Search any of ~4,700 US banks
Type at least 3 letters. Data: FDIC quarterly Call Reports, 2001–2026.
Every bank failure since 2019, ranked in advance
Run risk score vs. the regulatory Tier 1 risk-based capital ratio vs. a classic credit-failure ML model — all computed as of each quarter shown, using only data that would have been publicly filed by then.
Where things stand today
Loading…
How it works, in plain terms
When interest rates rise, bonds a bank bought earlier lose market value. Banks are allowed to hold some bonds as "held to maturity," which lets them carry those bonds at their original cost on the books — the loss never shows up in the bank's official capital ratio unless the bank is forced to sell.
Deposits above $250,000 are not covered by FDIC insurance. If uninsured depositors get nervous and a large share of them try to withdraw at once, the bank may be forced to sell those bonds — realizing the loss all at once, right when it can least afford to.
Marked's score is uninsured deposit share divided by mark-to-market equity (book equity minus the hidden bond loss). A bank with a lot of uninsured money sitting on capital that isn't really there scores high. This is two accounting inputs from public filings — no machine learning, no fitted weights, frozen before any historical backtest was run (see the repo's git history).
Honest limits
- Fraud and credit failures are invisible to this model by design. Heartland Tri-State Bank (2023) failed from fraud, not a run — this score didn't flag it, and shouldn't be read as a general failure predictor. The classic ML baseline (trained the standard way) catches credit failures like Citizens Bank (Sac City) instead.
- Banks under roughly $1B in assets don't report uninsured deposits to the FDIC in a way this API exposes. They're shown as "not reported," never treated as safe.
- Across all banks, the 2022Q3 score did not predict which banks lost the most deposits in the Q1 2023 panic (AUC ≈ 0.49, no better than chance) — small-bank deposit swings are seasonal and noisy. Restricted to the ~1,000 banks that report uninsured deposits, it does (AUC ≈ 0.64 vs. 0.39 for the regulatory ratio). The claim is scoped accordingly, not to "every bank."
- First Republic's losses were concentrated in low-rate mortgages, not bonds. This model only marks bond losses, so First Republic likely ranks lower than its true risk. We checked whether loan-level repricing data was available from the free FDIC API to fix this and it isn't — documented as a known gap, not silently patched.
- This is a research prototype, not financial advice, not a regulatory tool, and not a guarantee about any bank. It estimates exposure from public accounting data; it does not know about a bank's off-balance-sheet hedges, parent-company support, or private information.