Transform
Returns, differences and stationarity checks by asset class.
CASE 07 · ECONOMETRICS
GDP Forecasting with MIDAS
Mixed-frequency US GDP forecasting.
PREDICTOR UNIVERSE47 SERIES
Daily multi-asset and macro-financial data.
FULL COVERAGE1986–2025
Replication plus post-COVID extension.
DAILY BLOCK63 DAYS
Approximately one quarter of information.
RECENT OOS8 QUARTERS
A short, difficult 2024–2025 window.
RESEARCH DESIGN
Respect the information calendar.
The target is quarterly while predictors arrive daily or monthly. MIDAS preserves high-frequency ordering with a parsimonious lag function.
Compress
Individual predictors or PCA factors with mixed-frequency timing.
Estimate
ADL–MIDAS with exponential Almon weights and AIC lag selection.
Forecast
Recursive expanding-window evaluation without future information.
Compare
RMSFE against Random Walk, AR, ADL and factor benchmarks.
DATA & EVIDENCE
Diagnostics before model ranking.
Coverage, transformations, recursive weights and period-by-period errors remain visible alongside aggregate forecast scores.
Unequal histories and completeness thresholds are audited before factor extraction.
Forecast paths are compared with realized GDP under one recursive protocol.
The estimated decay reveals how much information is assigned across 63 trading days.
The extension removes the restriction that historical lags and current-quarter leads share one decay parameter.
A short recent sample can make model rankings depend on very few quarters.
CORE FINDING
Frequency helps under the right specification.
High-frequency financial information can improve the historical forecast exercise, but model choice, leads and the information block matter more than simply increasing complexity.
LIMIT
The recent ranking is fragile.
Eight out-of-sample quarters are insufficient for strong general claims. The Two-β extension fixes one weight restriction, not every nowcasting challenge.