After the deepest reserve declines, 5.8% of country-years start a currency crash; after the fastest reserve growth, 1.5%

What Frankel and Rose claimed

Their paper characterized currency crashes on annual data for over one hundred developing countries over 1971–1992, defining a crash as a nominal depreciation of at least 25% that is also at least a 10-percentage-point acceleration over the prior year’s depreciation, with a 3-year window so that one crisis is not counted twice. Two of its findings are tested here: that crashes come on the heels of falling reserves, and the null the paper is remembered for, that current-account deficits play little role in predicting crashes. Their event-study figures and probit estimates also involved the composition of capital inflows (a low FDI share of debt), domestic credit growth, northern interest rates, and real overvaluation; those predictors are outside this note’s two-claim scope, and their coefficients are not reproduced, because this note runs no regression and holds nothing constant.

Source: Crisis atlas panel (crises_panel): USDfx (GMD, LCU per USD, annual average) and CA_GDP (GMD, % of GDP); total reserves (reserves view, World Bank FI.RES.TOTL.CD, USD level); JST Macrohistory Database for the advanced-economy exclusion list. The claim under test is Frankel, J. and A. Rose (1996), “Currency Crashes in Emerging Markets: An Empirical Treatment,” Journal of International Economics 41(3–4), 351–366.

What we run it on

The crisis atlas panel crises_panel carries the GMD annual-average exchange rate (USDfx, LCU per USD, so an increase is a depreciation) and the current-account balance (CA_GDP); the reserves view carries total reserves in USD (World Bank FI.RES.TOTL.CD). The sample is every country-year with a valid exchange rate outside the 18 JST advanced economies, the developing-country proxy this estate supports, the same country-level cut the crisis atlas’s patterns page uses for its emerging-market panel: 217 economies, 1802–2025.

• The crash rule is the paper’s. The depreciation rate at year t is 100×(USDfx(t)/USDfx(t−1) − 1), computed only when the t−1 row is the immediately prior calendar year, and a crash is a year with a depreciation of at least 25% that is also at least 10 percentage points faster than the prior year’s. The prior year’s depreciation carries the same calendar guard, so classifying a year requires three consecutive exchange-rate observations, and a gap in the series drops the year rather than silently stretching the window.
• Onsets follow their 3-year window rule. A crash year counts as an onset only if no observed crash occurred in the prior 3 years (a year the data cannot classify counts as crash-free, the same convention the paper faced); crash years that are not onsets, the continuations inside a spell, are dropped from the counting panel entirely, so the tranquil side never includes mid-spell years. That leaves 19,226 classified country-years with 536 onsets (2.79%). As a check that the rule flags the crashes everyone remembers, 4 of the 4 celebrated onsets (Mexico 1995, Thailand 1998, Indonesia 1998, Argentina 2002) are flagged by the computed panel.
• Both predictors enter with a one-year lag. Reserve growth is 100×(reserves(t−1)/reserves(t−2) − 1), from USD reserve levels with the same calendar guard and both levels positive: a scale-free proxy for the paper’s falling-reserves claim. The paper’s own regressor was the level of reserves scaled by imports; an import series exists in the raw GMD file, but it comes from a different compiler than the World Bank reserve levels used here, and this note does not mix two compilers’ levels inside one ratio. The current-account predictor is CA_GDP at t−1, straight from the panel.
• Quartiles are pooled, and every benchmark is computed. Each table sorts its own joined sample into four near-equal groups by the lagged predictor (7,775 country-years for reserve growth, 7,847 for the current account) and reads each quartile’s onset rate against that sample’s own pooled rate (3.02%, 235/7,775 and 3.25%, 255/7,847), computed not assumed. The single pre-declared variant re-cuts the reserve-growth quartiles inside 1971–1992, the paper’s own era (2,190 country-years).

Source: Crisis atlas panel (crises_panel): USDfx (GMD, LCU per USD, annual average) and CA_GDP (GMD, % of GDP); total reserves (reserves view, World Bank FI.RES.TOTL.CD, USD level); JST Macrohistory Database for the advanced-economy exclusion list. The claim under test is Frankel, J. and A. Rose (1996), “Currency Crashes in Emerging Markets: An Empirical Treatment,” Journal of International Economics 41(3–4), 351–366. All numbers on this page are computed at build time from src/lib/currencyCrashes.ts; nothing is typed in.

Reserve growth in t−1: the onset rate falls quartile by quartile

The deepest-decline quartile, prior-year reserve changes of -98.7% to -4.9%, starts a crash in 5.8% of country-years (112 of 1,944), nearly twice the 3.02% benchmark, and the rate falls monotonically across the quartiles: 2.5% (48/1,944), 2.4% (46/1,944), then 1.5% (29/1,943) in the fastest-growth quartile, a 4.3pp gap from top to bottom of the reserve distribution.

Source: Crisis atlas panel (crises_panel): USDfx (GMD, LCU per USD, annual average) and CA_GDP (GMD, % of GDP); total reserves (reserves view, World Bank FI.RES.TOTL.CD, USD level); JST Macrohistory Database for the advanced-economy exclusion list. The claim under test is Frankel, J. and A. Rose (1996), “Currency Crashes in Emerging Markets: An Empirical Treatment,” Journal of International Economics 41(3–4), 351–366. The onset rate is crash onsets over country-years in the quartile; the sample's pooled rate is 3.02% (235/7,775), the benchmark every quartile is read against.

The 1971–1992 window: the paper’s own era

Keeping the country-years that fall in 1971–1992, the paper’s own sample years, and re-drawing the quartiles inside that set leaves 2,190 country-years with 92 onsets (4.20%), and the same ordering: 6.0% (33/548) after the deepest reserve declines, then 4.2% (23/548), 3.5% (19/547), and 3.1% (17/547) after the fastest growth, a 2.9pp gap inside the era the paper measured.

Source: Crisis atlas panel (crises_panel): USDfx (GMD, LCU per USD, annual average) and CA_GDP (GMD, % of GDP); total reserves (reserves view, World Bank FI.RES.TOTL.CD, USD level); JST Macrohistory Database for the advanced-economy exclusion list. The claim under test is Frankel, J. and A. Rose (1996), “Currency Crashes in Emerging Markets: An Empirical Treatment,” Journal of International Economics 41(3–4), 351–366. Quartiles and the pooled rate are recomputed within 1971–1992; the pooled rate there is 4.20% (92/2,190).

The current account in t−1: the paper’s null, tested

The identical design on the lagged current-account balance yields 7,847 country-years with 255 onsets (3.25%). The deepest-deficit quartile, balances of -240.5% to -7.8% of GDP, starts a crash in 3.3% of country-years (65 of 1,962), close to the 3.25% benchmark and below the second quartile’s 4.3% (85 of 1,962); only the surplus quartile stands clearly apart at 1.6% (31 of 1,961), 2.7pp under the riskiest bucket. The tails of the GMD ratio are extreme: the deepest recorded lagged deficit is -240.5% of GDP and the largest surplus 311.7%, and quartile membership, not the tail values, drives the counts.

Source: Crisis atlas panel (crises_panel): USDfx (GMD, LCU per USD, annual average) and CA_GDP (GMD, % of GDP); total reserves (reserves view, World Bank FI.RES.TOTL.CD, USD level); JST Macrohistory Database for the advanced-economy exclusion list. The claim under test is Frankel, J. and A. Rose (1996), “Currency Crashes in Emerging Markets: An Empirical Treatment,” Journal of International Economics 41(3–4), 351–366. Same panel, onset rule, pooled quartiles, and computed benchmark as the reserve tables, with CA_GDP(t-1) as the predictor; the pooled rate here is 3.25% (255/7,847).

Reading the result

• The reserves claim reproduces, and it is monotonic. The onset rate falls with every quartile of prior-year reserve growth, 5.8%/2.5%/2.4%/1.5% in the full sample and 6.0%/4.2%/3.5%/3.1% inside 1971–1992, so the deeper the prior-year reserve loss, the more often a crash follows, at both cuts this panel supports.
• The famous null largely reproduces. The deepest-deficit quartile is not the riskiest bucket: it starts crashes at 3.3% (65/1,962) against the second quartile’s 4.3% (85/1,962) and a 3.25% benchmark, and only the surplus quartile is clearly protected at 1.6% (31/1,961). The computed cells say exactly that much and no more: deficits do not rank country-years by crash risk, while surpluses mark the safe end.

What this cannot tell you

• Reserve growth is our proxy, not their regressor. The paper scaled the level of reserves by imports. A GMD import series exists in the estate, but joining it to the World Bank reserve levels would put two compilers inside one ratio, so the scale-free growth rate of USD reserves stands in for the falling-reserves claim, and a country holding low but stable reserves is invisible to it.
• Most of their regressor set is untested. The paper’s remaining predictors all exist in the estate in some form and are left untested here rather than claimed absent: FDI stocks (EWN and the IMF balance of payments, from which a composition share is one division away), domestic credit to the private sector (a World Bank series covering most of the sample countries), real effective exchange rates (GMD), and northern policy rates (the panel’s own US series). This note tests two claims; the paper’s findings on debt composition, credit, overvaluation, and external conditions are neither supported nor contradicted by it.
• The sample is a proxy for “emerging”. Excluding the 18 JST advanced economies is the developing-country cut this estate supports, not the paper’s own list of over one hundred developing countries, and the 217 economies here include small and micro states the paper did not cover.
• Annual averages understate crashes. USDfx is an annual average, so a collapse late in the year is smoothed into two smaller year-over-year changes, and a crash that begins and reverses within a year can fall below the 25% threshold entirely.
• The counts are serially dependent. Tranquil years within a country share persistent fundamentals, and the 3-year window makes onset eligibility depend on neighbouring years, so the 7,775 country-years overstate the number of independent observations and no rate on this page carries a standard error.
• Counting is not their probit. The paper’s claims are about probit marginal effects with the other regressors held constant; a quartile count with no controls is a weaker, assumption-light check, so agreement here supports the direction of their findings without reproducing their magnitudes, and disagreement would not have refuted the probit.