Banbura et al 2013 - Now-casting and the Real-Time Data Flow - Overview
Summary
Bańbura, Giannone, Modugno & Reichlin (2013, ECB Working Paper 1564; Handbook of Economic Forecasting vol. 2) review now-casting: estimating the present value of a slowly published, low-frequency target (such as GDP) from a ragged, mixed-frequency stream of releases. Their preferred framework is a joint state-space (dynamic factor) model fitted at high frequency, in which mixed frequency and ragged edges are missing-data problems solved by the Kalman filter and smoother, and in which each release produces model-based “news” that revises the now-cast.
Overview
Problem. Key statistics (GDP) are published with long lags, so the present must be inferred from earlier-released, higher-frequency indicators with varying publication delays (the “ragged” or “jagged” edge), possibly many series.
Contributions. (1) Notation for vintages and temporal aggregation; (2) joint state-space approaches (factor model, mixed-frequency VAR, daily model with flow variables); (3) the news decomposition of now-cast revisions; (4) partial models (bridge, MIDAS) for comparison; (5) an empirical application to US GDP with 24 series in real-time ALFRED vintages, finding that surveys matter, accuracy improves as information accumulates, and daily financial data add little.
Motivating PyMC examples
E98 (“Mixed-frequency sensor fusion on real-time data vintages”: monthly latent AR(1), Mariano-Murasawa aggregation, correlated GDP-GDI errors, ragged edge, revision noise), E96 (sensor fusion for a snow pillow), E97 (sea-level rise from three instruments), E99 (range-only localisation). The sensor-based ones are in Banbura 2013 - Sensor Fusion in State-Space Form (Concept Note).
Source Map
| Paper section | Covered in |
|---|---|
| Abstract, Introduction | this note |
| 2 Problem and overview; notation; 2.1 Temporal aggregation | Banbura 2013 - Mixed Frequency State Space and Mariano-Murasawa Aggregation |
| 2.2 Joint models in state space: factor model, estimation (PC, two-step, EM), daily model with flows, mixed-frequency VAR | Banbura 2013 - Factor Models, Estimation and the Ragged Edge |
| 2.3 Now-cast updates and news | Banbura 2013 - News Decomposition and Real-Time Vintages |
| 2.4 Partial models: bridge equations, MIDAS | Banbura 2013 - Partial Models and the Empirical GDP Application |
| 3 Empirical application: data (Table 1), real-time design, benchmark, uncertainty, news, financial variables | Banbura 2013 - News Decomposition and Real-Time Vintages; Banbura 2013 - Partial Models and the Empirical GDP Application |
| 4 Conclusions | Banbura 2013 - Partial Models and the Empirical GDP Application |
Connections
Direct application of Linear-Gaussian State-Space Models, The Kalman Filter (missing-observation updates), Marginal Likelihood via the Kalman Filter (ML/EM estimation) and The RTS Smoother (smoothed factors and uncertainty).
Relevance to marketing measurement
Weekly sales/spend with monthly or quarterly outcome series (survey tracker, finance data) is the same mixed-frequency problem. The news idea answers “why did our latest-weeks estimate of lift move after the finance data arrived?“. Ragged edges (offline channels arriving later than digital) are handled the same way.