Crisis early warning, honestly evaluated

Methodology

Target and onset definition

A row is a country-year (i, t). The label is 1 when a systemic banking-crisis onset falls in {t, t+1, t+2}, i.e. within 1-3 years after the last observed data year t−1. The onset definition is the crisis layer’s, quoted from the methodology: “the first year of a systemic banking-crisis episode for a country”, where distinct crisis start years are “merged when separated by no more than two tranquil years, i.e. a year y extends the current episode when y − episode.end ≤ 3; otherwise it opens a new episode. The onset is the episode start year.” One chronology per country, never mixed: crisisJST (JST Macrohistory R6) for the 18 JST advanced economies (86 merged onsets), the Laeven-Valencia 2026 banking vintage for every other country (144 merged onsets), the same split the patterns page uses. Unlike the event study, the label set keeps all merged episode starts; the event study’s extra 5-year de-overlap rule (84 + 142 kept onsets there) is replaced here by the post-crisis exclusion below, which removes the ambiguous rows instead of mislabeling them as tranquil.

Features: every lag and transformation

• Credit-to-GDP gap (pp). The BIS published Q4 gap (actual minus one-sided HP trend, private non-financial sector) at t−1 where BIS reports it; otherwise, for the 18 JST countries, a one-sided HP filter gap computed on annual JST credit-to-GDP (100·tloans/gdp) with smoothing λ = 1,562.5 (the Ravn-Uhlig annual equivalent of the BIS quarterly 400,000) and a 10-year burn-in. The artifact records which source each country-year used. The Global Macro Database carries no credit series at all (verified during the patterns build), so for non-JST countries without BIS coverage this feature is missing and stays missing.
• Real credit growth, 3y (%). JST tloans deflated by JST CPI, 3-year growth to t−1; where JST is unavailable, BIS Q4 nominal credit to the private non-financial sector (domestic currency, adjusted for breaks) deflated by the GMD CPI index. Source recorded per country-year.
• Current account / GDP (%). JST 100·ca/gdp at t−1, else GMD CA_GDP.
• Real house-price growth, 3y (%). JST hpnom/cpi 3-year growth to t−1, else GMD rHPI.
• Public debt/GDP change, 3y (pp). JST debtgdp (scaled to percent), else GMD govdebt_GDP; change from t−4 to t−1.
• Real GDP growth (%). One-year growth at t−1; JST rgdpmad, else GMD rGDP.
• CPI inflation (%). One-year rate at t−1; JST CPI growth, else GMD infl.

All growth and change windows are guarded on contiguous calendar years, so a gap in a source series yields a missing value, never a spurious multi-year “annual” change. Missing features are never fabricated: the logistic model winsorizes at the training 1st/99th percentiles, imputes the training median, standardizes with training moments, and carries one 0/1 missing-indicator column per feature so that missingness is itself a signal with an estimated coefficient. The gradient-boosting comparison consumes missing values natively. All preprocessing statistics come from the training rows of each block only.

Sample rules

• GMD observations are cut at 2024 because GMD merges IMF-consistent forecasts into the same columns beyond that; no model input is ever a forecast.
• Labels require a fully observable 3-year window inside the chronology: t+2 ≤ 2020 for JST countries, t+2 ≤ 2023 for the rest; right-censored years are dropped from estimation rather than labeled tranquil. Non-JST rows start at t ≥ 1970, the Laeven-Valencia window.
• War years (1914-1919, 1939-1946) are excluded, the JST-literature convention.
• Post-crisis exclusion: rows with an onset in the previous five years are dropped, so post-crisis collapse years are not learned as tranquil (the post-crisis bias of Bussière and Fratzscher 2006).
• A row needs at least 3 of the 7 features.

Models and evaluation

Headline: L2-regularized logistic regression (C = 1, fixed seed), published because its coefficients are inspectable above. Comparison: histogram gradient boosting (depth 3, 300 iterations, fixed seed). Both are fit unweighted so predicted probabilities are consistent with the sample base rate rather than inflated by class rebalancing. Evaluation is the era-blocked walk-forward described above; the pooled figure combines the seven walk-forward blocks with the final block’s post-2019 rows so no country-year is pooled twice. Deciles and the operating threshold come from the pooled out-of-sample distribution.

What the model has never seen

The chronologies end in 2020 (JST) and 2023 (Laeven-Valencia), so the model has never been shown a crisis after 2023 and cannot have learned anything from events since then. The current gauge is computed only from realized data (vintage: jst-r6(2020);lv-2026-wp2694(2023);bis-gap(2025-Q4);gmd-2026_06(cut2024), model ews-1.0), and its scores are conditional probabilities under a model of the past, not statements about current events.