Crisis layer methodology

The crisis atlas is built from five primary chronologies merged into one canonical event table (6,518 source rows, 4,798 distinct events) and a country-year panel (49,347 rows, 239 countries, 1800–2025). Every row traces to a flag in a source file; no events are invented, none are arbitrated away.

Event definition

An event is a distinct (country, year, crisis type) triple. The event table keeps one row per (event, source), so a crisis flagged by three chronologies appears as three rows of one event. Sources agreeing is the count of distinct chronologies flagging that triple. Union rule: a year flagged by any source is in the table. Types are banking, currency, and sovereign debt (defaults and restructurings combined); Reinhart-Rogoff inflation crises and ESRB residual (non-systemic) events are excluded.

What each source calls a crisis

The layer imposes no definition of its own. Each chronology’s definition, quoted or closely paraphrased from its own documentation:

Global Macro Database

All three types: GMD does not define crises independently: its flags are harmonized composites of other chronologies. Banking: “We combine four measures of banking crises. We take data from Baron, Verner, and Xiong (2019); Laeven and Valencia (2020); Jordà, Schularick, and Taylor (2017); and Reinhart and Rogoff (2019)”, in that priority order. Currency and sovereign debt combine Laeven-Valencia and Reinhart-Rogoff, prioritizing Laeven-Valencia; the sovereign flag counts any Laeven-Valencia default or restructuring year, else any Reinhart-Rogoff external or domestic default year. For all three, GMD counts a new crisis “only ... if there was no crisis in the previous three years”.

GMD replication code, github.com/KMueller-Lab/Global-Macro-Database, code/combine/ (BankingCrisis.do, CurrencyCrisis.do, SovDebtCrisis.do).

Jordà-Schularick-Taylor

Banking (the only type JST codes): The crisisJST variable “takes unity in the first year a country is experiencing a systemic banking crisis, and zero else. An episode of banking distress is coded as a systemic banking crisis if it is characterized by major bank failures, banking panics, substantial losses in the banking sector, significant recapitalization, and/or significant government intervention. Importantly, this definition excludes the failures or losses of individual/small banks without systemic implications”.

JST Financial Crisis Chronology (Jordà, Schularick and Taylor, February 25, 2021), p. 1, macrohistory.net.

Laeven-Valencia

Banking: “An event that meets two conditions: 1) Significant signs of financial distress in the banking system (as indicated by significant bank runs, losses in the banking system, and/or bank liquidations). 2) Significant banking policy intervention measures in response to significant losses in the banking system.” The start year is “the first year that both criteria are met”.
Currency: A “sharp” nominal depreciation of the currency vis-a-vis the U.S. dollar, meeting two thresholds: “i) a year-on-year depreciation of at least 30 percent; and ii) of at least 10 percentage points higher than the rate of depreciation observed in the year before” (building on Frankel and Rose 1996).
Sovereign debt: “The year of sovereign default to private creditors and/or restructuring. If public debt was restructured without a suspension of payments, the sovereign crisis year is recorded as the year of the restructuring.”

Laeven and Valencia, Systemic Banking Crises Revisited, IMF WP/18/206 (2018), Sections II and IV (published as IMF Economic Review 68(2), 2020).

Reinhart-Rogoff

Banking: “We mark a banking crisis by two types of events: (1) bank runs that lead to the closure, merging, or takeover by the public sector of one or more financial institutions; or (2) if there are no runs, the closure, merging, takeover, or large-scale government assistance of an important financial institution (or group of institutions), that marks the start of a string of similar outcomes for other financial institutions.”
Currency: A currency crash, following a variant of Frankel and Rose (1996): “an episode is counted for the entire period in which annual depreciations exceed the threshold of 15 percent per annum”.
Sovereign debt: “External debt crises involve outright default on payment of debt obligations incurred under foreign legal jurisdiction, repudiation, or the restructuring of debt into terms less favorable to the lender than in the original.” Domestic debt crises are coded with the same approach; many involve deposit freezes and/or the forcible conversion of foreign-currency deposits into local currency.

Reinhart and Rogoff, From Financial Crash to Debt Crisis, NBER Working Paper 15795 (2010), Section II (published in American Economic Review 101(5), 2011).

ECB/ESRB

Systemic financial crisis (mapped to banking here): Crises are identified “by combining a quantitative approach based on a financial stress index with expert judgement from national and European authorities”. A crisis is considered systemic when “a systemic crisis entails (i) the financial system acting as a shock originator or amplifier and/or (ii) systemic financial intermediaries experiencing distress or going bankrupt and/or (iii) substantial crisis management policy interventions”. The database’s 42 residual (non-systemic) stress events are excluded from the FinObservatory event table.

Lo Duca et al., A new database for financial crises in European countries, ESRB Occasional Paper No 13 (2017).

Source semantics are preserved, not homogenized

GMD and JST date crises by start year only. Laeven-Valencia banking episodes and ESRB systemic episodes carry start and end years and are expanded to one row per crisis year. Reinhart-Rogoff is an annual in-crisis chronology and is kept as-is. Consequence: agreement is highest on start years and drops on continuation years (for example, USA banking 2007 is flagged by 4 sources; 2008 by 2). Dating disagreements are kept, not arbitrated: RR dates the Thai currency crisis 1997, GMD and Laeven-Valencia 1998, and both years appear.

Cross-source agreement

Distribution over the 4,798 events:

Sources agreeing	Events
1	3,421
2	1,065
3	289
4	15
5	8

Single-source events are expected: the sources differ in country coverage, sample windows, and dating convention. Measured on start years, GMD matches 93% of Laeven-Valencia banking episodes within one year but only 78% of JST starts; GMD banking starts fall within one year of just 42% of ESRB systemic episode starts, because ESRB counts many European stress episodes that global banking chronologies do not classify as systemic banking crises.

Sources

Reinhart-Rogoff Global Crises Data (HBS BFFS distribution) (4,043 rows in the event table)
Annual in-crisis chronologies (banking, currency, sovereign external and domestic default), 70 countries, 1800-2016.
Reinhart and Rogoff (2009), This Time Is Different; data via the HBS Behavioral Finance & Financial Stability Project (file dated 2016-09-23), with coauthors Trebesch and V. Reinhart.
License: No explicit license stated; published by the HBS BFFS Project for free download. Cite Reinhart-Rogoff and the BFFS Project.
Global Macro Database (GMD), release 2026_06 (1,187 rows in the event table)
Banking, currency, and sovereign-debt crisis start years, 1800-2017, global coverage.
Müller, Xu, Lehbib and Chen (2025), NBER Working Paper 33714. Data release 2026_06, globalmacrodata.com, consulted 2026-07-09.
License: Research Use Terms: free for academic and non-profit research, not for commercial or for-profit institutional use. Citation required.
Laeven-Valencia Systemic Banking Crises Database II (914 rows in the event table)
Banking episodes with start and end years (expanded to annual rows), plus currency and sovereign-default/restructuring start years, 1970-2017.
Laeven and Valencia (2020), IMF Economic Review 68(2): 307-361. Working-paper version: IMF WP/18/206.
License: No explicit data license; copyright International Monetary Fund 2020. Distributed as the published article's supplementary file; cite the paper.
ECB/ESRB European Financial Crises Database (286 rows in the event table)
Systemic financial-crisis episodes with start and end dates (expanded to annual rows), 28 European countries, 1973-2021.
Lo Duca et al. (2017), ESRB Occasional Paper No 13; database vintage of 2022-01-20 (December 2023 update). Systemic episodes only.
License: Reproduction permitted provided the source is acknowledged (ESRB disclaimer).
Jordà-Schularick-Taylor Macrohistory Database, Release 6 (88 rows in the event table)
Banking-crisis start years (crisisJST), 18 advanced economies, 1870-2020.
Jordà, Schularick and Taylor (2017), NBER Macroeconomics Annual 31. Consulted 2026-07-09.
License: Free license to use and share “provided that it is for non-commercial (e.g., academic) purposes” (macrohistory.net terms). Citation with consultation date required.

Context datasets

Three further datasets sit in the FinObservatory data layer for indicator work. They do not enter the crisis event table.

BIS Data Portal
Credit-to-GDP gaps, total credit to the non-financial sector, debt service ratios, residential property prices (WS_CREDIT_GAP, WS_TC, WS_DSR, WS_SPP; downloaded 2026-07-09).
License: BIS terms: statistics may be used, reproduced and distributed free of charge provided the source is cited (“Source: BIS”).
FRED, Federal Reserve Bank of St. Louis
US financial-stress and recession series (STLFSI4, NFCI, ANFCI, USREC, T10Y2Y, T10Y3M, BAA10Y, VIXCLS, DRTSCILM, TEDRATE; retrieved 2026-07-09).
License: Publicly downloadable under the FRED Terms of Use; underlying source cited per series. VIXCLS is copyright Cboe (redistribution restricted).
World Bank Global Financial Development Database (GFDD)
108 financial-development indicators, 215 economies, 1960-2021 (September 2022 release). Citation: Čihák, Demirgüç-Kunt, Feyen and Levine (2012), World Bank Policy Research Working Paper 6175.
License: CC-BY 4.0.

Run-up patterns event study

The run-up patterns page is an event study of the macro-financial run-up to systemic banking crises, in two deliberately separate panels. For each variable it plots the median trajectory across kept crisis episodes over the eleven years centred on the onset (t-5 to t+5), with an interquartile band for cross-episode dispersion and a flat line for the tranquil-period benchmark. It describes what preceded past crises; it is not an early-warning signal, a forecast, or a probability for any country today. Panel A (below) is the advanced-economy study from JST; Panel B (further down) is the emerging-and-developing study from Laeven-Valencia and the Global Macro Database. The two are not merged and not directly comparable.

Panel A: advanced economies (1870–2020, Jordà-Schularick-Taylor)

Event definition. An onset is the first year of a systemic banking-crisis episode, taken only from the crisisJST chronology in the Jordà-Schularick-Taylor Macrohistory Database (18 advanced economies), not from the merged five-source event table above. Raw crisisJST start years are collapsed into episodes with the repository’s single toEpisodes rule (distinct crisis years bridged across gaps of at most two tranquil years), so a handful of adjacent start-year pairs (BEL 1931 and 1934, PRT 1920 and 1923) fold into one episode. A later onset within five years of a prior kept onset in the same country is then dropped, because its t-5 to t+5 window is already covered by the earlier onset’s window; this drops ESP 1924 (four years after ESP 1920) and ITA 1935 (five years after ITA 1930). The result is 84 kept onsets across 18 economies, spanning 1870–2008 over a JST sample of 1870–2020.

Window and statistics. Relative years t-5 to t+5 (11 points around t0 = onset). For each variable and relative year, across the contributing episodes: a median path, an interquartile band (25th to 75th percentile, linear interpolated), and the sample size n. Medians rather than means are used because the panel spans wars, hyperinflations, and gold-standard breaks, where a few extreme country-years would dominate a mean. The tranquil benchmark is the median of the variable across all country-years more than five years from every kept onset in the same country, over the same 18-country JST sample.

The five variables, each with its source, the number of kept onsets contributing an in-window observation, the tranquil benchmark, and the median in the year before onset (t-1):

Variable	Source	Episodes	Tranquil	Median at t-1
Credit-to-GDP change	JST	77	+0.64pp	+1.73pp
Current account / GDP	JST	81	-0.14%	-1.50%
Real house-price growth	JST	59	+1.97%	+2.33%
Real GDP growth	JST	84	+2.21%	+2.05%
BIS credit-to-GDP gap	BIS	25	-1.02pp	+13.51pp

Four variables (credit-to-GDP change, the current account ratio, real house-price growth, and real GDP growth) come from the JST table itself, so onsets and covariates share the same (country, year) grid. The fifth, the BIS credit-to-GDP gap, is the year-end (Q4) actual-minus-trend gap for the private non-financial sector; it covers only the modern onsets, because BIS quarterly data begin around 1961.

Reading. Credit-to-GDP change accelerates into the crisis, the median peaking at +2.13pp at t-2 against a +0.64pp tranquil pace, then contracting after the onset. The current account slips to a -1.50% median at t-1 versus a -0.14% tranquil balance. Real house-price growth crests near +5.92% at t-2 (tranquil +1.97%) and turns negative from the onset. Real GDP growth holds near its +2.21% tranquil pace until it collapses to +0.33% at t0 and -0.13% at t+1. The BIS credit-to-GDP gap sits at a +13.51pp median at t-1 on a 24-episode sample, far above both the tranquil -1.02pp level and the +10pp Basel upper buffer threshold, before falling sharply post-crisis.

Advanced-economy skew. JST is 18 advanced economies, so these are the run-ups of rich, financially deep economies; emerging-market banking crises are not represented here. Panel B below addresses this directly on a broad country set.
Annual granularity. Onset dating is to the year and the variables are annual, so sub-year lead-lag structure (the quarter credit turns, the month house prices peak) is invisible.
Not a prediction. These are unconditional median paths around crises that did happen, not a model. They do not identify triggers, control for confounders, or separate crises from booms that did not end in one, and a country can sit inside the band for years without a crisis.
Small modern-gap sample. The BIS-gap panel rests on about 25 episodes; its post-crisis tail is dominated by the 2007-08 cluster and should not be over-read.

License.Bound by the JST Macrohistory Database license: free open access, no commercial redistribution, citation with consultation date required. The BIS credit-to-GDP gap is shown with a “Source: BIS” attribution. Consistent with the crisis layer’s posture, the page is labelled non-commercial research use only.

Citations: Jordà, Schularick and Taylor (2017), NBER Macroeconomics Annual 31 (Release 6, macrohistory.net, consulted 2026-07-09); Bank for International Settlements, BIS Data Portal (credit-to-GDP gap).

Panel B: emerging and developing economies (1976–2023, Laeven-Valencia + GMD)

Panel A’s stated limit is advanced-economy skew. Panel B answers it with a second, deliberately separate event study on a broad country set, using a different chronology and a different macro panel. It keeps the identical window, episode definition, and median/IQR machinery, so the two panels are constructed the same way even though they are not merged and not directly comparable (see the comparability note below).

Onsets. The event is the first year of a systemic banking-crisis episode from a single named chronology, the laeven_valencia_2025 banking vintage (Laeven-Valencia 2026, one row per episode with an explicit start year). Chronologies are never mixed: Panel B uses only Laeven-Valencia onsets and only GMD macro, and Panel A uses neither.

Exclusion (no double-counting).Any Laeven-Valencia banking onset whose country is one of Panel A’s 18 JST economies is dropped from Panel B, so every crisis belongs to exactly one panel (the 2008 crises in Belgium, Germany, Spain and the rest; the Nordic 1991 crises; the US 1988 and 2007 crises, and the others). The rule is country-level rather than exact-year on purpose: Laeven-Valencia and JST sometimes date the same crisis a year or two apart (Laeven-Valencia dates Norway 1991 and the US S&L crisis 1988, JST 1988 and 1984), so an exact-year exclusion would leave those double-counted.

Merge and overlap.The surviving onsets are collapsed with the repository’s single toEpisodes rule, exactly as in Panel A; only the Democratic Republic of the Congo’s 1991 and 1994 starts (three years apart) fold into one episode. The overlap rule then drops a later onset within five years of a prior kept onset in the same country, removing Brazil 1994 (four years after 1990) and Chile 1981 (five years after 1976). The result is 142 kept onsets across 104 economies, spanning 1976–2023. Most of the 104 economies are middle- or low-income, but some are non-JST high-income economies (Korea, Greece, Iceland, the transition economies), so Panel B is honestly the complement of the 18 JST countries, not a pure emerging-market panel.

Variables and statistics. The window (t-5 to t+5), the median/IQR path, and the tranquil benchmark (median of the variable across country-years more than five years from every kept Panel B onset in the same country) are identical to Panel A. GMD carries no credit series and only advanced-economy-heavy house-price data, so Panel A’s credit-to-GDP change and real house-price growth cannot be reproduced for this sample; they are replaced by CPI inflation and two exchange-rate views, the dimensions that define emerging-market banking crises. All GMD macro is cut at 2024 to exclude its merged IMF-consistent forecasts. Below, each variable with its source, kept onsets contributing an in-window observation, the tranquil benchmark, and the median at onset (t0):

Variable	Source	Episodes	Tranquil	Median at t0
Real GDP growth	GMD	142	+3.84%	+2.17%
CPI inflation	GMD	142	+3.85%	+13.10%
Current account / GDP	GMD	141	-2.97%	-4.18%
REER change	GMD	140	+0.32%	-0.37%
Nominal FX depreciation vs USD	GMD	142	+0.00%	+11.76%
BIS credit-to-GDP gap	BIS	18	+0.44pp	+11.19pp

Reading. Emerging-market banking crises look different from the advanced-economy run-up. Real GDP growth holds near its +3.84% tranquil pace, then slides to a +1.43% trough at t1, a shallower median dip than Panel A but over a far wider band. Inflation runs an order of magnitude hotter, near +13.10% at t0 against a +3.85% tranquil level. The current account stays in deficit, widening to -4.18% at t0 (tranquil -2.97%). The real exchange rate appreciates into the crisis (REER +2.98% at t-1, tranquil +0.32%), then gives way to nominal devaluation (+11.76% at t0 and +14.67% at t+1, tranquil +0.00%): the classic twin-crisis signature. The thin BIS credit-to-GDP gap overlay, for the 18 larger emerging markets with quarterly BIS credit data, still spikes to +11.19pp at t0 from a +0.44pp tranquil level.

Not emerging-market only. Panel B is the complement of the 18 JST countries, tilted to middle- and low-income economies but including some non-JST high-income onsets; it is broad coverage, not a pure emerging-market panel.
No credit or house-price line. GMD carries neither, so Panel B cannot show the credit boom or the property cycle for most of its countries; the credit dimension survives only as the 18-episode BIS overlay, and real house-price growth is dropped entirely.
Chronology dependence. Onsets are Laeven-Valencia’s dating; a different chronology would shift some onsets by a year or two and change which cells populate each relative year.
Wide dispersion, thin overlay. The interquartile bands are far wider than Panel A’s, so the median hides large cross-country differences, and the BIS-gap overlay rests on a small sample that should not be over-read.

Comparability. Panel A and Panel B are not directly comparable. They use different chronologies (JST vs Laeven-Valencia), different macro panels (JST vs GMD), different country universes (18 advanced economies vs 104 mostly emerging and developing), different eras, and partly different variables. A level difference between a Panel A line and a Panel B line is as likely to reflect the panel, the chronology, or the era as any real advanced-vs-emerging difference. The two are complementary windows, not a matched comparison.

License.Panel B draws on GMD, the most restrictive input: free for academic and non-profit research only, citation required. Onsets are Laeven-Valencia’s. The BIS gap is shown with a “Source: BIS” attribution. Panel B is labelled non-commercial research use only, matching the GMD license.

Citations: Global Macro Database (Müller, Xu, Lehbib and Chen 2025, NBER Working Paper 33714, release 2026_06, globalmacrodata.com, consulted 2026-07-09); Laeven and Valencia (2026), “Systemic Banking Crises Database: 1970-2025”, IMF Working Paper WP/26/94; Bank for International Settlements, BIS Data Portal (credit-to-GDP gap).

Documented policy responses (Metrick-Schmelzing)

Episode pages carry an optional documented policy responses block drawn from the Metrick-Schmelzing Banking-Crisis Interventions Database (Yale Program on Financial Stability). It records 1,946 individual government and central-bank interventions across 140 countries and three global episodes, 1257–2019, each tagged with intervention-type flags (guarantees, lending, capital injections, restructuring, asset management, regulatory forbearance, and finer sub-types) and a short code, under a crisis code of the form <ISO3>-<year>. It is a separate dataset from the five crisis chronologies above and does not enter the crisis event table.

Matching rule.An intervention is shown on an episode page only when its country (ISO3) matches and its recorded intervention year falls inside the episode window, inclusive. The match is exact, not fuzzy; the database’s global-crisis rows (which carry no country) are never attached to a country episode, and an episode with no matching intervention shows no block. Intervention categories are the database’s own flags; the finer sub-type codes are shown verbatim rather than re-labelled, because the source workbook does not define them in prose.

Citation: Metrick, Andrew, and Paul Schmelzing, “Banking-Crisis Interventions Across Time and Space,” working paper, 2024 (dataset consulted 2026-07-10; also NBER Working Paper No. 29281, 2021, and Review of Financial Studies, forthcoming). License: Creative Commons attribution.

Caveats

The sources are not independent. GMD’s flags are composites of Baron-Verner-Xiong, Laeven-Valencia, JST, and Reinhart-Rogoff, with repeat flags suppressed within three years; JST and Laeven-Valencia also draw on overlapping literatures. Agreement counts overstate independent corroboration.
Sources disagree on dates and classification. Start years for the same episode can differ by a year or more (Thailand’s currency crisis: 1997 in RR, 1998 in GMD and Laeven-Valencia); ESRB systemic episodes include 2020-21 COVID crisis-management windows that other chronologies do not treat as banking crises. The union rule keeps all of them.
GMD forecast values are excluded from crisis evidence. GMD extends its macro series to 2030 with IMF-consistent forecasts merged into the same columns. Crisis flags end in 2017, and panel macro context for years after 2024 is forecast, not observed.
Coverage gaps. Flags end in 2016 (RR), 2017 (GMD, Laeven-Valencia), 2020 (JST), and 2021 (ESRB), so the late 2010s and 2020s are thin. ESRB covers 28 European countries only; JST covers 18 advanced economies; RR covers 70 countries; Laeven-Valencia’s window is 1970-2017. Absence of a flag is not evidence that no crisis occurred, especially outside a source’s window.
Mixed dating semantics. Start-year sources (GMD, JST) never mark continuation years, so multi-year episodes show lower agreement after their first year. Laeven-Valencia’s three “ongoing” banking episodes are expanded to 2017, the database window end, not to their true end dates.

Coverage and license

Crisis flags span 1800–2025 across 165 countries. The combined layer inherits its most restrictive input license (GMD: academic and non-profit research only; JST: non-commercial purposes only) and is published as non-commercial research data. Cite all five sources when reusing it. Full retrieval detail, exact file URLs, and SHA-256 checksums live in the repository's per-source SOURCE.md files.