US Recession Risk

What you see: the four key inputs standardized as z-scores for visual comparison.

What this is: The reliability gate outputs w, the weight placed on the macro-finance model when forming the hybrid probability.

How to read: When w is near 0, the system is mostly trusting the yield curve. When w is near 1, it is mostly trusting credit and labor-market stress signals.

Why it helps: The two models can genuinely disagree — the curve can invert while broader stress stays low, or stress can build while the curve still looks benign. The gate learns, out-of-sample, which read to weight more heavily.

Two models: The baseline uses the yield-curve spread as its only input — the classic Estrella-style recession probit. The augmented model is an additive macro-finance specification: the same spread plus five broader stress measures.

Why additive — and why not interaction terms: An inverted curve plausibly means different things depending on context, so we tested an explicit interaction version — letting the spread's effect bend with credit stress, financial conditions, and labor markets (SPREAD × CREDIT_Z, and so on). In a controlled out-of-sample validation those interaction terms added no distinguishable value: the elastic-net shrank them toward zero and kept the plain spread as the dominant signal, and the variant that dropped the standalone spread was actually less well calibrated. The deployed model therefore stays additive.

What this buys: The additive model is simpler, better calibrated out-of-sample, and matches the specification in the accompanying paper. The reliability gate still has a meaningful job: the baseline reads the curve alone, the augmented model reads the curve in the company of broader stress, and the gate learns which read has been more trustworthy over time.

Why this section exists: Transparency about model limitations is as important as the headline probability.

The out-of-sample evaluation covers only 3 recession episodes (roughly 28 recession months out of 356 total). Any statistical analysis of threshold selection, lead times, or calibration is inherently limited by this sample. Results should be treated as suggestive, not definitive.

The augmented model combines its features additively on the log-odds scale. It does not capture conditional effects — for example, an inversion meaning something different when credit stress is high versus low. We tested explicit interaction terms (SPREAD × CREDIT_Z, and others) and they added no distinguishable out-of-sample value, so they were left out for parsimony. A tree-based or neural approach could capture richer non-linearities, but at the cost of interpretability and with higher overfitting risk on 3 episodes.

With only 3 episodes and a low base rate, no single threshold achieves clean separation between true and false alarms. The probability level and direction matter more than whether it crosses a specific line.

This is a monitoring tool based on historical statistical relationships. It does not model causal mechanisms, it cannot anticipate unprecedented shocks (e.g., a pandemic), and it should never be used as the sole basis for financial decisions. It is one input among many.