Failing banks still look alike years in advance, until 2023

What Correia, Luck and Verner claimed

Their paper builds the longest micro-level panel of US bank balance sheets to date, 1863 through 2024, and documents that failing banks show the same signature across 160 years: rising asset losses, deteriorating solvency, and an increasing reliance on expensive noncore funding. From that signature, bank failures are highly predictable using simple accounting metrics from publicly available financial statements, years before the failure date, and deliberately simple models (logistic regressions on a handful of ratios) do most of the work. The staff-report version states in-sample AUCs for the full specification from 0.86 in the historical sample to 0.95 in the modern sample. Because weak fundamentals precede failures with and without depositor runs, the authors read failures as mostly fundamental insolvency rather than pure coordination failure.

Source: Correia, S., S. Luck, and E. Verner (2026), "Failing Banks," Quarterly Journal of Economics 141(1), 147-204 | NY Fed Staff Report 1117 (2024 version)

What we run it on

The panel is the union of the estate's two FDIC financials files: 1,131,818 bank-quarters for 17,362 banks, report dates 1992-03-31 through 2025-12-31, disjoint on (CERT, report date) by construction and asserted in the build. The outcome file is the FDIC's Failures and Assistance Transactions list: 847 resolutions typed FAILURE with a failure date in 1992-2025 (15 ASSISTANCE transactions excluded). Of those, 789 (93.2%) match a call report in the panel before the failure date. Every one of the 58 unmatched failures dates to 1992: the panel's first report date is 1992-03-31, so banks seized in early 1992 never filed a report inside it. From 1993 on, the match is 668 of 668, complete.

Eight features approximate the paper's ratio set from the fields this panel carries: equity/assets, the FDIC's annualized return on assets, noncurrent loans and loss reserves as shares of gross loans, brokered deposits over assets (the noncore-funding proxy), the interest expense yield on earning assets (the expensive-funding proxy), year-over-year asset growth (constructed by an exact-date self-join, so it requires a report exactly four quarters earlier), and log assets. Three of the paper's ingredients cannot be built from these fields and are simply absent: surplus and undivided profits as separate capital components (their historical national-bank items), time deposits and the interest expense on deposits specifically (so noncore funding here is brokered deposits only, and expensive funding is the all-in yield), and loans past due 30-89 days (the noncurrent field is 90-plus days and nonaccrual). Rows missing any feature are dropped, leaving 1,060,737 complete cases from 1993-03-31; the asset-growth requirement costs the whole first year of the panel, and with it the pre-failure quarters of the 121 matched 1992 failures.

The outcome is failure within 12 quarters of the report date, the horizon that carries the paper's years-in-advance claim, with failure within 4 quarters computed alongside. Failures are observable through 2026-05-01, so a bank-quarter enters a horizon's evaluation only when its full window fits inside the observed record: no censored negative is ever scored, and no partially-windowed positive either. The model is a plain logistic regression on standardized features, effectively unpenalized, matching the paper's deliberate simplicity. Discipline is out-of-sample twice over: train on 1992-2006 and predict the 2007-2012 wave, then train on 1992-2019 and predict 2020-2025. The full-sample in-sample fit is reported for comparison with the paper's own in-sample numbers, and for nothing else.

Source: FDIC BankFind Suite financials (fdic_financials_hist, fdic_financials) | FDIC Failures and Assistance Transactions (fdic_failures) Build: scripts/build/build_research_failing_banks.py; every AUC is recomputed inside the build by an independent rank-based implementation and asserted to match sklearn to three decimals.

Predictability: strong through 2012, weaker since 2020

The headline test is the harshest available: fit the model on 1992-2006, a training window that contains only 321 four-quarter failure positives, then score the global financial crisis. It works: AUC 0.947 at four quarters and 0.855 at twelve, on 1,928 and 5,516 positive bank-quarters respectively. The same exercise moved forward, trained on 1992-2019 and tested on 2020-2025, drops to 0.789 and 0.788, with wide standard errors (0.043 and 0.029) because the test window contains only 40 and 87 positives. That drop is the note's second finding, reported as it comes out: the recent failure cohort, dominated by the 2023 banks, fits the historical accounting signature far less well than the 2008 cohort did.

AUC of a logistic regression on the eight standardized ratios; SE is Hanley-McNeil. Evaluation windows end where the failure record ends (2026-05-01): the 12-quarter split-b window closes at 2023-03-31, the 4-quarter at 2025-03-31. "In-sample" fits and scores the same rows and is shown only against the paper's in-sample 0.86-0.95 range.

Which ratios carry the model, full-sample, 12-quarter horizon, in standardized units: equity/assets -1.07, return on assets -0.67, noncurrent loans +0.46, brokered deposits +0.41, log assets +0.21, loss reserves -0.27. Solvency and asset quality dominate, exactly the paper's ordering, with the noncore-funding proxy the strongest positive after them. Two features add almost nothing conditional on the rest: the expensive-funding yield (+0.02, though its unconditional gap against controls is visible in the paths below) and asset growth (-0.001, whose boom-then-bust shape the levels already absorb).

The signature: what the median failing bank looks like

The paper's central exhibit tracks failing banks' ratios in event time. Here is the same picture for the 789 matched failures: the median value of each ratio over the last 12 call reports before failure (failure at the right edge), against the median of matched controls, banks in the same calendar quarter and the same within-quarter size decile that did not fail in the following 12 quarters. Three years out the median future failure already reads differently: equity 8.8% of assets against 9.7% for controls, noncurrent loans 1.5% against 0.6%, brokered deposits 4.5% of assets against a control median of 0.0%, and assets growing 10.9% a year against 6.2%. By the last report the collapse is total: equity 1.4%, return on assets -4.4%, noncurrent loans 11.9%, assets shrinking 13.6% a year. That is the paper's sequence on this panel: boom, expensive noncore money, then solvency rot.

One wrinkle the paper's mechanism predicts: brokered deposits peak years before failure (4.5% of assets at 12 reports out) and then fall to 0.5% by the last report. That fall is not health. Section 29 of the FDI Act restricts brokered-deposit acceptance once a bank is no longer well capitalized, so the funding runs off exactly as the bank weakens; the interest expense yield stays elevated against controls to the end (2.7% vs 1.8% at the last report). The noncore-funding flag is a leading indicator here, not a coincident one.

2023: what the accounting ratios saw, and what they missed

The table shows every 2020-2025 failure with the percentile of its predicted failure probability among all banks filing the same quarter, under the model trained through 2019 (12-quarter horizon), at the bank's last call report before failure. These are historical, realized failures; this site publishes no live per-bank failure probabilities, the same rule the composite backtest follows. 10 of the 13 sat above the 90th percentile at their final report, Silicon Valley Bank (97.3), Signature (97.8) and First Republic (98.2) included. 2 sat below the median: Pulaski Savings Bank (32.6) and Santa Anna National Bank (39.8). Their final reports simply looked ordinary on these eight ratios; this note reports the ratios and does not assign failure causes.

Percentile of the split-b (12-quarter) predicted probability among all complete-case banks filing the same quarter. Green: 90th percentile or above at the last report; red: below the median. Assets are from the FDIC failures file at resolution.

The Silicon Valley Bank row deserves the decomposition, because its headline percentile flatters the model. At 2022-12-31, 128 of 4,739 filing banks scored higher, putting SVB at the 97th percentile. But in standardized log-odds contributions, 1.18 of that came from sheer size (log assets) and 0.48 from equity at 7.39% of assets, below the panel median, while the classic failing-bank markers pointed the other way: noncurrent loans contributed -0.23 (its loan book looked pristine at 0.19%) and brokered deposits -0.07 (it reported 0.00%). The ratios that made 789 historical failures predictable, credit losses and noncore funding, were absent at SVB; what its balance sheet actually carried, held-to-maturity securities losses and an uninsured-deposit concentration, sits outside these eight fields entirely. The model's rank and the model's reasons disagree, and the reasons are the finding.

Against the site's own composite backtest

This note complements the composite score backtest, which scores the same failure file with the FinObservatory composite rather than a fitted model. The two agree on the base pattern: failures carry visibly worse public ratios well before resolution (the backtest's median pre-failure composite is 20.7 against a panel median of 50.5; here, out-of-sample AUCs up to 0.947). They also agree on the exception: the composite missed SVB outright at 63.5 (8 of its 32 failures since 2016 scored 35 or better at the last report), and this note's logit ranked SVB high only for reasons unrelated to the failing-bank signature. The shared lesson, stated on both pages: public call-report ratios captured the 1992-2012 failure mechanism almost completely, and captured the 2023 mechanism poorly, because held-to-maturity losses and uninsured-deposit concentration are not among the ratios. The backtest's flag threshold (20) and its false-positive base rate live on that page; this note adds the formal out-of-sample discrimination numbers the composite page deliberately does not compute.

Where ours differs from theirs

• 33 years, not 160. Their panel spans 1863-2024 across multiple institutional regimes; this one is 1992-2025 with 789 matched failures, almost all from two waves (the early-1990s tail of the S&L era and 2008-2013). Nothing here tests their historical claims, the national-bank era, or failures before deposit insurance.
• Proxied features. Noncore funding is brokered deposits only (no time-deposit split exists in these fields), expensive funding is the all-in interest expense yield (not the deposit rate), and surplus/undivided profits are unavailable. The replication tests the signature, not their exact variable definitions.
• Pooled logit, their simplest specification. The paper also reports richer exercises (failure decompositions, recovery rates on failed assets, run analysis). This note replicates only the predictability core, which is the part the estate's data can support.
• FDIC failure dates, FDIC ratios. The outcome is an FDIC resolution typed FAILURE. Open-bank rescues, forced mergers and voluntary closures are negatives here; the paper's historical failure definition is broader in places.

What this cannot tell you

• The 2020-2025 AUCs are imprecise. With 40 positives at the 4-quarter horizon and 87 at 12, the Hanley-McNeil standard errors are 0.043 and 0.029. The drop from the 2007-2012 numbers is large relative to those errors, but the point estimates should not be read to three digits of meaning.
• Percentiles are not probabilities. A rank near the top of a quarter still implies a small absolute probability when only 0.7% of complete-case bank-quarters are followed by failure within 12 quarters. The table ranks; it does not forecast, and no live per-bank number appears anywhere on this site.
• Right-censoring trims the modern test. Failures are observed through 2026-05-01, so the 12-quarter evaluation ends with reports filed by 2023-03-31. Failures after that observation edge, should they come, could move the split-b numbers either way.
• Complete-case selection. Requiring a four-quarter asset-growth lag drops new banks' first year and the panel's 1992 vintage, 71,081 bank-quarters in all. De novo banks fail young often enough that this could shade the AUCs in either direction; the paths exhibit, which does not require the lag except in its growth panel, is unaffected.
• Prediction is not explanation. That accounting ratios rank future failures well is the paper's claim and it replicates; why banks reached those ratios, and whether supervision should have acted earlier (the authors' companion piece, "Supervising Failing Banks," takes that up), is outside what this exercise can say.