Lagged credit growth still predicts banking crises, and 12 added years sharpen it

The spec, as published

The dependent variable is crisisJST, the database’s revised coding of systemic banking-crisis start years. The regressors are lags 1 through 5 of the log change in real bank loans, nominal tloans deflated by cpi, plus country fixed effects, fit by logit. The money variant replaces loans with broad money (money), deflated the same way. Standard errors are cluster-robust by country. The war years 1914–1918 and 1939–1945 are dropped. The lag order, the deflation, the fixed effects, the clustering and the war exclusion are all the paper’s choices, held fixed here. One pre-declared robustness variant is run and no other: both models re-fit on the sample ending 2008, the paper’s terminal year.

Missing observations enter no regression. The panel has 2,718 country-years; requiring 5 consecutive lags of real loan growth, a coded crisis flag and non-war years leaves 2,137 for the credit model and 2,204 for the money model. The two models are therefore not fit on identical samples, and their onset counts differ: 72 against 76.

Source: Jorda-Schularick-Taylor Macrohistory Database Release 6, retrieved 2026-07-09. Statistics on the site are subject to revision; the consultation date is part of the citation. Data citation: Jorda, Schularick and Taylor (2017), "Macrofinancial History and the New Business Cycle Facts," NBER Macroeconomics Annual 31.

The horse race

On the full sample the credit model’s pseudo R-squared is 0.076 against the money model’s 0.043, its AUC 0.708 against 0.670, and its clustered joint chi-square 46.7 against 20.9. The paper-era sample orders the same three pairs the same way: 0.069 against 0.041, 0.693 against 0.666, 41.5 against 17.1. That ordering is the paper’s headline horse race.

Source: Jorda-Schularick-Taylor Macrohistory Database Joint tests are Wald tests that all 5 lag coefficients are zero, computed twice: on the cluster-robust covariance (the same one behind every SE on this page) and on the nonrobust ML covariance. Pseudo R2 is McFadden's. AUC is in-sample.

The lag structure: one lag carries the model

In the full-sample credit model, lag 2 enters at 6.74 with p = 8.9e-5; the smallest p-value among the other four credit lags is 0.15. The five credit-lag coefficients sum to 7.70 (6.83 in the paper-era sample, where lag 2 is 6.34 with p = 1.0e-4). That shape, a dominant second lag inside a positive five-lag sum, is the temporal pattern behind the paper’s title: the boom precedes the bust by years, not by months.

Source: Jorda-Schularick-Taylor Macrohistory Database Full sample, both models. p-values from the cluster-robust covariance.

Discrimination, and what a 0.708 AUC means at a 3.37% base rate

The ROC curves below are in-sample: the model is scored on the same observations it was fit on, so they overstate what a real-time forecaster would have had. The base rate makes the trade-off concrete. The credit curve passes through an operating point that catches 75% of the 72 onsets while flagging 49.4% of the 2,065 non-crisis country-years (the three-quarters target behind that reading is our illustrative choice, not the paper’s). The claim that survives is comparative: the credit model’s AUC of 0.708 exceeds the money model’s 0.670.

The joint test depends on the covariance, and both are printed

The sample is not the paper’s

The paper used 14 advanced economies, 1870–2008, and its own 2012 crisis chronology. This vintage carries 18 economies to 2020: the 4 additions relative to the paper’s list are Belgium, Finland, Ireland, Portugal, and they are not passengers, contributing 12 of the 72 onsets in the full-sample credit model. Realized spans differ by country: the credit model’s earliest usable year is 1876, and entry years and onset counts are in the table.

Source: Jorda-Schularick-Taylor Macrohistory Database Composition of the full-sample credit model after lag construction, crisis-flag coverage and the war exclusion. Gaps inside a span (the war years at minimum) mean n is less than the span length.

The dependent variable moved between vintages

The file carries two crisis codings: crisisJST, the revised chronology used in every regression above, and crisisJST_old, the database’s earlier coding. On the 2,568 country-years where both are coded (the earlier coding is null for Ireland in every year before 2020, and nowhere else), the revised chronology has 87 onsets against 90: 3 country-years gained onset status (Belgium 1876, Japan 1901, USA 1930) and 6 lost it (Netherlands 1893, Germany 1907, Netherlands 1907, USA 1929, Denmark 1931, Netherlands 1939). Whether the earlier column matches the 2012 paper’s published dates row for row cannot be checked from this file, so no such claim is made; what the comparison shows is that the dependent variable itself moves between vintages, including the United States’ Depression-era entry shifting from 1929 to 1930.

Source: Jorda-Schularick-Taylor Macrohistory Database Release 6, retrieved 2026-07-09. Statistics on the site are subject to revision; the consultation date is part of the citation.

What this cannot tell you

• This is not the paper’s dataset. Different country set (18 against 14), different years (1870–2020 against 1870–2008), and a revised crisis chronology. Whether the underlying loan and money series were also revised between vintages cannot be checked here, where only Release 6 exists. No coefficient on this page should be expected to match the paper’s to the second decimal, and none is presented as doing so.
• AUC is in-sample. The model is scored on the observations it was fit on. Nothing on this page is an out-of-sample or real-time forecasting claim.
• 18 clusters is few. Cluster-robust SEs and Wald tests rely on many-cluster asymptotics. Both covariances are printed for every joint test because they disagree about the money model.
• The regressors overlap. Adjacent observations share 4 of their 5 lag years, so regressor windows are serially dependent by construction. That is the paper’s design, kept here, and it is part of why clustering matters.
• Attrition is silent and unequal. Of 2,718 panel rows, 2,137 survive to the credit model and 2,204 to the money model. Countries with late loan coverage enter late, and the two horses do not run on identical tracks.
• The war exclusion is inherited, not tested. Dropping 1914–1918 and 1939–1945 is the paper’s specification choice. No variant with the wars included was run, because the discipline here is the paper’s spec plus exactly one pre-declared variant, the 2008 cutoff.