Bank-equity crashes without panics are followed by a 3.6% output loss after three years; with panics, 10.5%

Counting the crashes

The panel flags 262 country-years with a 30 percent real bank-equity decline (the paper’s threshold, carried in the panel’s bank_eq_crash30 column), in 45 of its 46 countries: 192 without a panic that year and 70 with one. Collapsing consecutive crash years within a country gives 218 distinct episodes, 149 quiet and 69 containing at least one panic year. Any “number of crises” depends on the unit, so each count below states whether it is country-years or episodes.

The companion crisis list carries 224 episodes: 192 with a panic, 138 with a 30 percent bank-equity decline (39 entries leave that flag empty), 193 with widespread bank failures, 35 newly added by the authors relative to prior chronologies, and 17 flagged for wartime exclusion.

Coverage is a real constraint: bank_eq_crash30 is NULL in 1,898 of the panel’s 6,177 country-years, exactly the country-years where the panel’s real bank-equity return is also missing, and the panic flag is NULL in 1,868. Those years are unmeasured, not crash-free, and every count and regression here conditions on measurement.

Quiet crashes and output: the local projections

The outcome is cumulative log real GDP from the year before the crash (t-1) to t+h, built by summing log(1 + real_gdp_growth) from the panel’s growth column, which is a decimal fraction; values are multiplied by 100 only for display. At each horizon h = 0 to 6, a Jorda local projection regresses that outcome on a quiet-crash dummy (crash, no panic) and a panic-crash dummy (crash with panic), with country fixed effects and 3 lags of log growth, standard errors clustered by country (46 clusters). The 30 percent threshold is the paper’s; the 3-lag control set is our approximation of their controls and may not match their published specification line by line. Both dummies stay missing wherever bank_eq_crash30 is NULL, since coding unmeasured years as 0 would pool them into the control group. One row drops on a source artifact: Germany’s 1946 growth is -1.0019, below -100 percent, a break at the WWII boundary in the source file, and log(1 + g) is undefined there.

The quiet-crash path bottoms out at h = 3 at -3.62 (SE 1.30), on 172 in-sample quiet-crash observations. The panic-crash path bottoms out at h = 5 at -11.46 (SE 1.60) and rests on 67 panic-crash observations at that horizon, so its precision should be read with that count in mind.

Source: Baron, Verner and Xiong (2021) replication panel (bvx_annual_panel, bvx_crisis_list) Coefficients in cumulative log real GDP x 100 relative to t-1; country-clustered SEs in parentheses. Postwar restricts to the panel's own postwar flag, 1946 onward (3,258 country-years), the one pre-declared robustness variant.

The postwar restriction is the single pre-declared robustness variant: at h = 3 the quiet-crash coefficient moves from -3.62 to -5.22 (SE 1.16) and the panic-crash coefficient from -10.49 to -12.69 (SE 1.36). The German 1946 row is inside the postwar flag (1946 onward), so the variant does not remove that artifact; the row drops in both runs because its growth value is below -100 percent.

Where the crashes sit against the chronologies

The paper’s pitch is that equity crashes catch distress that narrative chronologies miss, so the natural check is how often the 262 crash country-years land on a chronology date. A coding trap sits here: the panel’s imported crisis flags for BVX, Reinhart-Rogoff and Laeven-Valencia are coded 1-or-NULL, with zero recorded 0s (0, 0 and 0 zeros respectively), so treating NULL as “no crisis” would fabricate a denominator. Overlap is therefore reported as join counts over the 262 crash country-years, never as rates over uncovered years. The estate’s independent rr_crises (870 banking-crisis country-years across its banking and banking_systemic crisis types) and laeven_valencia (151 banking-crisis start years) tables allow a windowed second look.

Source: Baron, Verner and Xiong (2021) replication panel (bvx_annual_panel, bvx_crisis_list) | rr_crises (Reinhart-Rogoff chronology) | laeven_valencia (Laeven-Valencia banking crises) Counts of the 262 crash country-years matched by each chronology. The narrative flag is the panel's own genuine 0/1 column (192 ones across the panel; 81 of them on crash years).

Two rows need reading care. The Reinhart-Rogoff exact-year count jumps from 51 on the panel’s flag to 107 on the independent table. The independent table marks crisis years in progress: its 870 banking country-years collapse to 308 distinct runs, 179 of them longer than a single year. The two sources also disagree in coverage: 29 of the panel’s 215 Reinhart-Rogoff flag years do not appear in the independent table at all, so the joins compare related but not identical vintages of that chronology. And the BVX row is a consistency check rather than new information: each of the panel’s 208 bvx_crisis years falls on a crisis-list year (208 of 208), and each of the 208 list episodes with a same-year panel row is flagged (208 of 208), so the flag is the list restricted to panel coverage.

What this cannot tell you

• This is not an independent replication. bvx_annual_panel is the authors’ replication dataset. Reproducing their facts from it verifies the estate’s copy and our implementation of the spec, not the paper’s robustness to new data.
• The specification is an approximation. We rebuild cumulative GDP by summing log(1 + growth) from the panel’s growth column rather than from the authors’ underlying GDP levels, and our control set (country fixed effects plus 3 lags of log growth) is our reading of their setup, not a line-by-line copy of a published table. No coefficient here should be quoted as the paper’s.
• No published coefficient is compared line by line. We did not carry the paper’s own table values onto this page, so “reproduces” means the qualitative facts (crash counts, quiet-crash output losses, larger panic losses) come out of the panel, not that each number matches a published cell.
• The panic coefficients rest on 67 observations. Cumulative outcomes at neighbouring horizons overlap by construction and are serially correlated; country-clustered SEs (46 clusters) are the minimum honest treatment, and no significance is claimed beyond what those SEs give.
• Absence is not zero. 1,898 country-years lack the crash flag and 1,868 lack the panic flag; the growth column itself is missing in 815 rows. Frequency and overlap statements condition on measured years only.
• The index construction is inherited, not audited. The historical bank-equity returns behind bank_eq_crash30 are the authors’ reconstruction from historical sources, with whatever survivorship and listing biases that entails; this note cannot re-audit them.