A house-price boom with credit behind it is followed by a banking crisis 2.36× as often as one without

What Jordà, Schularick and Taylor claimed

Their paper distinguished asset-price booms that were accompanied by a credit expansion (“leveraged” bubbles) from those that were not. Estimating the path of output and the probability of a financial crisis after equity and housing booms on this same macrohistory panel, they found that a boom on its own carried modest risk, but a boom financed by rapid credit growth was followed by materially worse outcomes, and that credit-fuelled housing booms were the most dangerous category of all, the ones with the deepest and most persistent output losses. The specific coefficients, local-projection paths, and crisis probabilities they estimated are not reproduced here, because this note runs no regression and models no controls; only the qualitative ordering, credit-fuelled housing booms are the dangerous ones, is the object of comparison.

Source: Jordà-Schularick-Taylor Macrohistory Database Release 6 (18 advanced economies): real house prices (hpnom/cpi), real bank loans (tloans/cpi), and the JST banking-crisis flag (crisisJST). The claim is Jordà, Ò., M. Schularick, and A. M. Taylor (2015), "Leveraged Bubbles," Journal of Monetary Economics 76, S1-S20.

What we run it on

The JST panel view jst carries annual real house prices (nominal house prices hpnom deflated by cpi), real bank loans (tloans deflated by cpi), and JST’s own banking-crisis flag crisisJST, for 18 advanced economies over 1870–2020.

• A boom is a stated threshold, not a turning point. A house-price boom is a 3-year window whose average annual real house-price growth is at least 5% a year; the boom is dated to the window’s end year, and windows are kept non-overlapping so one long run is not counted as several booms. This is a simple rule chosen up front, not the paper’s HP-filter and turning-point dating; the 209 booms it finds span 1873–2020 across all 18 countries.
• Leveraged means above-median credit growth. For every boom with loan data, real bank-loan growth is measured over the identical 3-year window. The 200 booms with that data are split at the panel median real-loan-growth across them (7.1% a year at this threshold): the higher-credit half is “credit-fuelled”, the lower-credit half is not. The split is a median cut of the same set of house-price booms, so the two buckets are the same kind of event and differ only in how much credit grew alongside.
• The outcome is JST’s own crisis flag. A boom counts as followed by a crisis if a JST banking-crisis onset (the first year of a crisis episode under the site’s shared episode rule) falls within 1–3 or 1–5 years after the boom. The boom year itself is excluded. Using JST’s flag rather than the merged crisis atlas is deliberate: the boom inputs and the outcome are then dated by the same source, so a chronology offset between sources cannot manufacture or erase a lead.
• The baseline is matched to the window. Each bucket’s incidence is compared to the share of all covered country-years followed by an onset in the same forward window (9.6%, 250/2,614, at 1–3 years; 15.9%, 409/2,578, at 1–5). Booms and baseline years alike are scored only when the full forward window lies inside the country’s crisisJST coverage, so right-censoring at the panel end cannot bias one against the other.

Source: Jordà-Schularick-Taylor Macrohistory Database Release 6 (18 advanced economies): real house prices (hpnom/cpi), real bank loans (tloans/cpi), and the JST banking-crisis flag (crisisJST). Onset construction reuses src/lib/episodes.ts, the one merge rule behind every /crises page.

The headline: the leveraged booms are the ones a crisis follows

On the default 5%-a-year rule, a banking crisis follows 17.0% of credit-fuelled house-price booms within 1–3 years (17 of 100) against 7.2% of lower-credit booms (7 of 97), a 2.36× gap. Stretching the window to 1–5 years the split persists: 23.0% against 14.1%, a 1.63× gap. Against the matched baseline, the credit-fuelled booms run 1.78× the unconditional rate at 1–3 years while the lower-credit booms run 0.75×, at or below it: a house-price boom without credit behind it is, on this panel, no more crisis-prone than an average year.

Source: Jordà-Schularick-Taylor Macrohistory Database Release 6 (18 advanced economies): real house prices (hpnom/cpi), real bank loans (tloans/cpi), and the JST banking-crisis flag (crisisJST). "Credit-fuelled" and "lower-credit" are the halves of the booms above and below the median real-loan-growth; each cell is onsets that follow over booms whose forward window is observable.

Robustness: the ordering is not a threshold artifact

The 5%-a-year boom cut is a choice, so the same count is repeated at a looser and a stricter threshold. Across all 6 threshold-by-horizon cells the credit-fuelled bucket exceeds the lower-credit bucket in 6. The size of the lead moves with the threshold; the ordering is what holds.

Each rule finds a different number of booms: ≥5%/yr yields 200 classifiable booms (median real-loan-growth 7.1%), ≥3%/yr yields 301 classifiable booms (median real-loan-growth 6.3%), and ≥8%/yr yields 108 classifiable booms (median real-loan-growth 8.5%). The median split is recomputed within each rule’s own set of booms.

Source: Jordà-Schularick-Taylor Macrohistory Database Release 6 (18 advanced economies): real house prices (hpnom/cpi), real bank loans (tloans/cpi), and the JST banking-crisis flag (crisisJST). All three rules use the identical window, credit split, onset construction, and matched baseline; only the house-price growth threshold differs.

Where ours differs from theirs

This is a re-examination on a deliberately simpler method, not a replication of their model, and the differences are large enough that no number here should be expected to match the paper’s.

• A threshold rule, not their dating. Jordà, Schularick and Taylor dated booms off an HP-filtered house-price series and its turning points; this note dates them off a fixed 3-year average-growth threshold. The two procedures select overlapping but different episodes, so a boom here need not be a boom there.
• Counts, not a causal model. They estimated local projections with controls and reported crisis probabilities and output paths; this note counts co-occurrence and reports a frequency. The 2.36× gap is an association in the panel, not an estimate that the credit caused the crisis, and it carries no normalcy controls, no confidence interval, and no standard error.
• The housing arm only. Their result spanned both equity and housing booms and found housing the worse of the two; this note tests only house-price booms, so it can confirm the credit split within housing but says nothing about the equity-versus-housing comparison that was half of their finding.
• Annual, and 18 advanced economies. The panel is annual and confined to the 18 long-run advanced economies of the JST database, 1870–2020. Sub-year timing is invisible, and nothing here speaks to emerging markets or to the post-2020 period the flag does not cover.

What this cannot tell you

• The credit split is a median cut. “Credit-fuelled” is defined relative to the other booms in the same rule, at 7.1% real-loan-growth a year on the default cut, not against an external leverage standard. A different split point would move the two buckets’ membership; the median is stated so the reader can see where the line falls.
• The cells are thin. The credit-fuelled bucket rests on 17 crises among 100 booms at 1–3 years; the counts are printed beside every rate because a handful of episodes moves the percentage, and no interval is drawn.
• The baseline is a choice. The 9.6% unconditional rate pools every covered country-year and conditions on nothing but the forward window. It is a historical frequency for this panel, not a model output, and a different denominator would rescale every ratio on the page.
• Coverage edges are handled, not adjusted for. A boom in the last years of a country’s crisisJST coverage is dropped from a horizon when its full forward window would run past the coverage end, exactly as the baseline drops those years, rather than being counted as “not followed”. This keeps the conditional and baseline comparable but shrinks the late-sample cells.
• Non-overlap is a rule, not a fact. Keeping boom windows non-overlapping means a long, sustained property boom contributes a single dated episode rather than several; a rule that allowed overlapping windows would raise the boom count and change the buckets. The rule is stated rather than tuned.