US bank-failure nowcast: a transparent model, honestly backtested

The honest headline: high AUC, and a single ratio nearly matches it

Pooled over 422,098 labeled bank-quarters (2008-2025, 1,918 of them within four quarters of an actual failure), the model’s AUC is 0.950. The benchmark, ranking banks by nothing but equity/assets, scores 0.951: the single ratio is marginally better pooled. Most of what any call-report model knows about bank failure is already in the capital ratio; the model’s value is in the tails and in years where earnings, growth, and funding mattered too. Across the four report dates preceding an actual failure, 46% of those bank-quarters sat in the top 1% of their quarter’s scores and 82% in the top 5%.

Pooled AUC mixes score scales across 18 differently-trained yearly models. That is harmless for the logits (near-probability scale) but understates the gradient-boosting model, whose balanced-weight scores shift scale year to year; its per-year AUCs below are the meaningful comparison. The 8-feature model pools over 2019-2025 only.

The 2023 failures: what the model actually said

The practical question for any failure model is 2023. The answer, read verbatim from the out-of-sample scores, is that the model did not flag Silicon Valley Bank, Signature, or First Republic. On the 2022Q4 call reports, the last filed before the March runs, SVB scored in the 44th percentile of all banks, Signature the 56th, and First Republic the 16th, dropping to the 7th on its final 2023Q1 filing. None was near the top-1% or top-5% buckets that catch most credit-driven failures. These three failed from interest-rate and uninsured-deposit-run risk, concentrations that the eight call-report ratios do not measure; their capital, earnings, and credit quality looked ordinary to strong. The 7-feature model, which ignores credit quality, saw them as elevated but still unflagged (SVB at the 88th percentile, Signature the 84th, First Republic the 79th), and adding the nonperforming-loans feature pushed all three down: their loan books were clean, which is exactly why a credit-quality model missed a duration-and-run failure. The one adjacent name the model did flag is Silvergate (98.5th percentile in 2022Q4, 100th in 2023Q1 as its balance sheet collapsed), which wound itself down voluntarily in March 2023 and is therefore not on the FDIC failures list at all. The same limitation appears in the FinObservatory Composite’s SVB backtest; publishing it is the point.

Scores from the walk-forward models for test years 2022 and 2023 (trained only on data whose outcomes were known before those years). Washington Mutual and IndyMac (2008) never appear in this panel: as OTS-regulated thrifts they filed Thrift Financial Reports, not FFIEC call reports; they are counted among the 54 unmatched failures below.

Out-of-sample AUC by test year

Each row: models trained only on observations whose full four-quarter outcome window closed before that year began, then scored on that year’s filings. 2021 has no row entries because no bank that filed in 2021 failed within four quarters. Years with few failures carry wide uncertainty; they are reported anyway.

The 8-feature model exists from 2019 (it needs the RC-N nonperforming fields, first published in the 2017Q1 bulk files, for training). Its worst year is 2022, AUC 0.360: the 2023 failures had clean credit books, so the nonperforming feature actively pointed away from them.

The current quarter, as a distribution

Below is the full score distribution for 2026-03-31, published deliberately without names: a public percentile is a ranking of ratio vulnerability, not an accusation, and this page will not manufacture a watch-list. Counts by band, state, and size show where the tail sits.

Methodology