Second Brain

About

A structured knowledge base covering Bayesian statistics, econometrics, causal inference, and agent-based modeling. Notes are cross-linked by topic — use the graph view or search bar to explore connections across the collection.

For AI assistants

Reading this site with a language model? Start at llms.txt — a routed index of every note. Any page is also available as raw markdown by appending .md to its URL.

Explore the 3D knowledge graph →

Selected Analyses

ratio estimation?

September 18, 2026 · Agent Based Modeling · Likelihood Free Inference · Simulation Estimation · Calibration · Bayesian Statistics

The eleven methods differ on only three design choices: what is compared (hand-picked moments, an auxiliary model’s parameters, or learned summaries), how the comparison is scored (a quadratic distance to minimise, a threshold to pass, a Gaussian or neural density to evaluate) and whether the simulator is called inside the search loop or once up front.


In-context learning has been read as amortized (implicit) Bayesian inference. How does that reading compare with neural posterior estimation, variational autoencoders’ amortized encoders, deep adaptive design and hierarchical models — and where does the analogy break?

September 18, 2026 · Large Language Models · Likelihood Free Inference · Variational Inference · Bayesian Statistics · Calibration

The analogy is sound at the level of computational economics: NPE, the VAE encoder, deep adaptive design and in-context learning (ICL) all pay a large training cost once, over a distribution of datasets, so that adapting to a new dataset is a single forward pass with no gradient step.


deep adaptive design, dynamic treatment regimes with Q- and A-learning, switchback experiments and RLHF all choose actions from accumulating data. Laid out on one map, what is each optimizing, what is the state, and what feedback does it assume?

September 18, 2026 · Multi Armed Bandits · Bayesian Experimental Design · Causal Inference · Online Experimentation · Large Language Models · Machine Learning

All of these methods run the same loop (model the history, act, observe, update) and differ on four things: the objective (cumulative regret, final value, information, policy value, or a valid estimate), whether the state is only the learner’s belief or a physical state that actions change, whether data are gathered online by the algorithm or logged offline by someone else, and whether the reward is observed, delayed, or itself estimated.


VNMC bounds on expected information gain, the contrastive PCE and ACE bounds, neural ratio estimation and the forward-KL objective of neural posterior estimation relate? Which direction of KL does each use, is each an upper or lower bound, and what failure does that choice cause?

September 18, 2026 · Variational Inference · Bayesian Experimental Design · Likelihood Free Inference · Bayesian Statistics

Every one of these objectives is the same identity, intractable quantity = computable surrogate ± an expected KL, obtained by replacing an intractable density with a learned .


mROAS optimization if the outcome is customer lifetime value rather than sales, given that CLV is a model-based forecast observed with delay and censoring?

September 18, 2026 · Customer Lifetime Value · Market Response · Causal Inference · Forecasting · Uncertainty Quantification

The optimization keeps its shape (maximize posterior-expected value under a budget, equalize marginal returns) but the numerator stops being an observed quantity.


Knowledge Base

DomainCore Topics
Bayesian StatisticsInference fundamentals, hierarchical models, MCMC, model checking
EconometricsIdentification strategies, regression foundations, simulation-based estimation
Agent-Based ModelingCalibration methods, social dynamics, consumer behavior
Research MethodologyExperimental design, multiple comparisons, causal reasoning

Browse the full collection via the search bar or explore topic connections in the graph view.