Real-Time Recession Probability & GDP Nowcasting

A vintage-honest recession forecasting pipeline built on ALFRED, the Fed's real-time data archive. It estimates the probability of a US recession at the nowcast, 3-, 6-, and 12-month horizons, using only data that was actually published as of each forecast date.

Approach

Three models cover the horizon spectrum. An Estrella–Mishkin probit runs on the yield curve and credit spreads. A mixed-frequency dynamic factor model runs on real-activity data (payrolls, industrial production, income, sales, claims) and also produces a full predictive density for current-quarter GDP. A gradient-boosted classifier runs on the combined feature set.

The discipline throughout is vintage honesty. Every forecast in the backtest sees only the data available on that date: first-print GDP, unrevised payrolls, and NBER recession labels lagged to their actual announcement dates. Backtests on revised data flatter every model, since revisions contain information that did not exist in real time.

Results and caveats

The backtest covers 27 years: 330 monthly forecasts spanning three recessions. Results reproduce the literature honestly. The factor model leads at 3–12 months. The simple credit-spread probit is nearly unbeatable at 12 months. The ML model wins only at the nowcast horizon, and for a structural reason: with roughly three recessions of training data, boosted trees have little to learn about year-ahead turning points.