The debt service ratio’s short-horizon crown does not survive this sample

The spec, and whose choices are whose

Two indicators, one panel. The DSR is the BIS’s published debt service ratio for the private non-financial sector, a level in per cent (debt service payments over sector income). The credit gap is the BIS’s published credit-to-GDP gap, the ratio’s deviation from a one-sided HP trend in percentage points. The panel keeps only quarters where both are non-null: 3,363 of the 3,444 published DSR quarters (81 dropped), 32 economies, 1999-Q1 to 2025-Q4. The BIS publishes this DSR for 32 economies and the merged panel keeps 32 of them.

From the paper: the 20-quarter (five-year) evaluation window, the within-two-years reading of h 1-8, and AUC as the metric. Ours, because the estate forces them: crisis dates come from Laeven-Valencia banking-crisis start YEARS (the source is annual), anchored at Q1 of the start year; quarters inside crisis years plus 8 quarters after the end year are excluded from the tranquil negatives; and, as the one pre-declared robustness variant, the DSR minus its one-sided expanding country mean (no look-ahead), a proxy for the paper’s country-level adjustment. AUCs are Mann-Whitney with Hanley-McNeil standard errors, each horizon’s pre-crisis quarters against the same 2,704 tranquil quarters.

The only crises this panel can see

The merged panel begins in 1999-Q1. Of the 37 Laeven-Valencia banking crises recorded for these 32 economies, 23 (onsets 1977 to 2000) have zero pre-anchor quarters in the panel and cannot be tested. The 14 that remain all date to 2 in 2007 and 12 in 2008: one clustered global event, not 14 independent draws.

Source: BIS, debt service ratios for the private non-financial sector | BIS, credit-to-GDP gaps | Laeven and Valencia, Systemic Banking Crises Database, 2025 vintage Anchor = Q1 of the Laeven-Valencia start year (our convention; the source is annual).

The 23 untestable onsets: ESP 1977, MEX 1981, TUR 1982, THA 1983, USA 1988, BRA 1990, FIN 1991, HUN 1991, NOR 1991, SWE 1991, POL 1992, IND 1993, BRA 1994, MEX 1994, CZE 1996, IDN 1997, JPN 1997, KOR 1997, MYS 1997, THA 1997, CHN 1998, RUS 1998, TUR 2000. Their crisis-year and post-crisis quarters are excluded from the negatives, not counted as misses. RUS 2008 is only partially covered: its merged series starts 2005-Q2, leaving 11 of the 20 pre-anchor quarters.

The race, horizon by horizon

At h = 1, the quarter just before the anchor and the closest this panel gets to the event the DSR is supposed to own, the credit gap’s AUC is 0.751 against 0.702 for the demeaned DSR and 0.633 for the raw level. Pooling the within-two-years window (h 1-8): gap 0.717 (SE 0.028), demeaned DSR 0.620 (0.029), raw DSR 0.588 (0.029), on 112 pre-crisis against 2,704 tranquil quarters. At h 9-20: gap 0.717 (0.023), raw DSR 0.566 (0.024), demeaned DSR 0.518 (0.024).

Source: BIS, debt service ratios for the private non-financial sector | BIS, credit-to-GDP gaps | Laeven and Valencia, Systemic Banking Crises Database, 2025 vintage Raw DSR is a level in per cent; the demeaned DSR is percentage points relative to its one-sided expanding country mean; the gap is percentage points of the credit-to-GDP ratio relative to a one-sided HP trend. AUC is unit-free, so the three columns are comparable; the inputs are not.

What does reproduce

The paper characterises the DSR as a short-fuse indicator, weak far from the event and sharpening into it. That shape is here. Beyond the two-year window the demeaned DSR’s AUC drifts between 0.471 and 0.554, around coin-flip. Inside it, the AUC rises at every step from 0.535 at h = 8 to 0.702 at h = 1. The failure is relative, not absolute: in the within-two-years window the demeaned DSR ends 0.10 of AUC below the gap.

Three data reasons the gap likely wins here

1. Left-censoring leaves a single crisis wave. Every testable crisis here dates to 2007 or 2008; the USA credit gap stood at 11.6 points in 2007-Q4. The paper’s authors constructed their own DSR series back to 1,980; 22 of the 23 onsets this panel cannot test fall inside the span of a series starting then.
2. Annual crisis dating clips the DSR’s best quarters. The Q1 anchor can mislabel true horizons by up to 3 quarters, and a dating error costs a short-fuse indicator more than a slow-moving one. The USA DSR reaches its panel maximum of 18.5 per cent in 2007-Q3 and holds it in 2007-Q4, 3 quarters after our 2007-Q1 anchor, so the DSR’s loudest readings sit inside the excluded crisis window rather than in the pre-crisis sample.
3. The level adjustment is a proxy on a short history. The paper adjusts DSRs for country-specific levels using its long constructed histories. Our expanding mean has 8 years of history at the first testable onset. It still helps inside the two-year window (0.620 demeaned vs 0.588 raw) and hurts beyond it (0.518 vs 0.566).

Why the raw level needed adjusting at all

Raw DSR levels are not comparable across countries. Over tranquil quarters, NLD averages 31.6 per cent and IDN 4.1 per cent, a spread of 27.5 points. A pooled comparison of raw levels carries that country difference into every crisis-vs-tranquil contrast; the demeaned variant exists to remove it before pooling.

What this cannot tell you

• Whether the difference is significant. The Hanley-McNeil SEs treat the 112 short-window pre-crisis quarters as independent. They are 14 crises’ quarters, up to 8 stacked per crisis, and the crises themselves are one global wave. Every SE on this page is optimistic, and no significance is claimed for the gap-vs-DSR difference.
• Whether the paper is wrong. The paper’s authors built their own DSR series back to 1,980 for 26 economies and dated crises quarterly. This panel starts in 1999-Q1, dates crises to a year, and its 14 testable events all fall in 2007 or 2008. A replication on a sample that cannot see any crisis before 1999 tests the indicator on this era, not the paper’s inference on its own data.
• Anything at the paper’s dating precision. The Q1 anchor is ours, forced by the annual source, and it shifts true horizons by up to 3 quarters in a direction that systematically handicaps the DSR (see reason 2 above).
• Out-of-sample warning value after 2025. The tranquil sample runs through 2025-Q4; the paper appeared in 2,014, so the years after that are new to this test. The last banking-crisis onset Laeven-Valencia records for these economies is 2008, so those added years contribute tranquil quarters only.
• Full BIS coverage. The panel keeps only quarters where both indicators are non-null: 3,363 of 3,444 published DSR quarters. The latest-starting economy, BRA, enters in 2006-Q1.
• A like-for-like unit comparison. The raw DSR is a level in per cent of sector income; the gap is already a detrended quantity in percentage points. The demeaned DSR exists to soften that asymmetry and is labelled a proxy, not the paper’s adjustment.