LiquiLens US scores every FDIC insured bank for run risk from free public filings, point in time, quarter by quarter since 2004. It knows who is fragile, watches when the reaching for deposits starts, reads the funding tide from our open Seiche terminal, and maps where trouble travels, from stablecoin reserves to private credit. The validation is published in full, false positives and misses included.
of all 552 FDIC failures since 2008 caught in advance
median warning before the FDIC arrived
false positive rate, published next to the recall
balance sheet failures of the rate shock era flagged early, both 2026 misses published by name
A warning system that never says how often it cries wolf is not a warning system. Incumbent bank raters publish hit rates and stop there. We publish the entire confusion matrix on all 527,760 bank quarters, because the trade off is the product: we place roughly ten percent of the industry on watch to catch 73 percent of failures one to two years early.
| failure within 24m | no failure | |
|---|---|---|
| flagged | 2,179 caught | 50,637 false alarms |
| not flagged | 2,209 missed | 472,735 correctly clear |
| false positive rate | 9.7% |
| precision | 4.1% |
| failure base rate | 0.83% |
| lift over base rate | 5.0x |
| pooled AUC | 0.854 |
| recall, 95% interval | 69.0% to 76.6% |
Failures are rare, so most flags are not followed by one. That is what a 0.83 percent base rate means, and any vendor whose numbers sound better than this either fit their model to the past or is not telling you their false positive rate. Score versions are kept honest too: the preregistered score reads 72.8 percent, adding private credit exposure reads 72.0 percent, adding deposit desperation reads 75.5 percent, all at the same alarm budget.
every balance sheet driven failure of the rate shock era, from public filings alone, misses shown at zero
Nothing in this pipeline was fitted to any outcome, in any era. The weights were frozen before the first backtest ever ran, so the same rule set that catches the 2008 failures catches the 2026 ones. From here the record runs forward: every quarterly watchlist is sealed when published, the same discipline as our open Seiche terminal.
Five misses of the 2023 to 2026 era are small banks whose failures were publicly attributed to fraud or insider conduct. Embezzlement is invisible to balance sheet fundamentals; this engine predicts run risk and solvency erosion, it does not detect crime. The two July 2026 misses are different, and harder to write. Both were tiny non fraud banks, $3.7M and $73M in assets, whose capital eroded while their deposit base stayed quiet, and a run risk lens ranks that low. We replayed both on their final filings the week they failed, found neither had ever reached the watch decile, and published exactly that instead of an excuse.
Loans to nondepository financial institutions, the private credit and fund finance pipeline, are the fastest growing corner of US bank lending, and the one regulators started forcing into disclosure in 2024. Analysts cover the twenty largest banks by hand. This engine ranks every one of the 1,052 banks above one billion dollars in assets, every quarter, against their own capital. The ten most concentrated, straight from Q1 2026 filings:
| Institution | State | Assets | NDFI loans | NDFI / tier 1 | Uninsured deposits |
|---|---|---|---|---|---|
| Northpointe Bank | Michigan | $7.4B | $3.9B | 5.64x | 8% |
| Axos Bank | California | $28.2B | $10.1B | 3.96x | 19% |
| EverBank | Florida | $47.2B | $16.4B | 3.96x | 13% |
| Stifel Bank | Missouri | $12.1B | $3.3B | 3.90x | 15% |
| Merchants Bank of Indiana | Indiana | $20.3B | $7.9B | 3.50x | 27% |
| Texas Capital Bank | Texas | $33.2B | $10.1B | 2.72x | 41% |
| Customers Bank | Pennsylvania | $25.9B | $6.0B | 2.57x | 43% |
| Forbright Bank | Maryland | $8.3B | $2.0B | 2.42x | 14% |
| Apple Bank | New York | $19.2B | $3.2B | 2.17x | 22% |
| Goldman Sachs Bank USA | New York | $751.8B | $122.8B | 1.96x | 42% |
In March 2023 USDC broke its peg because Circle's cash sat at Silicon Valley Bank. The GENIUS Act now regulates issuers like banks, but it sets no limit on where the cash reserves sit, and the New York Fed itself notes issuers are not required to name every bank servicing their deposits. So we map the edges: issuer by issuer, bank by bank, only from citable public disclosures, joined to the same run risk engine that scores every US bank above.
| Issuer | Token | Named institution | Relationship | NDFI / tier 1 |
|---|---|---|---|---|
| Circle | USDC | Bank of New York Mellon | reserve fund custody | 0.40x |
| Circle | USDC | Customers Bank | cash deposits | 2.57x |
| Circle | USDC | Cross River Bank | cash deposits | 1.50x |
| Tether | USDT | Cantor Fitzgerald | treasuries custody, non bank | · |
| SoFi | sofiUSD | SoFi Bank, N.A. | bank issued | 0.00x |
| JPMorgan | JPMD | JPMorgan Chase Bank, N.A. | deposit token | 0.81x |
Every edge above carries a citable public source: issuer transparency disclosures, SEC money market fund filings, bank quarterly reports. Where an issuer has not named its banks, we say undisclosed rather than guess. Nothing here alleges wrongdoing by any institution.
Reaching for deposits is one of the oldest tells in banking and one of the best documented in the academic record. As a quarterly signal from filings alone it catches 64.7 percent of all 552 failures by itself, and adding it lifts the full engine from 72.0 to 75.5 percent recall at the same alarm budget. The live layer sharpens it to weeks: every bank's offered rates against the FDIC national baseline, where a legal ceiling already exists for banks that are less than well capitalized.
FDIC national deposit rates and section 337.7 rate caps, as of June 15, 2026. The caps mark national rate plus 75 basis points, the legal ceiling for banks that are less than well capitalized. An institution bidding far above its peers, and rising fast, moves up this engine's watchlist weeks before the next quarterly filing lands. Deposit pricing vendors sell this data to banks to help them price. Nobody sells it as a distress signal, because their customers are the banks. Ours are not.
Scores here are conditioned on the live funding regime read by Seiche, our free and open source terminal for US money market stress. Incumbent risk vendors condition on hypothetical scenario menus. Seiche infers the regime that is actually in force, right now, from the plumbing: repo, reserves, the Fed's own balance sheet prints. It publishes its calls before outcomes and keeps a public scoreboard, misses included.
A score at any date uses only filings available at that date. Signal weights were fixed before the first backtest ran and never tuned on outcomes. Post hoc additions, private credit exposure and deposit desperation, are validated standalone and reported as separate score versions.
Five of the seven 2023 to 2026 misses were fraud driven failures; balance sheet fundamentals cannot see embezzlement, and we will not pretend otherwise. The July 2026 microbank misses state a second limit: slow solvency erosion in very small banks with a quiet deposit base ranks low on a run risk lens. A conduct signal and a dedicated solvency lens are roadmap items, not tuning fixes.
The Fed watches actual payment flows in real time. Nobody outside can. Public data nowcasting approximates run pressure rather than observing it, and the honest word for that is proxy.
Uninsured deposit estimates have a reporting gap for smaller banks between 2015 and 2019. The score reweights over available signals and records how many were present for every bank quarter.
LiquiLens US outputs are screening percentiles computed from public FDIC filings. They are research, not credit ratings, not a prediction that any institution will fail, and not investment, legal or deposit placement advice. Named institutions are identified because their filings are public records; nothing on this page alleges wrongdoing by anyone. Data: FDIC BankFind Suite, public domain. Validation reproducible end to end; the full report travels with the product.
On April 17, 2026 the Federal Reserve, OCC and FDIC replaced SR 11-7 with SR 26-2, revised model risk guidance effective immediately. It drops the fixed annual validation ritual and asks for something harder: risk based tailoring, outcomes analysis against your own book, a documented validation cadence and effective challenge, with vendor models held to the same principles as internal ones. That is the shape this page already has. Every LiquiLens engine ships an SR 26-2 ready validation pack: design and intended use, limitations stated plainly, data lineage, the pre registered gate studies as a prospective validation protocol, and the sealed publication ledger as the outcomes analysis audit trail your examiner can replay.
Fintechs watching a sponsor bank, treasurers sizing uninsured exposure, credit desks screening counterparties, stablecoin risk teams. One engine, published validation, misses included.