A banking crisis still raises the odds of a currency crisis, but only 18.6% of banking crises are twins

What Kaminsky and Reinhart claimed

Their paper argued that banking and balance-of-payments crises are linked: banking-sector problems typically precede a currency crisis, the currency collapse then deepens the banking crisis, and both are preceded by recessions and, often, financial liberalization. Their central probability statement was that knowing a banking crisis has begun raises the estimated probability of a currency crisis above its unconditional level, and that the banking→currency link is the more prominent of the two directions. The specific conditional probabilities they reported were computed on monthly data over a 24-month window for their 20-country panel and are not reproduced here, because this atlas is annual and global; only their qualitative pattern is the object of comparison.

Source: FinObservatory crisis atlas (crises_events): currency chronologies from Reinhart-Rogoff, Laeven-Valencia, and the Global Macro Database; banking chronologies from those three plus JST and the ESRB. The claim is Kaminsky, G. L., and C. M. Reinhart (1999), "The Twin Crises: The Causes of Banking and Balance-of-Payments Problems," American Economic Review 89(3), 473-500.

What we run it on

The atlas view crises_events holds one row per (country, year, crisis type, source); those rows are merged events, not episodes, so counting them overstates history. Every figure here first collapses to distinct (country, year, type) and then to episode onsets with the site’s one shared episode rule (distinct crisis years of a type merged when separated by at most two non-crisis years; the onset is the episode’s first year), applied per type exactly as the crisis-patterns event study derives its banking onsets. That leaves 442 banking onsets and 566 currency onsets across the 160 countries the atlas covers for these two crisis types, 1800–2025.

The unconditional base rate needs an honest denominator. We use each country’s observed event span (its first to its last recorded banking or currency event, inclusive) rather than the full padded1800–2025 grid, because a country that never enters the chronology should not count as a run of tranquil years. That gives 9,960 country-years and an unconditional currency-onset rate of 5.7% per country-year (4.4% for banking). This denominator is bracketed by crises at both ends, so it runs the base rate high, which makes every lift below a conservative reading rather than a flattering one.

The conditional probability that a currency onset falls within h years of a banking onset is compared to a baseline computed the same way: the probability that a currency onset falls within h years of a random covered country-year. Both apply the identical rule that the full h-year window must lie inside the country’s coverage, so right-censoring at a span’s end cannot bias the conditional against its baseline. A lift is the ratio of the two; a lift of one means the onset carries no information beyond the base rate.

Source: FinObservatory crisis atlas (crises_events): currency chronologies from Reinhart-Rogoff, Laeven-Valencia, and the Global Macro Database; banking chronologies from those three plus JST and the ESRB. Onset construction reuses src/lib/episodes.ts, the same merge behind every /crises page.

The headline: banking leads currency, weakly

On the full atlas, a banking onset raises the near-term currency-crisis probability above baseline: 7.0% the same year against 5.7%, 14.6% within a year against 9.9%, a 1.47× lift. The reverse direction is present but weaker (a banking onset follows within a year of 9.9% of currency onsets, a 1.21× lift). That asymmetry, banking leading currency more than currency leads banking, is the shape Kaminsky and Reinhart reported.

Source: FinObservatory crisis atlas (crises_events): currency chronologies from Reinhart-Rogoff, Laeven-Valencia, and the Global Macro Database; banking chronologies from those three plus JST and the ESRB. Full atlas. "Onsets that follow" is the count of conditioning onsets whose observable h-window contains a target onset, over all conditioning onsets whose window is observable.

Most banking crises are not twins

The lift is real but modest, and the reason is the twin share. Of the 442 banking onsets in the atlas, only 82 (18.6%) have a currency onset within 1 year on either side. Kaminsky and Reinhart’s picture of banking and currency crises as two faces of one event is, on this annual global chronology, the minority case: the typical banking crisis in the atlas arrives without a currency crisis beside it. The lift says a banking onset shifts the odds; the twin share says the shift is from uncommon to still-uncommon.

Robustness: the KR era and corroborated events

Two cuts test how much the headline depends on the sample. On the authors’ own 1970–1995 window the lift nearly disappears: the same-year currency-after-banking probability is 6.7% against a 8.7% baseline, a 0.76× lift below one, and even the one-year lift is only 1.10×. In that era currency crises were so common (8.7% per country-year against 5.7% on the full atlas) that a banking onset adds almost nothing to the forecast. Restricting to events flagged by at least two sources moves the other way: the one-year banking→currency lift rises to 1.85×.

Source: FinObservatory crisis atlas (crises_events): currency chronologies from Reinhart-Rogoff, Laeven-Valencia, and the Global Macro Database; banking chronologies from those three plus JST and the ESRB. All three cuts use the identical onset construction and matched-baseline rule; only the year window and the source-agreement floor differ.

Where ours differs from theirs

This is a re-examination on different measurement, not an exact replication, and the differences are large enough that no number here should be expected to match the paper’s.

• Annual, not monthly. Kaminsky and Reinhart dated crises to the month off an exchange-market-pressure index; the atlas dates onsets to the year off merged published chronologies. Sub-year lead-lag, the core of their monthly evidence, is invisible here, and a same-year (h = 0) twin can order either way within the year.
• A global panel, not their 20 countries. Their sample was 20 mostly middle- and high-income countries, 1970–1995; this is 160 countries over 1800–2025. On their own window the lift is 0.76× to 1.10×; the 1.47× full-sample lift comes from the wider era and country set they did not use.
• Chronology onsets, not a constructed pressure index. Currency crises here are whatever the merged Reinhart-Rogoff, Laeven-Valencia and Global Macro Database chronologies flag, not a threshold on a market-pressure series. The two definitions overlap but are not the same event, so a currency crisis in this atlas need not be one KR would have dated, and vice versa.
• A base rate, not a causal model. KR modelled the joint dynamics and the role of liberalization and reserves; this note counts co-occurrence. A 1.47× lift is an association in the chronology, not evidence that the banking crisis caused the currency crisis.

What this cannot tell you

• The denominator is a choice. The 5.7% base rate uses each country’s observed event span, which brackets the country between crises and so runs high; a wider window would lower the base rate and raise every lift. The span is stated so the reader can see what the lift is measured against.
• Merged chronologies disagree. The atlas unions sources that date the same crisis a year or two apart, which is why the 1-year twin window and the two-source robustness cut both matter: at 19.7% the twin share barely moves when noisy single-source events are dropped, but the underlying counts do.
• Onsets are per type. Banking and currency onsets are merged into episodes separately, so a country in a long serial-crisis run can contribute several onsets of each type. The site’s cross-type episode URLs merge differently; this note deliberately does not, because the twin question is about the timing of one type relative to the other.
• Coverage edges cut both ways. The 1970–1995 window clips episodes in progress at its left edge, so an ongoing crisis can re-date its onset to the first window year; and the twin window counts an onset at the very end of a country’s coverage as a non-twin even though the following year is unobservable. Both choices are stated rather than adjusted for.
• No inference is attached. These are counts and ratios, not model coefficients with standard errors. With 82 twins the sampling uncertainty is real and no confidence interval is drawn; the lift is reported as a descriptive frequency.