame	integer	200	–	Scored rows averaged per frame for the density (≥ 1, ≤ 20000)
score_draws	integer	400	–	Posterior draws kept per score (evenly spaced) (≥ 1, ≤ 20000)

Input data contract (schema template)

{
  "params": {
    "kind": {
      "type": "string",
      "enum": [
        "von_mises",
        "projected_normal",
        "regime_mixture",
        "activity"
      ],
      "default": "von_mises"
    },
    "group_dim": {
      "type": "string",
      "default": "group"
    }
  },
  "template": {
    "layout": "long",
    "dims": {
      "obs": {
        "kind": "integer",
        "policy": "extend",
        "description": "One row per report or record"
      }
    },
    "vars": {
      "direction": {
        "dims": [
          "obs"
        ],
        "dtype": "float64",
        "role": "target"
      }
    }
  },
  "cases": [
    {
      "when": {
        "kind": "projected_normal"
      },
      "merge": {
        "vars": {
          "hour": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "h",
            "bounds": [
              0,
              24
            ]
          },
          "day": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "d",
            "bounds": [
              0,
              367
            ]
          }
        }
      }
    },
    {
      "when": {
        "kind": "regime_mixture"
      },
      "merge": {
        "vars": {
          "hour": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "h",
            "bounds": [
              0,
              24
            ]
          },
          "day": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "d",
            "bounds": [
              0,
              367
            ]
          }
        }
      }
    },
    {
      "when": {
        "kind": "von_mises"
      },
      "merge": {
        "vars": {
          "hour": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "h",
            "bounds": [
              0,
              24
            ]
          },
          "day": {
            "dims": [
              "obs"
            ],
            "dtype": "float64",
            "role": "covariate",
            "units": "d",
            "bounds": [
              0,
              367
            ]
          }
        }
      }
    },
    {
      "when": {
        "kind": "activity"
      },
      "merge": {
        "dims": {
          "{group_dim}": {
            "kind": "categorical",
            "policy": "fixed"
          }
        },
        "vars": {
          "{group_dim}_idx": {
            "dims": [
              "obs"
            ],
            "dtype": "int64",
            "role": "index",
            "index_of": "{group_dim}"
          }
        }
      }
    }
  ]
}

sci.cmb_field

Field-level inference of a masked CMB temperature map: every pixel of the sky on a periodic flat patch is sampled jointly with the bandpowers by a Wandelt-Eriksen Gibbs sampler (constrained realisations by preconditioned conjugate gradients, an inverse-gamma bandpower step and a rescaling move). Gives the Wiener-filtered map, sky draws behind the mask, a per-pixel sd and the bandpowers.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: every pixel depends on one latent sky field with tens of thousands of parameters, so pointwise leave-one-out is not defined; compare variants by their bandpowers, amplitude and tilt
Default plots	core.convergence, sci.power_spectrum, sci.field_map, sci.image_residuals, sci.ppc_image_grid
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
sky	(row, col)	uK	Posterior draws of the beamed sky (constrained realisations; unconstrained skies with the posterior spectrum for simulate)
temperature_hat	(row, col)	of temperature	The sky draw (expected), plus noise with noise_sd (predictive)
bandpower	(bin)	uK^2	Bandpower D_ell of each bin
bandpower_ratio	(bin)	1	Bandpower over the reference bandpower, exp(q)
log_amplitude	()	1	Log amplitude of the bandpowers relative to the reference
tilt	()	1	Tilt of the bandpowers relative to the reference, at tilt_pivot

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
sky_moments	(stat, row, col)	uK	fit	Posterior mean (the Wiener filter averaged over the spectrum) and sd of each pixel over every kept Gibbs draw

Settings (model.settings)

Field	Type	Default	Level	Description
pixel_arcmin	number	7.5	–	Side of a map pixel (arcmin) (x-ui-level: basic) (> 0, ≤ 600)
beam_fwhm_arcmin	number	5.0	–	FWHM of the Gaussian beam of the map (arcmin) (≥ 0, ≤ 600)
pixel_subsamples	integer	8	–	Point samples per pixel side that the map pixels average (box window; 1 = none) (≥ 1, ≤ 64)
pixel_window	boolean	false	–	The data has a pixel_window(multipole) variable (for example the HEALPix window of the source map)
bin_edges	list of number	–	–	Multipole edges of the bandpower bins (x-ui-level: basic) (min items 2, max items 101)
projection	gnomonic or flat	gnomonic	–	gnomonic: correct the multipoles for the mean stretch of a tangent-plane projection; flat: no correction (x-ui-level: basic)
border	integer	8	–	Unobserved pixels around the raster on the periodic torus (≥ 0, ≤ 128)
ell_dense	number	300.0	–	Modes below this multipole get the dense CG preconditioner (≥ 0)
n_alias	integer	3	–	Aliased replicas folded into the prior power (≥ 0, ≤ 6)
mean_prior_sd	number	100.0	–	Prior sd of the patch mean (uK) (> 0)
cg_tol	number	1e-05	–	Relative residual that stops CG (> 0, ≤ 0.1)
cg_max_iter	integer	5000	–	Most CG iterations per solve (≥ 10, ≤ 100000)
rescale	boolean	true	–	Use the non-centred rescaling move
q_init_sd	number	0.3	–	Sd of the dispersed chain starts (≥ 0, ≤ 3)
tilt_pivot	number	500.0	–	Pivot multipole of the tilt (> 0)
tilt_max_ell	number	1450.0	–	Bins with a centre at or above this are left out of the tilt (> 0)
tilt_prior_sd	number	0.5	–	Prior sd of log amplitude and tilt (> 0)
field_draws	integer	40	–	Sky maps drawn per score (over all chains) (≥ 1, ≤ 1000)

Input data contract (schema template)

{
  "params": {
    "pixel_window": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "row": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Map rows (bottom up)"
      },
      "col": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Map columns"
      },
      "multipole": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Multipole ell = 0, 1, 2, ... of the reference spectrum"
      }
    },
    "vars": {
      "temperature": {
        "dims": [
          "row",
          "col"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow",
        "description": "CMB temperature of each pixel (uK); NaN where masked"
      },
      "noise_sd": {
        "dims": [
          "row",
          "col"
        ],
        "dtype": "float64",
        "role": "weight",
        "description": "Noise sd of each pixel (uK)"
      },
      "reference_dl": {
        "dims": [
          "multipole"
        ],
        "dtype": "float64",
        "role": "constant",
        "description": "Reference spectrum D_ell = ell (ell + 1) C_ell / 2 pi (uK^2)"
      }
    }
  },
  "cases": [
    {
      "when": {
        "pixel_window": true
      },
      "merge": {
        "vars": {
          "pixel_window": {
            "dims": [
              "multipole"
            ],
            "dtype": "float64",
            "role": "constant",
            "description": "Pixel window of the source map at each multipole"
          }
        }
      }
    }
  ]
}

sci.colour_naming

Colour naming in one language (World Color Survey): the probability that a speaker gives each term to each colour chip, as a softmax over quadratic CIELAB logits with zero-sum weights, pooled or with speaker-level intercepts and linear weights. Predicts new speakers and maps each term's share of speakers on the chip palette.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.swatches, sci.field_map
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
p_name	(speaker, chip, term)	probability	Probability that the speaker gives the term to the chip (new speakers: fresh effects per draw); predictive: one sampled term per draw
p_term	(term, lightness, hue)	probability	Share of the language's speakers who give the term to the chip, on the palette grid (cells without a chip: empty in the posterior, 0 in scores)
mode_term_rgb	(chip, rgb)	sRGB	Each chip in the display colour of its most probable term in the draw (the posterised map)
p_top	(chip)	probability	Share of speakers who give the chip the term that is most probable on average
k_eff_speaker	()	categories	Effective number of categories of one speaker
k_eff_language	()	categories	Effective number of categories of the language (all speakers)

Settings (model.settings)

Field	Type	Default	Level	Description
kind	hierarchical or pooled	hierarchical	–	hierarchical: every speaker has their own intercept and linear weights around the language's; pooled: all speakers share one naming (x-ui-level: basic)
features	cielab or grid_line or grid_circle	cielab	–	cielab: a quadratic surface in CIELAB; grid_line / grid_circle: on the printed palette, with the hue column as a line or as a circle (x-ui-level: basic)
w_sd	number	3.0	–	Prior sd of the zero-sum weights (> 0, ≤ 20)
sd_scale	number	1.0	–	Half-normal scale of the between-speaker sds (> 0, ≤ 10)
n_random	integer	4	–	Leading features with speaker effects (4: the intercept and L, a, b) (≥ 1, ≤ 10)
hue_columns	integer (optional)	null	–	Hue columns around the circle (default: the largest grid_col; grid_circle only)
new_speakers	integer	20	–	Simulated new speakers per draw for the language-level shares (≥ 1, ≤ 200)
score_draws	integer	200	–	Posterior draws used by one score request (≥ 10, ≤ 4000)

Input data contract (schema template)

{
  "template": {
    "dims": {
      "speaker": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "hyperprior"
      },
      "chip": {
        "kind": "integer",
        "policy": "fixed"
      }
    },
    "vars": {
      "naming": {
        "dims": [
          "speaker",
          "chip"
        ],
        "dtype": "str",
        "role": "target",
        "missing": "allow",
        "description": "The term each speaker gave each chip ('' = no answer)"
      },
      "lab_l": {
        "dims": [
          "chip"
        ],
        "dtype": "float64",
        "role": "covariate",
        "description": "CIELAB L* of the chip"
      },
      "lab_a": {
        "dims": [
          "chip"
        ],
        "dtype": "float64",
        "role": "covariate",
        "description": "CIELAB a* of the chip"
      },
      "lab_b": {
        "dims": [
          "chip"
        ],
        "dtype": "float64",
        "role": "covariate",
        "description": "CIELAB b* of the chip"
      },
      "grid_row": {
        "dims": [
          "chip"
        ],
        "dtype": "int64",
        "role": "covariate",
        "description": "Row of the chip on the palette (WCS: A-J as 0-9)"
      },
      "grid_col": {
        "dims": [
          "chip"
        ],
        "dtype": "int64",
        "role": "covariate",
        "description": "Hue column of the chip on the palette (0 = greys)"
      }
    }
  }
}

sci.fisher_kpp

Cell counts across a closing wound (a scratch assay) as a Fisher-KPP reaction-diffusion equation: cells move with diffusivity D, divide logistically at rate lam up to a capacity K, optionally after a start-up delay. The PDE runs inside the model from the counts at the first time; outputs are the expected density field (a kymograph over time and position) and the front speed 2 sqrt(D lam).

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.profile_bands, sci.field_map
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
density	(t, x)	of count	Cells per column: the expected count K u(x, t), or a Poisson draw
wave_speed	()	um/h	Speed of the travelling invasion front, 2 sqrt(D lam)

Settings (model.settings)

Field	Type	Default	Level	Description
steps_per_hour	integer	10	–	Explicit Euler steps per hour; more steps are slower but stable for a larger diffusivity (x-ui-level: basic) (≥ 1, ≤ 1000)
delay	boolean	true	–	Multiply movement and growth by tanh(beta t): the cells start slowly after the scratch (x-ui-level: basic)
max_extension	integer	4	–	Observation periods a forecast may reach past the last fitted time (≥ 0, ≤ 1000)
diffusivity_median	number	1000.0	–	Prior median of D (um^2/h) (> 0)
diffusivity_log_sd	number	1.0	–	Log-scale sd of the prior of D (> 0)
growth_median	number	0.05	–	Prior median of lam (1/h) (> 0)
growth_log_sd	number	1.0	–	Log-scale sd of the prior of lam (> 0)
capacity_mu	number	122.0	–	Prior mean of K (cells per column) (> 0)
capacity_sigma	number	15.0	–	Prior sd of K (> 0)
delay_rate_median	number	0.1	–	Prior median of beta (1/h) (> 0)
delay_rate_log_sd	number	1.0	–	Log-scale sd of the prior of beta (> 0)
initial_floor	number	0.1	–	Smallest initial count, so an empty column can still fill (> 0)
cfl_quantile	number	0.99	–	Quantile of the prior of D that must satisfy D dt / dx^2 <= 1/2 (> 0.5, < 1.0)

Input data contract (schema template)

{
  "params": {
    "max_extension": {
      "type": "integer",
      "default": 4
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "integer",
        "policy": "extend",
        "max_extension": "{max_extension}",
        "description": "Hours since the start of the assay"
      },
      "x": {
        "kind": "float",
        "policy": "fixed",
        "description": "Centre of each column across the wound (um, equal spacing)"
      }
    },
    "vars": {
      "count": {
        "dims": [
          "t",
          "x"
        ],
        "dtype": "int64",
        "role": "target",
        "units": "cells",
        "bounds": [
          0,
          null
        ]
      }
    }
  }
}

sci.graphical_model

Learns which variables are directly linked: partial correlations (the correlation of two variables given all the others) and the probability of an edge, P(|rho| > threshold). kind dense is a multivariate normal with an LKJ prior on the correlations (it also predicts and simulates new samples); kind neighbourhood is a sparse horseshoe regression of each variable on the others with AND/OR symmetrisation.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, net.graph, sci.field_map
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
pcor	(var, var_)	1	Partial correlation of each pair of variables given all the others (unit diagonal)
edge_prob	(var, var_)	1	Edge indicator per draw (|partial correlation| > threshold); its mean over the draws is the edge probability
value	(cell, var)	of value	dense only: the values of each cell (a draw from the model with kind predictive or mode simulate)

Settings (model.settings)

Field	Type	Default	Level	Description
kind	dense or neighbourhood	dense	–	dense: one multivariate normal with an LKJ prior (predicts and simulates cells); neighbourhood: sparse horseshoe regressions of each variable on the others (x-ui-level: basic)
threshold	number	0.1	–	Relevance threshold delta: an edge is |partial correlation| > delta (x-ui-level: basic) (> 0, < 1)
standardise	boolean	true	–	Centre and scale each variable with the mean and sd of the fit data
reference	boolean	false	–	The data has a known network reference(var, var_) (1 = edge) that the plots compare with (x-ui-level: basic)
log_likelihood	auto or joint or conditional	auto	–	Pointwise log-likelihood kept for elpd_loo: joint (dense only) or conditional (each variable given the others; both kinds). auto: joint for dense, conditional for neighbourhood
lkj_eta	number	2.0	–	dense: LKJ shape of the correlation prior (> 0)
sd_rate	number	1.0	–	dense: rate of the Exponential prior of the standardised sds (> 0)
mean_sd	number	1.0	–	dense: sd of the Normal prior of the means (> 0)
expected_neighbours	number	2.0	–	neighbourhood: prior guess p0 of the neighbours of each variable (sets the global horseshoe scale) (> 0)
slab_alpha	number	2.0	–	neighbourhood: c^2 ~ InvGamma(alpha, beta) (> 0)
slab_beta	number	2.0	–	neighbourhood: c^2 ~ InvGamma(alpha, beta) (> 0)
edge_rule	and or or or rho	and	–	neighbourhood: an edge when both coefficients pass the threshold (and), either one (or), or their symmetric combination (rho)

Input data contract (schema template)

{
  "params": {
    "reference": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "cell": {
        "kind": "integer",
        "policy": "new_levels",
        "new_level_strategy": "pooled",
        "description": "Exchangeable samples (cells); new ones may be scored"
      },
      "var": {
        "kind": "categorical",
        "policy": "fixed",
        "description": "The measured variables: the nodes of the graph"
      }
    },
    "vars": {
      "value": {
        "dims": [
          "cell",
          "var"
        ],
        "dtype": "float64",
        "role": "target",
        "description": "The measured value of each variable in each cell"
      }
    }
  },
  "cases": [
    {
      "when": {
        "reference": true
      },
      "merge": {
        "dims": {
          "var_": {
            "kind": "categorical",
            "policy": "fixed",
            "description": "The variables again (the second node of a pair)"
          }
        },
        "vars": {
          "reference": {
            "dims": [
              "var",
              "var_"
            ],
            "dtype": "int64",
            "role": "constant",
            "description": "A known network to compare with (1 = edge)"
          }
        }
      }
    }
  ]
}

sci.heat_plate

Thermocouple readings of a heated plate as the 2-D steady heat equation: conduction k, face loss h, edge loss h_edge and heater efficiency eta, with a finite-volume linear system solved inside the model for each run. Outputs are the temperature field of each run (or of any new heater layout) and the heat-spreading length sqrt(k delta / 2h).

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, core.pair, sci.field_map, sci.image_residuals
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
field	(run, y, x)	of temperature	Temperature of each cell (degC): the steady field, or a simulated thermocouple reading
spread_length	()	cm	Heat-spreading length sqrt(k delta / 2h)

Settings (model.settings)

Field	Type	Default	Level	Description
thickness_mm	number	5.0	–	Thickness of the plate in mm (x-ui-level: basic) (> 0, ≤ 1000)
room_temperature	number	20.0	–	Temperature of the room in degC: the plate without heat (x-ui-level: basic) (≥ -100, ≤ 1000)
use_insulation	boolean	false	–	The data has a bool insulation(y, x): True where the back face is insulated (x-ui-level: basic)
insulated_back_loss	number	0.1	–	Back-face loss under insulation as a fraction of the bare loss (≥ 0, ≤ 1)
k_median	number	50.0	–	Prior median of the conductivity k (> 0)
k_log_sd	number	0.5	–	Log-scale sd of the prior of k (> 0)
h_median	number	10.0	–	Prior median of the face coefficient h (> 0)
h_log_sd	number	0.5	–	Log-scale sd of the prior of h (> 0)
h_edge_median	number	10.0	–	Prior median of the edge coefficient (> 0)
h_edge_log_sd	number	0.7	–	Log-scale sd of the prior of h_edge (> 0)
eta_alpha	number	40.0	–	Beta prior of the heater efficiency: alpha (> 0)
eta_beta	number	4.0	–	Beta prior of the heater efficiency: beta (> 0)
sigma_scale	number	0.5	–	Half-normal scale of the thermocouple noise sd (degC) (> 0)

Input data contract (schema template)

{
  "params": {
    "use_insulation": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "run": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "pooled",
        "description": "One run of the plate with one heater layout (score new runs to predict new layouts)"
      },
      "y": {
        "kind": "float",
        "policy": "fixed",
        "description": "Centre of each row of cells (cm, equal steps)"
      },
      "x": {
        "kind": "float",
        "policy": "fixed",
        "description": "Centre of each column of cells (cm, the same step as y)"
      }
    },
    "vars": {
      "heater_power": {
        "dims": [
          "run",
          "y",
          "x"
        ],
        "dtype": "float64",
        "role": "covariate",
        "units": "W",
        "bounds": [
          0,
          null
        ],
        "description": "Electrical power put into each cell by the heaters"
      },
      "temperature": {
        "dims": [
          "run",
          "y",
          "x"
        ],
        "dtype": "float64",
        "role": "target",
        "units": "degC",
        "missing": "allow",
        "description": "Thermocouple reading of each cell (NaN where there is none)"
      }
    }
  },
  "cases": [
    {
      "when": {
        "use_insulation": true
      },
      "merge": {
        "vars": {
          "insulation": {
            "dims": [
              "y",
              "x"
            ],
            "dtype": "bool",
            "role": "covariate",
            "description": "True where the back face of the cell is insulated"
          }
        }
      }
    }
  ]
}

sci.ibp_features

Latent binary features of small images (handwritten digits): every image is the average image plus the features it owns. A stick-breaking Indian buffet process truncated at K features, with every image's binary pattern summed out over 2^K patterns, so NUTS samples the feature images and their probabilities. Outputs are the feature images, which images own them, and completions of unseen pixels.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.field_map, sci.ppc_image_grid, sci.image_residuals
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
ink_hat	(digit, y, x)	of ink	Expected image given the seen pixels (completion where unseen), or a draw of pattern and noise
features	(feature, y, x)	ink	Feature images: the ink a feature adds to (or removes from) the average image
feature_prob	(feature)	1	Probability pi_k that an image owns feature k
owns	(digit, feature)	1	Probability that each image owns each feature, given its seen pixels

Settings (model.settings)

Field	Type	Default	Level	Description
n_features	integer	8	–	Truncation K: the largest number of features. Each image sums over 2^K patterns, so each extra feature doubles the cost (x-ui-level: basic) (≥ 1, ≤ 10)
kind	fixed_noise or free_noise	fixed_noise	–	fixed_noise keeps the pixel noise at noise_sd (the notebook); free_noise learns it (x-ui-level: basic)
noise_sd	number	0.25	–	Pixel noise sigma_X (ink units). It sets the resolution: a smaller value makes more features worth keeping (x-ui-level: basic) (> 0)
alpha_shape	number	1.0	–	Gamma shape of the IBP mass alpha (> 0)
alpha_rate	number	1.0	–	Gamma rate of the IBP mass alpha (> 0)
feature_scale_sd	number	1.0	–	HalfNormal scale of sigma_A (> 0)
start	digits or default	digits	–	digits: every chain starts at half of K random images with nu = 0.8 and no jitter (avoids the 'no features' trap); default: the jittered platform start
score_draws	integer	200	–	Posterior draws (spread over the chains) per score (≥ 1, ≤ 4000)
shown_mask	boolean	false	–	The data has a shown(digit, y, x) mask: pixels where it is False are hidden from the fit (held-out halves)
truncation_warn	number	0.05	–	Warn when the posterior mean of the last feature probability is above this (> 0, < 1)

Input data contract (schema template)

{
  "params": {
    "shown_mask": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "digit": {
        "kind": "categorical",
        "policy": "fixed",
        "description": "One small image (a digit) per level"
      },
      "y": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel row (top = 0)"
      },
      "x": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel column"
      }
    },
    "vars": {
      "ink": {
        "dims": [
          "digit",
          "y",
          "x"
        ],
        "dtype": "float64",
        "role": "target",
        "units": "ink",
        "missing": "allow",
        "description": "Ink of each pixel (0 blank, 1 full); empty = not seen"
      }
    }
  },
  "cases": [
    {
      "when": {
        "shown_mask": true
      },
      "merge": {
        "vars": {
          "shown": {
            "dims": [
              "digit",
              "y",
              "x"
            ],
            "dtype": "bool",
            "role": "mask",
            "description": "False hides the pixel from the fit (held out)"
          }
        }
      }
    }
  ]
}

sci.lgcp_grid

A log-Gaussian Cox process for event counts on a regular space-time grid: a random-walk total per period times a smooth 2-D map (Matérn-5/2 HSGP) that can change from era to era, with negative-binomial counts. Forecasts reach at most max_extension periods past the fit.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, core.ppc_fan, sci.field_map, sci.ensemble
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
rate	(period, row, col)	of count	Events per cell and period: the expected intensity, or a predictive count draw
total	(period)	of count	Expected events per period in the whole grid

Settings (model.settings)

Field	Type	Default	Level	Description
max_extension	integer	8	–	Periods a forecast may reach past the fitted range without a refit (x-ui-level: basic) (≥ 1, ≤ 400)
era_periods	integer	4	–	Periods in one era: the spatial field changes from era to era (4 quarters = one map per year) (≥ 1, ≤ 1000)
field	era_walk or static	era_walk	–	era_walk: a new map each era (a random walk of maps); static: one map for all periods (the separable model)
anchor	busiest or first or last	busiest	–	The era whose map gets the marginal prior; the other eras walk away from it (busiest = the fit era with the most events)
likelihood	negbin or poisson	negbin	–
cell_km	number	25.0	–	Width of one grid cell in km (> 0)
m_row	integer	10	–	HSGP basis functions along rows (≥ 2, ≤ 60)
m_col	integer	14	–	HSGP basis functions along columns (≥ 2, ≤ 60)
c	number	1.5	–	HSGP boundary factor (> 1.0, ≤ 5.0)
ell_lower_km	number	15.0	–	5% quantile of the lengthscale prior (> 0)
ell_upper_km	number	100.0	–	95% quantile of the lengthscale prior (> 0)
alpha_sigma	number	1.5	–	Sd of the log total per period (centred on the fit mean) (> 0)
sigma_g_scale	number	0.5	–	Half-normal scale of the RW step (> 0)
eta0_scale	number	2.0	–	Half-normal scale of the anchor map (> 0)
eta_d_scale	number	1.0	–	Half-normal scale of the era increments (> 0)
inv_phi_rate	number	1.0	–	Exponential rate of 1/phi (0 = Poisson) (> 0)

Input data contract (schema template)

{
  "params": {
    "max_extension": {
      "type": "integer",
      "default": 8
    }
  },
  "template": {
    "dims": {
      "period": {
        "kind": "time",
        "policy": "extend",
        "max_extension": "{max_extension}"
      },
      "row": {
        "kind": "integer",
        "policy": "fixed"
      },
      "col": {
        "kind": "integer",
        "policy": "fixed"
      }
    },
    "vars": {
      "count": {
        "dims": [
          "period",
          "row",
          "col"
        ],
        "dtype": "int64",
        "role": "target"
      }
    }
  }
}

sci.network_latent_space

An undirected graph (who works with whom) as one Bernoulli tie for each pair of nodes. The log odds add pair covariates (homophily), a sociality effect per node and, for kind latent_space, minus the distance between latent positions. Outputs are the tie probability of every pair, the positions (aligned by Procrustes), the sociality, simulated graphs and their statistics.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, net.graph, sci.field_map
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
tie_prob	(node, node_)	probability	Probability of a tie between the two nodes (0 on the diagonal)
tie	(node, node_)	of tie	The tie: its probability, or a simulated graph (kind predictive, mode simulate)
degree	(node)	ties	Ties of each node with the scored nodes (node_ of the score data; all nodes for the degree in the graph): expected, or in a simulated graph
sociality	(node)	log odds	How much more often the node has ties with anyone (kind sociality or latent_space)
position	(node, latent_dim)	latent units	Latent position, aligned by Procrustes (kind latent_space)
graph_stat	(statistic)	statistic	Statistics of each simulated graph of the whole node set (density, sd_degree, max_degree, isolates, triangles, transitivity, esp0, mean_geodesic, unreachable); mode simulate or kind predictive

Settings (model.settings)

Field	Type	Default	Level	Description
kind	covariates or sociality or latent_space	latent_space	–	Model of the log odds: covariates only, + a sociality effect per node, + latent positions (x-ui-level: basic)
covariates	boolean	true	–	Use the pair covariates dyad_cov(node, node_, cov); false leaves them out (x-ui-level: basic)
node_groups	boolean	false	–	The data has node_group(node), a group code kept for the plots
latent_dims	integer	2	–	Dimensions of the latent space (≥ 1, ≤ 3)
alpha_mu	number	-2.0	–	Prior mean of the intercept alpha
alpha_sigma	number	2.0	–	Prior sd of the intercept alpha (> 0)
beta_sigma	number	1.0	–	Prior sd of each covariate effect (> 0)
sociality_scale	number	1.0	–	Scale of the HalfNormal prior of sigma_s (> 0)
latent_scale	number	2.0	–	Scale of the HalfNormal prior of sigma_z (> 0)
procrustes_iters	integer	10	–	Passes of generalised Procrustes alignment (≥ 1, ≤ 100)

Input data contract (schema template)

{
  "params": {
    "covariates": {
      "type": "boolean",
      "default": true
    },
    "node_groups": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "node": {
        "kind": "integer",
        "policy": "fixed",
        "description": "The nodes (actors)"
      },
      "node_": {
        "kind": "integer",
        "policy": "fixed",
        "description": "The second node of a pair (the same labels as node)"
      }
    },
    "vars": {
      "tie": {
        "dims": [
          "node",
          "node_"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow",
        "bounds": [
          0,
          1
        ],
        "description": "1 when the two nodes have a tie, 0 when not, missing when not observed (symmetric)"
      }
    }
  },
  "cases": [
    {
      "when": {
        "covariates": true
      },
      "merge": {
        "dims": {
          "cov": {
            "kind": "categorical",
            "policy": "fixed",
            "description": "The covariates of a pair"
          }
        },
        "vars": {
          "dyad_cov": {
            "dims": [
              "node",
              "node_",
              "cov"
            ],
            "dtype": "float64",
            "description": "Covariates of each pair, for example 1 when the two nodes are in the same group"
          }
        }
      }
    },
    {
      "when": {
        "node_groups": true
      },
      "merge": {
        "vars": {
          "node_group": {
            "dims": [
              "node"
            ],
            "dtype": "int64",
            "role": "constant",
            "description": "A group code of each node, used for colours in plots"
          }
        }
      }
    }
  ]
}

sci.phylo_regression

A regression of one trait of species on other traits whose residuals are correlated by a phylogeny: i.i.d., Brownian motion, Pagel's lambda or Ornstein-Uhlenbeck on the shared branch length of the tree. Predicts a missing trait from relatives and stores the leave-one-out conditional log density for elpd_loo.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.phylogeny
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
trait	(species)	of trait	The trait of each species predicted from its relatives (Gaussian conditioning on the other species; the median when the trait is logged), or a draw with the conditional noise
slope	(predictor)	1	Slope of each predictor (sd of the trait per sd of the predictor when standardised)

Settings (model.settings)

Field	Type	Default	Level	Description
kind	iid or brownian or lambda or ou	lambda	–	Residual correlation: iid (no tree), brownian (Brownian motion on the tree), lambda (Pagel's lambda: tree part plus a species part) or ou (Ornstein-Uhlenbeck) (x-ui-level: basic)
predictors	list of string	–	–	Names of the predictor variables on the species dim (x-ui-level: basic) (max items 16)
log_target	boolean	true	–	Take the log of the trait (it must be positive)
log_predictors	boolean	true	–	Take the log of each predictor (they must be positive)
standardise	boolean	true	–	Centre and scale the trait and the predictors by the fitted species (the slopes are then in sd of the trait per sd of the predictor)
intercept_sd	number	1.0	–	Prior sd of the intercept (> 0)
slope_sd	number	0.5	–	Prior sd of each slope (> 0)
sigma_rate	number	1.0	–	Rate of the exponential prior of sigma (> 0)
lambda_alpha	number	1.0	–	Beta prior of lambda: first shape (> 0)
lambda_beta	number	1.0	–	Beta prior of lambda: second shape (> 0)
ou_parameterisation	rate or variance	rate	–	ou: put the prior on the Brownian rate eta2*rho (rate, recommended) or on eta2 and rho (variance)
rate_sd	number	1.0	–	ou (rate): HalfNormal sd of eta2*rho (> 0)
rho_sd	number	5.0	–	ou: HalfNormal sd of rho (pull to the optimum) (> 0)
eta2_sd	number	5.0	–	ou (variance): HalfNormal sd of eta2 (> 0)
jitter	number	1e-06	–	ou: added to the diagonal of K (≥ 0, ≤ 0.01)

Input data contract (schema template)

{
  "params": {
    "predictors": {
      "type": "string_list",
      "default": [
        "body",
        "group_size"
      ]
    }
  },
  "template": {
    "dims": {
      "species": {
        "kind": "categorical",
        "policy": "fixed",
        "description": "One level per species; the labels are tip labels of the tree"
      },
      "node": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Every node of the tree (tips and internal nodes)"
      }
    },
    "vars": {
      "trait": {
        "dims": [
          "species"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow",
        "description": "The trait to explain; missing values are predicted from relatives"
      },
      "{*predictors as p}": {
        "dims": [
          "species"
        ],
        "dtype": "float64",
        "role": "covariate",
        "description": "A predictor of the trait"
      },
      "parent": {
        "dims": [
          "node"
        ],
        "dtype": "int64",
        "role": "constant",
        "description": "Position of the parent node; -1 at the root"
      },
      "branch_length": {
        "dims": [
          "node"
        ],
        "dtype": "float64",
        "role": "constant",
        "description": "Length of the branch above the node (time, e.g. Myr)"
      },
      "is_tip": {
        "dims": [
          "node"
        ],
        "dtype": "bool",
        "role": "constant",
        "description": "True for a tip (a species), false for an internal node"
      },
      "tip_label": {
        "dims": [
          "node"
        ],
        "dtype": "str",
        "role": "constant",
        "description": "The species label of each tip (empty for internal nodes)"
      }
    }
  }
}

sci.storm_localisation

Single-molecule localisation microscopy (STORM/PALM): finds isolated spots in raw EMCCD frames and fits each with a pixel-integrated Gaussian PSF on a background, with Poisson counts after the camera's photon-transfer calibration. Outputs are positions with nm error bars, photons, the PSF width, a super-resolved render and the fitted patches.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, sci.point_pattern, sci.image_residuals, sci.field_map
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
x	(spot)	nm	Column position of each spot
y	(spot)	nm	Row position of each spot
photons	(spot)	photons	Photons that each molecule emitted into the frame
background	(spot)	photons/px	Background at the centre of each patch
psf_sd	(spot)	nm	PSF sd of each spot
render	(render_y, render_x)	localisations	Super-resolved image: the spots at the drawn positions, blurred by half a render pixel
patch_fit	(spot, py, px)	counts	Expected effective counts (photons / 2) of each fitted patch, or Poisson draws
frames_hat	(frame, row, col)	of frames	Camera image (ADU): the fitted spots on the background

Settings (model.settings)

Field	Type	Default	Level	Description
pixel_nm	number	100.0	–	Size of one camera pixel in the sample (nm) (x-ui-level: basic) (> 0)
kind	flat or plane	flat	–	Background in each patch: flat (one value) or plane (a value and a slope in x and y) (x-ui-level: basic)
offset_adu	number (optional)	null	–	Camera offset (ADU). Set it with scale_adu; unset, both come from the photon-transfer curve of the frames
scale_adu	number (optional)	null	–	ADU per effective count: twice the EM gain (the photon-transfer slope)
max_spots	integer	200	–	Isolated spots fitted, evenly spaced in frame order (x-ui-level: basic) (≥ 1, ≤ 5000)
threshold_sd	number	5.0	–	Detection threshold (robust noise sds) (> 0)
smooth_px	number	1.0	–	Gaussian smoothing before detection (px) (> 0)
patch_px	integer	11	–	Side of each fitted patch (odd, px) (≥ 5, ≤ 31)
min_blob_px	integer	13	–	Fewest pixels of a fitted blob (≥ 1)
max_blob_px	integer	39	–	Most pixels of a fitted blob (≥ 1)
amplitude_band	list of any	[0.0, 1.0]	–	Quantile band of blob amplitude that is fitted (0, 1: every isolated blob) (min items 2, max items 2)
photons_median	number	1500.0	–	Prior median of the photons per spot (> 0)
photons_log_sd	number	1.0	–	Log-scale prior sd of the photons (> 0)
psf_sd_px	number	1.0	–	Prior median of the PSF sd (px) (> 0)
psf_log_sd	number	0.2	–	Log-scale prior sd of the PSF sd (> 0)
position_sd_px	number	2.0	–	Prior sd of the position around the blob centre (px) (> 0)
background_photons	number (optional)	null	–	Prior median of the background (photons per pixel); unset: twice the median effective count of the movie
background_log_sd	number	0.5	–	Log-scale prior sd of the background (> 0)
gradient_sd	number	0.05	–	Prior sd of the background slopes (log per px; kind plane) (> 0)
score_draws	integer	40	–	Posterior draws in each score (over all chains) (≥ 1, ≤ 1000)
ppc_draws	integer	20	–	Predictive patch draws kept in the bundle (0: none) (≥ 0, ≤ 200)
render_px_nm	number	16.0	–	Pixel size of the super-resolved render (nm; the render of score_draws draws must stay within 50,000,000 values) (≥ 1.0)

Input data contract (schema template)

{
  "template": {
    "dims": {
      "frame": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Camera frame"
      },
      "row": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel row"
      },
      "col": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel column"
      }
    },
    "vars": {
      "frames": {
        "dims": [
          "frame",
          "row",
          "col"
        ],
        "dtype": "int32",
        "role": "target",
        "units": "ADU",
        "bounds": [
          0,
          null
        ],
        "description": "Raw camera counts of each frame"
      }
    }
  }
}

sci.strong_lens

A pixel forward model of a galaxy-scale strong gravitational lens: a singular isothermal ellipsoid with external shear traces every pixel to a Sersic source; the lens light, a PSF and correlated noise are modelled; the light amplitudes are integrated out. Gives the Einstein radius, the unlensed source and critical curves. Reads plain or tree datasets (the PSF from a /psf node).

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: the light amplitudes are integrated out over all pixels and the whitened noise couples neighbouring pixels, so pointwise leave-one-out is not defined
Default plots	core.convergence, core.posterior_forest, sci.field_map
Resources	class M (min S)

Outputs

Output	Dims	Units	Description
model_image	(row, col)	of image	Expected image (lens light, lensed source, sky, PSF), or a simulated observation with noise
source	(source_row, source_col)	of image	The unlensed source (parametric Sersic of each draw) on a source-plane grid
einstein_radius	()	arcsec	Einstein radius of the isothermal ellipsoid
einstein_mass	()	solar_mass	Mass inside the Einstein radius (needs z_lens and z_source)

Derived quantities (outputs.derived)

Name	Dims	Units	Cached	Description
critical_curves	(sample, curve_point)	arcsec	fit	Tangential critical curve (image plane), its caustic and the cut (source plane) for posterior draws, in arcsec and grid pixel indices

Settings (model.settings)

Field	Type	Default	Level	Description
pixel_scale	number	0.05	–	Pixel size in arcsec (> 0, ≤ 10)
subsample	integer	2	–	Rays per pixel along each axis (sub-pixel sampling) (≥ 1, ≤ 4)
lens_light_components	integer	2	–	Concentric Sersic profiles for the lens galaxy's light (≥ 0, ≤ 2)
external_shear	boolean	true	–	Add an external shear to the lens
noise_correlation	number	0.0	–	Lag-one correlation of the noise between neighbouring pixels (separable rows x columns; drizzled images have about 0.2). 0 treats pixels as independent (≥ 0.0, < 0.45)
fit_noise_scale	boolean	true	–	Scale the noise map by a free factor
use_mask	boolean	false	–	The data has a bool mask(row, col) of pixels
psf_tolerance	number	0.001	–	Relative error allowed in the separable PSF expansion (> 0, < 1)
psf_max_rank	integer	12	–	Most separable PSF terms (≥ 1, ≤ 40)
psf_source	auto or tree	auto	–	Where the PSF comes from: 'auto' uses the /psf node of a tree dataset when there is one, else the flat psf variable (which the input then needs); 'tree' reads it only from the /psf node, so the flat input needs no psf
theta_e_median	number	1.0	–	Prior median of the Einstein radius (> 0)
theta_e_log_sd	number	0.5	–	Prior sd of log Einstein radius (> 0)
mass_center_sd	number	0.1	–	Prior sd of the mass centre (> 0)
light_center_sd	number	0.1	–	Prior sd of the light centre (> 0)
ellipticity_sd	number	0.2	–	Prior sd of ellipticity components (> 0)
shear_sd	number	0.05	–	Prior sd of the shear components (> 0)
light_size_medians	list of any	[0.5, 1.5]	–	Prior medians of the lens-light half-light radii (min items 2, max items 2)
source_size_median	number	0.2	–	Prior median source radius (> 0)
source_offset_sd	number	0.3	–	Prior sd of the source position (> 0)
noise_scale_log_sd	number	0.2	–	Prior sd of log noise scale (> 0)
theta_e_start	number (optional)	null	–	Start the optimisation at this Einstein radius instead of the radius of the ring found in the light-subtracted image
n_starts	integer	5	–	Starts of the full-model optimisation (source positions) (≥ 1, ≤ 9)
max_iter	integer	3000	–	Iterations per optimisation (≥ 10, ≤ 100000)
source_grid	integer	30	–	Source-plane grid side (pixels) (≥ 4, ≤ 128)
source_half_width	number	0.45	–	Half-width of the source-plane grid (arcsec) (> 0)
curve_points	integer	128	–	Points per critical curve (≥ 16, ≤ 2048)
curve_draws	integer	100	–	Posterior draws with cached critical curves (≥ 1, ≤ 2000)
z_lens	number (optional)	null	–	Lens redshift (for the mass)
z_source	number (optional)	null	–	Source redshift (for the mass)
h0	number	70.0	–	Hubble constant (km/s/Mpc) (> 0)
omega_m	number	0.3	–	Matter density (flat LCDM) (> 0, < 1)

Input data contract (schema template)

{
  "params": {
    "use_mask": {
      "type": "boolean",
      "default": false
    },
    "psf_source": {
      "type": "string",
      "enum": [
        "auto",
        "tree"
      ],
      "default": "auto"
    }
  },
  "template": {
    "dims": {
      "row": {
        "kind": "integer",
        "description": "Image rows (row 0 at the bottom)"
      },
      "col": {
        "kind": "integer",
        "description": "Image columns"
      }
    },
    "vars": {
      "image": {
        "dims": [
          "row",
          "col"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow",
        "description": "Surface brightness of each pixel (for example electrons/s)"
      },
      "weight": {
        "dims": [
          "row",
          "col"
        ],
        "dtype": "float64",
        "role": "weight",
        "description": "Inverse variance of each pixel; 0 leaves the pixel out"
      }
    }
  },
  "cases": [
    {
      "when": {
        "psf_source": "auto"
      },
      "merge": {
        "dims": {
          "psf_row": {
            "kind": "integer",
            "description": "Rows of the PSF kernel"
          },
          "psf_col": {
            "kind": "integer",
            "description": "Columns of the PSF kernel"
          }
        },
        "vars": {
          "psf": {
            "dims": [
              "psf_row",
              "psf_col"
            ],
            "dtype": "float64",
            "role": "constant",
            "description": "The point-spread function (odd sides, centred)"
          }
        }
      }
    },
    {
      "when": {
        "use_mask": true
      },
      "merge": {
        "vars": {
          "mask": {
            "dims": [
              "row",
              "col"
            ],
            "dtype": "bool",
            "role": "mask",
            "description": "True where the pixel is used in the fit"
          }
        }
      }
    }
  ]
}

sci.traction_force

Traction force microscopy: recovers the 2-D traction field under a cell from the displacement field of the gel. A Gaussian prior in Fourier space (white, squared exponential or Matern spectrum); NUTS on the hyperparameters with the field integrated out, then exact posterior draws of the field.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: the likelihood is a potential over Fourier modes of the whole raster; pixels are not conditionally independent, so pointwise leave-one-out is not defined
Default plots	core.convergence, core.posterior_forest, sci.vector_field
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
traction	(component, row, col)	Pa	Joint posterior draws of the traction field
displacement_hat	(component, row, col)	of displacement	Displacement of each traction draw (expected), plus noise (predictive)
strain_energy	()	fJ	Elastic energy the cell put into the gel, inside the force mask
contractility	()	uN	Sum of the forces towards the force epicentre, inside the mask

Settings (model.settings)

Field	Type	Default	Level	Description
grid_um	number	1.0	–	Grid spacing of the displacement raster (um) (> 0)
young_pa	number	10000.0	–	Young's modulus of the gel (Pa) (> 0)
poisson	number	0.49	–	Poisson ratio of the gel (≥ 0, < 0.5)
young_cv	number	0.0	–	Relative uncertainty of Young's modulus (lognormal sd); 0 = known exactly (≥ 0, ≤ 1)
prior	white or smooth or matern	matern	–	Prior spectrum of the tractions: white (Tikhonov), smooth (squared exponential) or matern ((1 + k^2 l^2)^-alpha)
noise_um	number	0.01	–	Prior median of the noise sd (um) (> 0)
traction_pa	number	300.0	–	Prior median of the traction scale (Pa) (> 0)
lengthscale_um	number	5.0	–	Prior median of the length scale (um; smooth and matern) (> 0)
field_draws	integer	100	–	Traction fields drawn per score (over all chains) (≥ 1, ≤ 1000)
force_mask	boolean	false	–	The data has a force_mask(row, col) variable: strain energy and contractility are summed inside it (else over the whole grid)

Input data contract (schema template)

{
  "params": {
    "force_mask": {
      "type": "boolean",
      "default": false
    }
  },
  "template": {
    "dims": {
      "component": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "row": {
        "kind": "integer",
        "policy": "fixed"
      },
      "col": {
        "kind": "integer",
        "policy": "fixed"
      }
    },
    "vars": {
      "displacement": {
        "dims": [
          "component",
          "row",
          "col"
        ],
        "dtype": "float64",
        "role": "target",
        "description": "Displacement of the gel surface (x and y components)"
      }
    }
  },
  "cases": [
    {
      "when": {
        "force_mask": true
      },
      "merge": {
        "vars": {
          "force_mask": {
            "dims": [
              "row",
              "col"
            ],
            "dtype": "bool",
            "role": "mask",
            "description": "Region that strain energy and contractility cover"
          }
        }
      }
    }
  ]
}

sci.turing_pattern

Turing patterns: recovers the reaction and diffusion rates of a Schnakenberg activator-substrate system from early images of a growing pattern. The likelihood is exact for the linearised dynamics (independent complex Gaussian Fourier modes, a Woodbury identity per wavenumber). Gives the dispersion relation, the pattern wavelength, P(Turing conditions hold) and simulated patterns from a spectral solver.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict, simulate
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.convergence, core.posterior_forest, sci.field_map, sci.ppc_image_grid, sci.profile_bands
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
pattern	(t, y, x)	of activator	Activator field at the requested times: a draw given the fitted images (predict), or a new pattern from the nonlinear solver (simulate)
settled	(y, x)	of activator	Simulate only: the pattern of each run at settle_time
growth	(wavenumber)	1/time	Dispersion relation: growth rate Re lambda(k) of a mode with wavenumber k
p_turing	()	probability	1 when the Turing conditions hold for the draw (the mean is the probability)
wavelength	()	length	Wavelength 2 pi / k_m of the fastest-growing mode
turing_space	(b_supply, a_supply)	probability	1 where the supply rates (a, b) are Turing unstable at the draw's diffusion ratio (the mean is P(Turing))

Settings (model.settings)

Field	Type	Default	Level	Description
box_length	number	100.0	–	Side of the square image in the length unit of the model (the activator diffuses with D = 1 in these units) (x-ui-level: basic) (> 0)
k2_max	number	1.5	–	Use only the Fourier modes with |k|^2 below this value; faster modes have decayed below the camera noise (x-ui-level: basic) (> 0)
max_extension	integer	2	–	Image times a prediction may reach past the last fitted time (the linear theory holds only while the pattern is small) (≥ 0, ≤ 100)
d_median	number	10.0	–	Prior median of the diffusion ratio d (> 0)
d_log_sd	number	1.0	–	Log-scale sd of the prior of d (> 0)
initial_noise_median	number	0.001	–	Prior median of the initial noise sd s0 (> 0)
camera_noise_median	number	0.001	–	Prior median of the camera noise sd sigma (> 0)
field_draws	integer	40	–	Posterior draws per score (over all chains) (≥ 1, ≤ 1000)
simulate_draws	integer	4	–	Nonlinear simulations per simulate score (each one is a full PDE run) (≥ 1, ≤ 200)
simulate_dt	number (optional)	null	–	Time step of the spectral solver in simulate (default: min(0.025, 0.1 / (g (1 + s^2))))
settle_time	number	300.0	–	Time of the settled pattern in simulate (> 0, ≤ 5000.0)
wavenumber_max	number	1.2	–	Largest wavenumber of the dispersion relation output (> 0)
wavenumber_points	integer	120	–	Points of the dispersion relation output (≥ 10, ≤ 2000)
a_max	number	0.6	–	Largest supply rate a of the Turing-space map (> 0)
b_max	number	3.0	–	Largest supply rate b of the Turing-space map (> 0)
supply_points	integer	60	–	Grid points along a and b of the Turing-space map (≥ 5, ≤ 300)
precondition	boolean	true	–	Sample a parameterisation whitened at the posterior mode (Laplace approximation); the model is the same. False samples the notebook's parameters directly (needs a long warm-up, for example tune 1500)

Input data contract (schema template)

{
  "params": {
    "max_extension": {
      "type": "integer",
      "default": 2
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "float",
        "policy": "extend",
        "ordered": true,
        "max_extension": "{max_extension}",
        "description": "Time since the uniform state was perturbed"
      },
      "y": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel row"
      },
      "x": {
        "kind": "integer",
        "policy": "fixed",
        "description": "Pixel column"
      }
    },
    "vars": {
      "activator": {
        "dims": [
          "t",
          "y",
          "x"
        ],
        "dtype": "float64",
        "role": "target",
        "description": "Activator concentration in each pixel (early, small pattern)"
      }
    }
  }
}

sklearn.multiclass_classifier

Multiclass classification with an allow-listed scikit-learn, xgboost, LightGBM or CatBoost classifier: the probability of each class of a text or integer target.

Version	1.0.0
Inference	point
Representation	point
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.roc_pr, core.confusion, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(class)	1	Probability of each class (sums to 1 over class)

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	NoUncertainty	–	–
class_order	list of string (optional)	null	–
max_classes	integer	50	–	≥ 2, ≤ 200
target_dtype	str or int64	str	–

Nested settings types

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

NoUncertainty

Point predictions (representation point).

Field	Type	Default	Description
kind	none	none

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    },
    "target_dtype": {
      "type": "string",
      "enum": [
        "str",
        "int64"
      ],
      "default": "str"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "{target_dtype}",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

sklearn.tabular_bootstrap

Tabular regression with a bootstrap ensemble: one joint draw per replicate, so any aggregation is valid.

Version	1.0.0
Inference	bootstrap
Representation	draws
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.fit_vs_actual, core.residuals_over_time, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(chain, draw)	–	Bootstrap draws of the prediction over the row dims
contributions	(chain, draw, component)	–	Per-replicate feature contributions, in model space

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	BootstrapUncertainty	–	–
contributions	none or auto or exact or zeroing	auto	–

Nested settings types

BootstrapUncertainty

Bootstrap ensemble (representation draws, one draw per replicate). by resamples whole levels of a row dim (cluster bootstrap) instead of rows. expected outputs are the replicate predictions; predictive outputs add a resampled out-of-bag residual to each draw.

Field	Type	Default	Description
kind	bootstrap	bootstrap
n	integer	50	≥ 2, ≤ 1000
by	string (optional)	null

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

sklearn.tabular_classifier

Binary classification: the probability that the target is the positive class (scikit-learn, xgboost, LightGBM or CatBoost classifiers).

Version	1.0.0
Inference	point
Representation	point
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.calibration, core.roc_pr, core.confusion, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	()	1	Probability that the target equals the positive class
contributions	(component)	–	Counterfactual-zeroing contributions to the probability

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	NoUncertainty	–	–
positive_class	number (optional)	null	–
contributions	none or zeroing	none	–

Nested settings types

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

NoUncertainty

Point predictions (representation point).

Field	Type	Default	Description
kind	none	none

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

sklearn.tabular_conformal

Tabular regression with split-conformal prediction intervals (per-cell coverage 1 - alpha; intervals cannot be aggregated).

Version	1.0.0
Inference	custom
Representation	interval
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.fit_vs_actual, core.residuals_over_time, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(bound)	–	Split-conformal prediction interval (lower, upper) and point

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	ConformalUncertainty	–	–

Nested settings types

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ConformalUncertainty

Split-conformal prediction intervals (representation interval, never draws). A calibration fraction of the training rows is held out (the most recent periods of time_dim for calibration_by: last, a seeded random subset for random); the interval is point ± q where q is the ceil((n+1)(1-alpha))-th smallest absolute calibration residual. refit refits the point model on all rows afterwards (sharper, but the finite-sample guarantee then only holds approximately).

Field	Type	Default	Description
kind	conformal	conformal
alpha	number	0.1	> 0, < 1
calibration	number	0.2	≥ 0.05, ≤ 0.5
calibration_by	last or random	last
refit	boolean	false

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

sklearn.tabular_quantiles

Tabular quantile regression: native xgboost multi-quantile or one quantile model per level; quantiles are per cell and cannot be aggregated.

Version	1.0.0
Inference	custom
Representation	quantiles
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.fit_vs_actual, core.residuals_over_time, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(quantile)	–	Conditional quantiles of the target over the row dims

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	QuantileUncertainty	–	–

Nested settings types

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

QuantileUncertainty

Conditional quantiles (representation quantiles). xgboost fits one native multi-quantile model; sklearn's quantile-capable estimators fit one model per level. rearrange sorts crossing quantiles per cell (monotone rearrangement), which never makes the pinball loss worse.

Field	Type	Default	Description
kind	quantile	quantile
quantiles	list of number	–	min items 1, max items 99
rearrange	boolean	true

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

sklearn.tabular_regressor

Tabular regression with an allow-listed scikit-learn or xgboost estimator: rows are stacked from labelled dims, predictions are point estimates (with exact linear or counterfactual-zeroing contributions).

Version	1.0.0
Inference	point
Representation	point
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.fit_vs_actual, core.residuals_over_time, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	()	–	Prediction of the target over the row dims (settings.rows)
contributions	(component)	–	Per-feature contributions (exact for linear estimators, counterfactual zeroing otherwise), in model space

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	LinearRegressionChoice or RidgeChoice or LassoChoice or ElasticNetChoice or QuantileRegressorChoice or LogisticRegressionChoice or RandomForestRegressorChoice or RandomForestClassifierChoice or GradientBoostingRegressorChoice or HistGradientBoostingRegressorChoice or XGBRegressorChoice or XGBClassifierChoice or HistGradientBoostingClassifierChoice or LGBMClassifierChoice or CatBoostClassifierChoice	required	–
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	NoUncertainty	–	–
contributions	none or auto or exact or zeroing	auto	–

Nested settings types

CatBoostClassifierChoice

Field	Type	Default	Description
kind	catboost.CatBoostClassifier	required
params	CatBoostClassifierParams	{"iterations": 500, "learning_rate": null, "depth": 6, "l2_…

CatBoostClassifierParams

Field	Type	Default	Description
iterations	integer	500	≥ 1, ≤ 10000
learning_rate	number (optional)	null
depth	integer	6	≥ 1, ≤ 16
l2_leaf_reg	number	3.0	≥ 0
min_data_in_leaf	integer	1	≥ 1, ≤ 1000000
rsm	number	1.0	> 0, ≤ 1
random_strength	number	1.0	≥ 0
border_count	integer	254	≥ 1, ≤ 65535
auto_class_weights	Balanced or SqrtBalanced (optional)	null

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

ElasticNetChoice

Field	Type	Default	Description
kind	sklearn.ElasticNet	required
params	ElasticNetParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

ElasticNetParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic
l1_ratio	number	0.5	≥ 0, ≤ 1

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingClassifierChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingClassifier	required
params	HistGradientBoostingClassifierParams	{"learning_rate": 0.1, "max_iter": 100, "max_leaf_nodes": 3…

HistGradientBoostingClassifierParams

Field	Type	Default	Description
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1
class_weight	balanced (optional)	null

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

LGBMClassifierChoice

Field	Type	Default	Description
kind	lightgbm.LGBMClassifier	required
params	LGBMClassifierParams	{"n_estimators": 100, "learning_rate": 0.1, "num_leaves": 3…

LGBMClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
learning_rate	number	0.1	> 0, ≤ 1
num_leaves	integer	31	≥ 2, ≤ 131072
max_depth	integer	-1	≥ -1, ≤ 64
min_child_samples	integer	20	≥ 1, ≤ 1000000
min_child_weight	number	0.001	≥ 0
min_split_gain	number	0.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
subsample_freq	integer	0	≥ 0, ≤ 1000
colsample_bytree	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	0.0	≥ 0
max_bin	integer	255	≥ 2, ≤ 4096
class_weight	balanced (optional)	null

LassoChoice

Field	Type	Default	Description
kind	sklearn.Lasso	required
params	LassoParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "m…

LassoParams

Field	Type	Default	Description
alpha	number	1.0	> 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
max_iter	integer	1000	≥ 1, ≤ 1000000
tol	number	0.0001	> 0, ≤ 1
selection	cyclic or random	cyclic

LinearRegressionChoice

Field	Type	Default	Description
kind	sklearn.LinearRegression	required
params	LinearRegressionParams	{"fit_intercept": true, "positive": false}

LinearRegressionParams

Field	Type	Default	Description
fit_intercept	boolean	true
positive	boolean	false

LogisticRegressionChoice

Field	Type	Default	Description
kind	sklearn.LogisticRegression	required
params	LogisticRegressionParams	{"C": 1.0, "l1_ratio": 0.0, "fit_intercept": true, "solver"…

LogisticRegressionParams

Field	Type	Default	Description
C	number	1.0	> 0, ≤ 1000000000000.0
l1_ratio	number	0.0	≥ 0, ≤ 1
fit_intercept	boolean	true
solver	lbfgs or liblinear or newton-cg or newton-cholesky or sag or saga	lbfgs
max_iter	integer	100	≥ 1, ≤ 100000
tol	number	0.0001	> 0, ≤ 1
class_weight	balanced (optional)	null

NoUncertainty

Point predictions (representation point).

Field	Type	Default	Description
kind	none	none

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

RandomForestClassifierChoice

Field	Type	Default	Description
kind	sklearn.RandomForestClassifier	required
params	RandomForestClassifierParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	gini or entropy or log_loss	gini
max_features	sqrt or log2 or number (optional)	sqrt
class_weight	balanced or balanced_subsample (optional)	null

RandomForestRegressorChoice

Field	Type	Default	Description
kind	sklearn.RandomForestRegressor	required
params	RandomForestRegressorParams	{"n_estimators": 100, "max_depth": null, "min_samples_split…

RandomForestRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 5000
max_depth	integer (optional)	null
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
bootstrap	boolean	true
max_samples	number (optional)	null
min_impurity_decrease	number	0.0	≥ 0
ccp_alpha	number	0.0	≥ 0
criterion	squared_error or absolute_error or friedman_mse or poisson	squared_error
max_features	sqrt or log2 or number (optional)	1.0

RidgeChoice

Field	Type	Default	Description
kind	sklearn.Ridge	required
params	RidgeParams	{"alpha": 1.0, "fit_intercept": true, "positive": false, "s…

RidgeParams

Field	Type	Default	Description
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
positive	boolean	false
solver	auto or svd or cholesky or lsqr or sparse_cg or sag or saga or lbfgs	auto
max_iter	integer (optional)	null
tol	number	0.0001	> 0, ≤ 1

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBClassifierChoice

Field	Type	Default	Description
kind	xgboost.XGBClassifier	required
params	XGBClassifierParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBClassifierParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	binary:logistic	binary:logistic
scale_pos_weight	number	1.0	> 0

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

tabular.hierarchical_glm

A hierarchical generalised linear model in long layout: group intercepts (and optional slopes) partially pooled towards a population mean that may depend on group-level covariates. Unseen groups are drawn from the fitted hyperprior.

Version	1.1.0
Inference	mcmc
Representation	draws
Score modes	counterfactual, evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.ppc, core.convergence, core.posterior_forest
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(obs)	of y	Expected response (mean) or a predictive draw per observation

Settings (model.settings)

Field	Type	Default	Level	Description
group_dim	string	county	–	The grouping dim (index var: _idx)
family	normal or student_t or poisson or bernoulli or negative_binomial or zero_inflated_poisson or zero_inflated_negative_binomi… or beta or gamma or ordinal	normal	–	Likelihood family (its canonical link)
covariates	list of string	[]	–	Observation-level predictors over obs (max items 256)
varying_slopes	list of string	[]	–	Covariates whose slope varies by group (max items 32)
group_covariates	list of string	[]	–	Group-level predictors of the group intercept (max items 64)
centered	boolean	false	–	Centred group effects (non-centred default)
intercept_sigma	number	2.0	–	Prior sd of the population intercept (> 0)
group_sigma_scale	number	1.0	–	Half-normal scale of group sds (> 0)
beta_sigma	number	1.0	–	Prior sd of the population slopes (> 0)
noise_sigma	number	1.0	–	Half-normal scale of the noise sd (> 0)
n_categories	integer (optional)	null	–	Number of ordered categories (ordinal family only)

Input data contract (schema template)

{
  "params": {
    "group_dim": {
      "type": "string",
      "default": "county"
    },
    "covariates": {
      "type": "string_list",
      "default": []
    },
    "group_covariates": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "layout": "long",
    "dims": {
      "obs": {
        "kind": "integer",
        "policy": "extend"
      },
      "{group_dim}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "hyperprior"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "obs"
        ],
        "dtype": "float64",
        "role": "target"
      },
      "{group_dim}_idx": {
        "dims": [
          "obs"
        ],
        "dtype": "int64",
        "role": "index",
        "index_of": "{group_dim}"
      },
      "{*covariates as v}": {
        "dims": [
          "obs"
        ],
        "dtype": "float64",
        "role": "covariate"
      },
      "{*group_covariates as v}": {
        "dims": [
          "{group_dim}"
        ],
        "dtype": "float64",
        "role": "covariate"
      }
    }
  }
}

tabular.image_logistic

A Bayesian multinomial logistic regression for small images of the same size: principal component scores (or block means) of the pixels predict the class, with zero-sum normal priors on the class weights.

Version	1.1.0
Inference	mcmc
Representation	draws
Score modes	evaluate, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: the model stores no pointwise log-likelihood
Default plots	core.roc_pr, core.confusion, core.convergence
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(class)	1	Probability of each class per image (sums to 1 over class)

Settings (model.settings)

Field	Type	Default	Level	Description
pixels	string	pixels	–	The image variable (image, y, x, channel)
target	string	label	–	The class label of each image
image_dim	string	image	–	The dim that lists the images
features	pca or pixels	pca	–	pca: principal component scores; pixels: the block means
n_components	integer	16	–	PCA components (features: pca) (≥ 2, ≤ 64)
downsample	integer	1	–	Block size of the mean over y and x (≥ 1, ≤ 16)
max_features	integer	256	–	features: pixels: the most block means allowed (≥ 4, ≤ 1024)
prior_sigma	number	1.0	–	Prior sd of the class weights (> 0, ≤ 10)
class_order	list of string (optional)	null	–	The classes in order (default: sorted)
input	pixels or features	pixels	–	pixels: an image variable (image, y, x, channel); features: image features (image, feature_dim), for example from the dl.embed step
feature_dim	string	feature	–	input: features: the dim of the features

Input data contract (schema template)

{
  "params": {
    "pixels": {
      "type": "string",
      "default": "pixels"
    },
    "target": {
      "type": "string",
      "default": "label"
    },
    "image_dim": {
      "type": "string",
      "default": "image"
    },
    "input": {
      "type": "string",
      "enum": [
        "pixels",
        "features"
      ],
      "default": "pixels"
    },
    "feature_dim": {
      "type": "string",
      "default": "feature"
    }
  },
  "template": {
    "dims": {
      "{image_dim}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "pooled"
      }
    },
    "vars": {
      "{target}": {
        "dims": [
          "{image_dim}"
        ],
        "dtype": "str",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "input": "pixels"
      },
      "merge": {
        "dims": {
          "y": {
            "kind": "integer",
            "policy": "fixed"
          },
          "x": {
            "kind": "integer",
            "policy": "fixed"
          },
          "channel": {
            "kind": "categorical",
            "policy": "fixed"
          }
        },
        "vars": {
          "{pixels}": {
            "dims": [
              "{image_dim}",
              "y",
              "x",
              "channel"
            ],
            "dtype": "uint8",
            "role": "covariate"
          }
        }
      }
    },
    {
      "when": {
        "input": "features"
      },
      "merge": {
        "dims": {
          "{feature_dim}": {
            "kind": "integer",
            "policy": "fixed"
          }
        },
        "vars": {
          "{pixels}": {
            "dims": [
              "{image_dim}",
              "{feature_dim}"
            ],
            "dtype": "float32",
            "role": "covariate"
          }
        }
      }
    }
  ]
}

tabular.regressor

Tabular quantile regression for panels such as store x sku x week demand: rows are stacked from labelled dims, coordinates become categorical features, new levels are pooled and forecasts are direct. Quantiles are per cell and never aggregated.

Version	1.0.0
Inference	custom
Representation	quantiles
Score modes	counterfactual, evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: scikit-learn and xgboost estimators provide no pointwise log-likelihood
Default plots	core.fit_vs_actual, core.residuals_over_time, ml.feature_importance
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(quantile)	–	Conditional quantiles of the target over the row dims

Settings (model.settings)

Field	Type	Default	Level	Description
estimator	XGBRegressorChoice or HistGradientBoostingRegressorChoice or GradientBoostingRegressorChoice or QuantileRegressorChoice	{"kind": "xgboost.XGBRegressor", "params": {"colsample_byle…	–	A quantile-capable estimator from the allow-listed registry
target	string	y	–
rows	list of string	required	–	min items 1, max items 8
time_dim	string	time	–
time_kind	time or integer or index or none	time	–
columns	ColumnsSettings	–	–
coord_features	map of category or onehot or ordinal or numeric	–	–
new_levels	pooled or zero or error	pooled	–
pooled_max_levels	integer	256	–	≥ 1, ≤ 10000
weight	string (optional)	null	–
mask	string (optional)	null	–
forecast_strategy	direct or recursive	direct	–
retransform	median or smearing	median	–
n_jobs	integer	1	–	≥ 1, ≤ 64
uncertainty	QuantileUncertainty	{"kind": "quantile", "quantiles": [0.1, 0.5, 0.9], "rearran…	–	Quantile levels (the default gives an 80% band around the median)

Nested settings types

ColumnsSettings

Which variables become feature columns. vars (explicit list) wins over roles; exclude removes names from either. Variables with dims beyond rows are flattened into one column per label (spend[tv], promo_lag[1]).

Field	Type	Default	Description
roles	list of Role	–	max items 16
vars	list of string (optional)	null
exclude	list of string	–	max items 1024

GradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.GradientBoostingRegressor	required
params	GradientBoostingRegressorParams	{"loss": "squared_error", "alpha": 0.9, "learning_rate": 0.…

GradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or huber or quantile	squared_error
alpha	number	0.9	> 0, < 1
learning_rate	number	0.1	> 0, ≤ 10
n_estimators	integer	100	≥ 1, ≤ 10000
subsample	number	1.0	> 0, ≤ 1
min_samples_split	integer	2	≥ 2, ≤ 1000000
min_samples_leaf	integer	1	≥ 1, ≤ 1000000
max_depth	integer (optional)	3
max_features	sqrt or log2 or number (optional)	null
n_iter_no_change	integer (optional)	null
validation_fraction	number	0.1	> 0, < 1
tol	number	0.0001	> 0, ≤ 1

HistGradientBoostingRegressorChoice

Field	Type	Default	Description
kind	sklearn.HistGradientBoostingRegressor	required
params	HistGradientBoostingRegressorParams	{"loss": "squared_error", "quantile": null, "learning_rate"…

HistGradientBoostingRegressorParams

Field	Type	Default	Description
loss	squared_error or absolute_error or gamma or poisson or quantile	squared_error
quantile	number (optional)	null
learning_rate	number	0.1	> 0, ≤ 10
max_iter	integer	100	≥ 1, ≤ 10000
max_leaf_nodes	integer (optional)	31
max_depth	integer (optional)	null
min_samples_leaf	integer	20	≥ 1, ≤ 1000000
l2_regularization	number	0.0	≥ 0
max_features	number	1.0	> 0, ≤ 1
max_bins	integer	255	≥ 2, ≤ 255
early_stopping	boolean	false
validation_fraction	number	0.1	> 0, < 1
n_iter_no_change	integer	10	≥ 1, ≤ 10000
tol	number	1e-07	≥ 0, ≤ 1

QuantileRegressorChoice

Field	Type	Default	Description
kind	sklearn.QuantileRegressor	required
params	QuantileRegressorParams	{"quantile": 0.5, "alpha": 1.0, "fit_intercept": true, "sol…

QuantileRegressorParams

Field	Type	Default	Description
quantile	number	0.5	> 0, < 1
alpha	number	1.0	≥ 0, ≤ 1000000000000.0
fit_intercept	boolean	true
solver	highs or highs-ds or highs-ipm	highs

QuantileUncertainty

Conditional quantiles (representation quantiles). xgboost fits one native multi-quantile model; sklearn's quantile-capable estimators fit one model per level. rearrange sorts crossing quantiles per cell (monotone rearrangement), which never makes the pinball loss worse.

Field	Type	Default	Description
kind	quantile	quantile
quantiles	list of number	–	min items 1, max items 99
rearrange	boolean	true

Role

What a variable means to a model.

One of target, covariate, media, spend, control, index, mask, constant, weight, exposure.

XGBRegressorChoice

Field	Type	Default	Description
kind	xgboost.XGBRegressor	required
params	XGBRegressorParams	{"n_estimators": 100, "max_depth": 6, "learning_rate": 0.3,…

XGBRegressorParams

Field	Type	Default	Description
n_estimators	integer	100	≥ 1, ≤ 10000
max_depth	integer	6	≥ 0, ≤ 64
learning_rate	number	0.3	> 0, ≤ 1
min_child_weight	number	1.0	≥ 0
subsample	number	1.0	> 0, ≤ 1
colsample_bytree	number	1.0	> 0, ≤ 1
colsample_bylevel	number	1.0	> 0, ≤ 1
colsample_bynode	number	1.0	> 0, ≤ 1
reg_alpha	number	0.0	≥ 0
reg_lambda	number	1.0	≥ 0
gamma	number	0.0	≥ 0
max_bin	integer	256	≥ 2, ≤ 4096
tree_method	hist or approx or exact	hist
grow_policy	depthwise or lossguide	depthwise
max_leaves	integer	0	≥ 0, ≤ 1000000
max_cat_to_onehot	integer (optional)	null
objective	reg:squarederror or reg:squaredlogerror or reg:absoluteerror or reg:pseudohubererror or reg:quantileerror or count:poisson or reg:gamma or reg:tweedie	reg:squarederror
quantile_alpha	number (optional)	null
tweedie_variance_power	number (optional)	null
huber_slope	number (optional)	null

Input data contract (schema template)

{
  "params": {
    "target": {
      "type": "string",
      "default": "y"
    },
    "rows": {
      "type": "string_list",
      "required": true
    },
    "time_dim": {
      "type": "string",
      "default": "time"
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer",
        "index",
        "none"
      ],
      "default": "time"
    },
    "new_levels": {
      "type": "string",
      "enum": [
        "pooled",
        "zero",
        "error"
      ],
      "default": "pooled"
    }
  },
  "template": {
    "dims": {
      "{*rows as d}": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "{new_levels}"
      }
    },
    "vars": {
      "{target}": {
        "dims": "{rows}",
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "time"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "time",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "integer",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    },
    {
      "when": {
        "time_kind": "index"
      },
      "merge": {
        "dims": {
          "{time_dim}": {
            "kind": "index",
            "policy": "extend",
            "new_level_strategy": "error"
          }
        }
      }
    }
  ]
}

ts.ardl

statsmodels ARDL: the result regressed on its own lags and current and past values of drivers by least squares; optional lag selection by information criterion; one model per series of a panel; forecasts need future drivers.

Version	1.0.0
Inference	point
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	required	basic	Driver variables (at least one); their future values are needed for forecasts (min items 1, max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
lags	integer	1	–	Past values of the result to use (≥ 0, ≤ 12)
order	integer	1	–	Past values of each driver to use (0 = same period only) (≥ 0, ≤ 12)
trend	n or c or t or ct	c	–	n none, c constant, t trend, ct both
auto	boolean	false	–	Choose lags and driver orders per series by an information criterion (up to lags/order)
ic	aic or bic or hqic	aic	–	Criterion for auto
causal	boolean	false	advanced	Use only past values of the drivers (no same-period effect)

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.arima

statsmodels ARIMA(p,d,q)(P,D,Q)s without drivers (state-space form, exact likelihood); one model per series of a panel; prediction intervals as quantiles.

Version	1.0.0
Inference	kalman
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
maxiter	integer	200	advanced	Optimiser iterations (≥ 10, ≤ 5000)
method	lbfgs or bfgs or powell or nm	lbfgs	advanced	Likelihood optimiser
order	list of any	[1, 0, 0]	basic	(p, d, q): past values used (AR), times differenced (d), past errors used (MA) (min items 3, max items 3)
seasonal_order	list of any	[0, 0, 0, 0]	basic	(P, D, Q, s): the same for the seasonal pattern repeating every s periods (min items 4, max items 4)
trend	auto or n or c or t or ct	auto	–	Deterministic trend: n none, c constant, t linear, ct both; auto uses a constant when the series is not differenced

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.bsts

A Bayesian structural time series: a trend, seasonal patterns and drivers, fitted with the Kalman filter of pymc-extras. Forecasts and filters start from the stored terminal state.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, filter, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	scaling
LOO	invalid: the Kalman likelihood is a sum over the whole series; steps depend on each other, so pointwise leave-one-out is not defined
Default plots	core.ppc_fan, core.decomposition, core.convergence, core.posterior_forest
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Expected value or predictive draw: one-step-ahead inside the fitted range, a forecast after it
level	(t)	of y	The trend and level component (expected)

Settings (model.settings)

Field	Type	Default	Level	Description
trend	level or local_linear or smooth	local_linear	basic	level: a level that wanders; local_linear: a level and a slope that both wander; smooth: a slope that wanders slowly (a smooth trend)
seasonality	list of SeasonSettings	–	basic	Seasonal patterns (at most three) (max items 3)
level_sigma	number	0.1	–	Half-normal scale of level changes (> 0, ≤ 5)
slope_sigma	number	0.02	–	Half-normal scale of slope changes (> 0, ≤ 5)
obs_sigma	number	0.5	–	Half-normal scale of the noise (> 0, ≤ 10)
initial_sd	number	1.0	–	Prior sd of the initial states and seasonal effects (> 0, ≤ 10)
regression_prior	normal or horseshoe	normal	–	normal: every driver has an effect; horseshoe: most drivers have none
regression_sigma	number	1.0	–	Prior scale of the driver effects (standardised) (> 0, ≤ 10)
regressors	list of string	–	basic	Driver variables (prices, promotions) whose future values are supplied for forecasts (max items 32)

Nested settings types

SeasonSettings

One seasonal pattern.

Field	Type	Default	Description
period	number	required	Length of the cycle in periods of the time dim (12 for monthly data with a yearly cycle, 52.18 for weekly data, 7 for daily data) (> 1, ≤ 1000)
harmonics	integer (optional)	null	Number of sine/cosine pairs. Empty: one effect per season (needs a whole-number period of at most 24)
stochastic	boolean	false	Let the seasonal pattern change over time
sigma	number	0.05	Half-normal scale of the seasonal changes (standardised) (> 0, ≤ 5)

Input data contract (schema template)

{
  "params": {
    "regressors": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "{*regressors as x}": {
        "dims": [
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  }
}

ts.dynamic_factor

statsmodels DynamicFactor: a few latent factors following a VAR drive many series; estimated by maximum likelihood in state-space form; stores the smoothed factors. Standardise the series first (core.standardize).

Version	1.0.0
Inference	kalman
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
maxiter	integer	200	advanced	Optimiser iterations (≥ 10, ≤ 5000)
method	lbfgs or bfgs or powell or nm	lbfgs	advanced	Likelihood optimiser
k_factors	integer	1	–	Number of hidden common factors (≥ 1, ≤ 5)
factor_order	integer	1	–	Lags in the factors' own dynamics (≥ 0, ≤ 4)
error_order	integer	0	–	Lags in each series' own noise (≥ 0, ≤ 2)

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "variable": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "variable",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.ecm

Single-equation error-correction model: an ARDL in levels reparameterised as an unrestricted ECM (statsmodels UECM) with long-run coefficients, the speed of adjustment and the Pesaran-Shin-Smith bounds test; forecasts need future drivers.

Version	1.0.0
Inference	point
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	required	basic	Driver variables (at least one); their future values are needed for forecasts (min items 1, max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
lags	integer	1	–	Past values of the result in the levels ARDL (≥ 1, ≤ 12)
order	integer	1	–	Past values of each driver in the levels ARDL (≥ 1, ≤ 12)
trend	n or c or ct	c	–	n none, c constant (bounds case 3), ct constant and trend (case 5)
bounds_case	1 or 2 or 3 or 4 or 5 (optional)	null	advanced	Pesaran-Shin-Smith case of the bounds test (default from the trend)
bounds_alpha	number	0.05	–	Significance of the bounds test (> 0, < 0.5)

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.ets

statsmodels ETSModel: exponential smoothing with additive or multiplicative error, trend (optionally damped) and season, estimated by maximum likelihood; one model per series of a panel; closed-form or simulated prediction intervals as quantiles.

Version	1.0.0
Inference	custom
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time, core.decomposition
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
error	add or mul	add	–	Error: additive (add) or multiplicative (mul)
trend	add or mul (optional)	null	–	Trend: none, additive or multiplicative
damped_trend	boolean	false	–	Let the trend flatten out over the horizon
seasonal	add or mul (optional)	null	–	Season: none, additive or multiplicative
seasonal_periods	integer (optional)	null	–	Season length in periods (e.g. 12 for monthly data)
simulate_repetitions	integer	1000	advanced	Simulated paths for intervals of multiplicative models (≥ 100, ≤ 20000)
maxiter	integer	1000	advanced	≥ 10, ≤ 10000

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.gaussian_hmm

A K-state Gaussian hidden Markov model over padded sequences (e.g. animal tracks). Hidden states are marginalised with the forward algorithm; forecasts and filters start from the stored terminal state.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, filter, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: the HMM likelihood is a potential over whole sequences; steps within a sequence are dependent, so pointwise leave-one-out is not defined
Default plots	core.convergence, core.posterior_forest, core.state_sequence
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(animal, step)	of y	One-step-ahead expected value or predictive draw of y
state_prob	(animal, step, state)	probability	Hidden-state probabilities (predicted, or filtered in filter mode)

Settings (model.settings)

Field	Type	Default	Level	Description
n_states	integer	2	–	Number of hidden states K (≥ 2, ≤ 8)
mu_center	number	0.0	–	Prior mean of the state means
mu_sigma	number	2.0	–	Prior sd of the state means (> 0)
sigma_scale	number	1.0	–	Half-normal scale of the emission sds (> 0)
stickiness	number	2.0	–	Extra Dirichlet weight on staying in the same state (≥ 0, ≤ 1000)
pi_concentration	number	1.0	–	Dirichlet weight of pi (> 0)
identification	relabel or ordered	relabel	–	relabel: sort states by mean after sampling; ordered: ordered prior

Input data contract (schema template)

{
  "template": {
    "dims": {
      "animal": {
        "kind": "categorical",
        "policy": "new_levels",
        "new_level_strategy": "pooled"
      },
      "step": {
        "kind": "integer",
        "policy": "extend"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "animal",
          "step"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "mask": {
        "dims": [
          "animal",
          "step"
        ],
        "dtype": "bool",
        "role": "mask"
      }
    },
    "ragged": [
      {
        "group_dim": "animal",
        "within_dim": "step",
        "form": "padded",
        "mask_var": "mask"
      }
    ]
  }
}

ts.hsgp_trend

A smooth Hilbert-space GP trend with optional linear drift and Fourier seasonality for one regular time series (Mauna-Loa style). Forecasts reach at most max_extension periods past the fit; further horizons need a refit.

Version	1.0.0
Inference	mcmc
Representation	draws
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	valid
Default plots	core.ppc_fan, core.convergence, core.posterior_forest, core.decomposition
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	The series: expected value (trend + seasonality) or a predictive draw
trend	(t)	of y	The long-term component (intercept + drift + GP), no seasonality

Settings (model.settings)

Field	Type	Default	Level	Description
max_extension	integer	24	–	Periods a forecast may reach past the fitted range without a refit; sizes the HSGP domain (x-ui-level: basic) (≥ 1, ≤ 10000)
kernel	expquad or matern52 or matern32	expquad	–	Covariance kernel of the smooth trend
m	integer	30	–	Number of HSGP basis functions (≥ 4, ≤ 400)
c	number	1.5	–	Boundary factor (L = c on the scaled axis) (> 1.0, ≤ 5.0)
lengthscale_mu	number	0.4	–	Median of the lengthscale prior on the scaled axis [-1, 1] (> 0)
lengthscale_sigma	number	0.5	–	Log-scale sd of the lengthscale (> 0)
eta_sigma	number	1.0	–	Half-normal scale of the GP amplitude (> 0)
linear_trend	boolean	true	–
intercept_sigma	number	1.0	–	> 0
slope_sigma	number	1.0	–	> 0
season_period	number (optional)	12.0	–	Seasonal period in periods of t (12 for monthly data)
season_order	integer	3	–	Fourier order of the seasonality (≥ 0, ≤ 20)
season_sigma	number	0.5	–	> 0
noise_sigma	number	0.5	–	> 0

Input data contract (schema template)

{
  "params": {
    "max_extension": {
      "type": "integer",
      "default": 24
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend",
        "max_extension": "{max_extension}"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      }
    }
  }
}

ts.markov_switching

statsmodels Markov switching: a regression or autoregression whose level, driver effects and noise switch between hidden regimes (Hamilton filter, maximum likelihood); stores regime probabilities and durations; forecasts are regime mixtures (simulated for autoregressions).

Version	1.0.0
Inference	custom
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time, core.state_sequence
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
k_regimes	integer	2	–	Number of regimes (≥ 2, ≤ 4)
order	integer	0	–	Autoregressive order (0 = no AR terms) (≥ 0, ≤ 4)
trend	n or c or t or ct	c	–	n, c (level), t or ct
switching_trend	boolean	true	–	Each regime has its own level/trend
switching_exog	boolean	true	–	Each regime has its own driver effects
switching_variance	boolean	true	–	Each regime has its own noise level
switching_ar	boolean	true	–	Each regime has its own AR coefficients
search_reps	integer	0	advanced	Random restarts to avoid poor local optima (seeded) (≥ 0, ≤ 50)
simulations	integer	2000	advanced	Simulated paths for forecasts of autoregressive models (≥ 200, ≤ 20000)
maxiter	integer	100	advanced	≥ 10, ≤ 2000

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.sarimax

statsmodels SARIMAX: ARIMA(p,d,q)(P,D,Q)s with drivers, estimated by exact maximum likelihood with the Kalman filter; one model per series of a panel; prediction intervals as quantiles; future drivers are required for forecasts.

Version	1.0.0
Inference	kalman
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
maxiter	integer	200	advanced	Optimiser iterations (≥ 10, ≤ 5000)
method	lbfgs or bfgs or powell or nm	lbfgs	advanced	Likelihood optimiser
order	list of any	[1, 0, 0]	basic	(p, d, q): past values used (AR), times differenced (d), past errors used (MA) (min items 3, max items 3)
seasonal_order	list of any	[0, 0, 0, 0]	basic	(P, D, Q, s): the same for the seasonal pattern repeating every s periods (min items 4, max items 4)
trend	auto or n or c or t or ct	auto	–	Deterministic trend: n none, c constant, t linear, ct both; auto uses a constant when the series is not differenced
enforce_stationarity	boolean	true	advanced
enforce_invertibility	boolean	true	advanced

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.theta

statsmodels ThetaModel (Assimakopoulos-Nikolopoulos theta method, equivalent to simple exponential smoothing with drift) with optional seasonal adjustment; a strong, fast baseline; one model per series of a panel.

Version	1.0.0
Inference	custom
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
period	integer (optional)	null	–	Season length; leave empty for no seasonal adjustment
deseasonalize	boolean	true	–	Remove the season before forecasting
use_test	boolean	true	advanced	Only remove the season when a test finds one
method	auto or additive or multiplicative	auto	advanced	How the season is removed
theta	number	2.0	–	Weight of the trend line (2 = classic) (≥ 1.0, ≤ 10.0)
use_mle	boolean	false	advanced	Estimate by maximum likelihood (ARIMA(0,1,1) with drift) instead of OLS + SES

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.unobserved_components

statsmodels UnobservedComponents: a structural model with a level/trend, optional stochastic seasonal, cycle, AR noise and drivers, estimated by maximum likelihood; smoothed components are stored with the fit.

Version	1.0.0
Inference	kalman
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time, core.decomposition
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
maxiter	integer	200	advanced	Optimiser iterations (≥ 10, ≤ 5000)
method	lbfgs or bfgs or powell or nm	lbfgs	advanced	Likelihood optimiser
level	irregular or fixed intercept or deterministic constant or local level or random walk or fixed slope or deterministic trend or local linear deterministic tr… or random walk with drift or local linear trend or smooth trend or random trend	local linear trend	basic	How the level and trend move: 'local level' (level wanders), 'local linear trend' (level and slope wander), 'smooth trend', 'deterministic trend', ...
seasonal	integer (optional)	null	–	Season length in periods (e.g. 52 for weekly data)
stochastic_seasonal	boolean	true	–	Let the seasonal pattern change slowly
cycle	boolean	false	–	Add a cycle (business-cycle like swings)
stochastic_cycle	boolean	false	advanced
damped_cycle	boolean	false	advanced
autoregressive	integer (optional)	null	–	Order of an AR term for the remaining noise

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "allow"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.var

statsmodels VAR(p) by least squares with the lag order chosen by AIC/BIC/HQIC/FPE; joint forecasts with intervals, orthogonalised impulse responses with 95% bands, forecast-error variance decomposition and pairwise Granger causality tests.

Version	1.0.0
Inference	point
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
irf_horizon	integer	12	–	Periods of the impulse responses and the variance decomposition (≥ 1, ≤ 60)
lags	integer (optional)	null	–	Lag order; leave empty to choose it by the criterion
max_lags	integer	8	–	Largest lag order tried when choosing (≥ 1, ≤ 24)
ic	aic or bic or hqic or fpe	aic	–	Criterion for the lag order
trend	n or c or ct or ctt	c	–	n none, c constant, ct + trend

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "variable": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "variable",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.varmax

statsmodels VARMAX(p, q) with drivers (VARX when q = 0) estimated by maximum likelihood in state-space form; joint forecasts need future drivers; stores orthogonalised impulse responses and the variance decomposition.

Version	1.0.0
Inference	kalman
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
irf_horizon	integer	12	–	Periods of the impulse responses and the variance decomposition (≥ 1, ≤ 60)
maxiter	integer	200	advanced	Optimiser iterations (≥ 10, ≤ 5000)
method	lbfgs or bfgs or powell or nm	lbfgs	advanced	Likelihood optimiser
order	list of any	[1, 0]	–	(p, q): VAR lags and MA lags (min items 2, max items 2)
trend	n or c or t or ct	c	–	n, c, t or ct
measurement_error	boolean	false	advanced
enforce_stationarity	boolean	true	advanced
enforce_invertibility	boolean	true	advanced

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "variable": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "variable",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}

ts.vecm

statsmodels VECM (Johansen maximum likelihood) for trending series tied together in the long run; the cointegration rank comes from the Johansen trace or max-eigenvalue test unless set; stores alpha/beta, the rank test table, impulse responses, FEVD and Granger tests.

Version	1.0.0
Inference	point
Representation	quantiles
Score modes	evaluate, forecast, predict
Contributions	none
Decision objective	–
Fast scoring	–
Owns	–
LOO	invalid: statsmodels estimates are maximum-likelihood or least-squares point estimates without posterior draws; use holdout or backtest metrics instead of LOO
Default plots	core.ppc_fan, core.fit_vs_actual, core.residuals_over_time
Resources	class S (min S)

Outputs

Output	Dims	Units	Description
y_hat	(t)	of y	Prediction-interval quantiles of y: one-step-ahead inside the fitted range, a multi-step forecast after it

Settings (model.settings)

Field	Type	Default	Level	Description
exog	list of string	–	basic	Driver variables (prices, promotions, holidays) whose future values are supplied for forecasts (max items 64)
series_dims	list of string	–	basic	Panel dims (geo, product, ...): one model per series cell. Empty for a single series. (max items 4)
quantiles	list of number	–	advanced	Quantiles of the prediction intervals (must include 0.05, 0.5 and 0.95) (min items 3, max items 16)
time_kind	time or integer	time	advanced	Whether the time index holds dates ('time') or period numbers ('integer')
irf_horizon	integer	12	–	Periods of the impulse responses and the variance decomposition (≥ 1, ≤ 60)
k_ar_diff	integer	1	–	Lagged differences in each equation (≥ 0, ≤ 12)
coint_rank	integer (optional)	null	–	Number of long-run relationships; leave empty to take it from the Johansen test
rank_test	trace or maxeig	trace	–	Johansen test statistic
rank_signif	number	0.05	–	Significance of the rank test: 0.01, 0.05 or 0.1
deterministic	n or co or ci or lo or li or cili or colo	ci	–	Constant/trend: n none, co outside / ci inside the long-run relation, lo/li linear trend outside/inside, combinations

Input data contract (schema template)

{
  "params": {
    "series_dims": {
      "type": "string_list",
      "default": []
    },
    "time_kind": {
      "type": "string",
      "enum": [
        "time",
        "integer"
      ],
      "default": "time"
    },
    "exog": {
      "type": "string_list",
      "default": []
    }
  },
  "template": {
    "dims": {
      "t": {
        "kind": "time",
        "policy": "extend"
      },
      "{*series_dims as d}": {
        "kind": "categorical",
        "policy": "fixed"
      },
      "variable": {
        "kind": "categorical",
        "policy": "fixed"
      }
    },
    "vars": {
      "y": {
        "dims": [
          "{*series_dims}",
          "variable",
          "t"
        ],
        "dtype": "float64",
        "role": "target",
        "missing": "forbid"
      },
      "{*exog as x}": {
        "dims": [
          "{*series_dims}",
          "t"
        ],
        "dtype": "float64",
        "role": "covariate",
        "future_required": true
      }
    }
  },
  "cases": [
    {
      "when": {
        "time_kind": "integer"
      },
      "merge": {
        "dims": {
          "t": {
            "kind": "integer"
          }
        }
      }
    }
  ]
}