Recession Nowcasting: Real-Time Probability Bands for the US and Canada
Calibrated real-time recession probabilities built only from data available at each historical date — with uncertainty bands, a vintage-vs-final refutation, and a weekly-updated US & Canada dashboard.
The Question
How likely is a US or Canadian recession right now, if you can only use data that was actually published at this moment?
That constraint — no look-ahead, no hindsight revisions — makes the problem much harder than it looks. A model that backfills revised data can appear to have predicted every recession. One built on genuine vintage data often cannot.
What This Is
A mixed-frequency dynamic factor model (statsmodels DynamicFactorMQ, Kalman filter / EM algorithm) evaluated in a pseudo-real-time expanding window: every backtest decision at month τ uses only the data that had been released by τ, approximated by per-series publication lags. The vintage-vs-final refutation measures how much apparent skill disappears when you remove hindsight.
Three comparators run alongside the DFM in the same backtest harness:
- Term-spread probit — the classic recession leading indicator (T10Y3M shifted 12 months)
- Sahm rule — the real-time labor-market trigger
- Penalised MIDAS — elastic-net logistic on multi-lag indicators
All models are calibrated (no class_weight='balanced'; ECEs ≤ 0.10) and evaluated on Brier score, AUC, PR-AUC, and reliability curves, with Diebold–Mariano / Giacomini–White tests for predictive ability and block-bootstrap confidence intervals.
Key Finding
The DFM scores AUC 0.963 on full-information data but only 0.742 in real-time — a 22-point drop that comes entirely from the ragged publication edge and the absence of data revisions. The term-spread probit (AUC 0.826) outperforms the DFM in real-time at h=0, which is the plan’s stated falsifiable outcome. Gradient boosting achieves the best real-time AUC (0.884); penalised MIDAS the best Brier score (0.048).
The output is a probability band (5th–95th percentile envelope over 500 MC draws that jointly perturb probit coefficients and factor-smoother uncertainty), not a single number.
Current Call (June 2026)
US recession probability: ~6% [2.3–13.9%]. Canada: ~5%.
Updated automatically each Monday at 11:00 UTC via GitHub Actions.