Measured, published, misses included.

The United States: 552 failures, quarter by quarter

Every FDIC insured bank scored point in time from free call report filings, 2004 to 2026. Weights were fixed before the first backtest ran, so the whole record is out of sample by construction. Two of the four 2026 failures were on the watchlist more than a year before the FDIC arrived. The seven misses of the 2023 to 2026 wave are named: five books concealed by fraud, which balance sheets cannot see, and the two July 2026 microbanks, Kentland Federal S&L and Small Business Bank, which never reached the watch decile and were published as misses the week they failed. We say so either way.

The trade off, stated as the product: roughly ten percent of the industry sits on watch to catch 73 percent of failures one to two years early. Precision 4.1 percent against a 0.83 percent base rate is a 5.0x lift, and we publish the entire confusion matrix because a warning system that never says how often it cries wolf is not a warning system.

527,760 bank quarters · 9,903 institutions · free FDIC data end to end · the full US record →

India: 48 institutions, two decades

Failed institutions, stressed survivors and healthy controls from vetted primary filings, banks and co-operatives next to NBFCs and microlenders, post default quarters walled off so nothing leaks from the future. 16 of 18 non fraud failures flagged at a median lead of 21 months, significant at p < 0.001, discriminating at AUC 0.713 against healthy controls.

Both dials ship. Conservative mode caught 11 of 16 with zero false alarms on healthy books. Sensitive mode adds the funding lens and caught 16 of 18 at the price of three healthy names warned. A second, fully out of sample replay on the 2025 microfinance cycle covered 11 more institutions with no retuning.

every dossier cited to primary filings · the Evidence tab replays it, on request →

Europe: seven case files, no recalibration

Northern Rock to Credit Suisse, replayed through the same lenses with nothing refit to the new market. Deliberately case studies rather than a cohort statistic, because seven collapses do not make a distribution. Every figure was audited against primary filings after the fact: 142 checked, 3 corrected in the open. The corrections stayed in the record, because a record you can quietly fix is not a record.

Northern Rock · Banco Popular · Credit Suisse and four more · the validation record →

Deposit desperation: the liability side talks first

Banks that start reaching for deposits, paying up and growing brokered funding against the tide, are telling you something their asset side has not admitted yet. On the US corpus the reaching for deposits signal alone catches 64.7 percent of all failures. Added to the score, recall rises from 72.0 to 75.5 percent at the same alarm budget, which is the honest way to state a lift: same number of alarms, more failures caught.

study on 552 failures, 2008 to 2026 · served on the US layer →

NDFI exposure: the $1.65 trillion side door

US banks now lend about $1.65 trillion to nondepository financial institutions, private credit funds among them, and that exposure concentrates hard. We rank all 1,052 banks above one billion dollars in assets by NDFI lending against tier 1 capital. The top concentration runs at 5.64 times tier 1. The watch is display only by our own gate discipline: it states risk concentration, it does not yet move a score.

from public call report schedules · the NDFI watch →

Stablecoin reserves: everyone rates the coin, nobody scores the banks holding its cash

In March 2023 USDC broke its peg not because of anything on chain, but because part of Circle's reserve cash sat at Silicon Valley Bank. That is a bank run risk problem wearing a crypto costume. We map which banks hold stablecoin reserve cash from citable public disclosures and join that map to the same failure engine that scores every other bank, for the era in which reserve custody is regulated business.

built from public disclosures only · the reserve custodian map →

The monsoon study: published because it failed

Indian microlender repayments live and die with rural cash flows, so monsoon stress into loan stress is an obvious hypothesis. We ran it as a pre registered lift study on the crisis record, thresholds frozen before scoring. The result came back below the bar, and the gate stayed closed: monsoon data reads on the board as context and does not move any score. Publishing the negative result is the point. A research page with only wins is marketing.

pre registered, thresholds frozen before scoring · the environment layer →

PROOF: forecasts written down before the outcome

The terminal publishes forward event odds on funding stress and then keeps the score, hits and misses both, on a sealed record. Entries are hash chained and anchored externally so the history cannot be quietly rewritten, and the scoreboard is served live on the site. This is the standard we think every early warning vendor should be held to, so we hold ourselves to it first, in public, on the hardest market there is.

free, no sign in · the PROOF scoreboard →

The dispatches: plumbing leads price

Every day the terminal writes a short dispatch on what the plumbing did, and the archive is quietly building a study: documented episodes where funding stress showed in repo operations and money market prints before it showed in price. When the strain tape catches one, the dispatch says so with the data attached. Reading the archive back to back is the closest thing we can offer to watching the thesis prove itself in real time.

every dispatch a static, linkable page · read today's →

The methods tier: uncertainty you can audit

The forecasting stack is built from methods with guarantees rather than vibes: conformal prediction for distribution free coverage with the coverage accounting done per regime, expert aggregation across interval forecasters, a calendar gated Hawkes process on the shock catalog, threshold free AUROC with permutation nulls for significance, and a one switch leakage audit we run against our own pipeline. All of it is in the open repository, so the claim is checkable at the code level.

methods cited to their papers in the repo · read the source →

Harbors: five money markets, one tide

Dollar funding does not live alone. The terminal now watches five harbors, India, China, the euro area, Japan and Korea, from each market's own public prints, and estimates directional spillover between them so a reader can see which market is exporting stress and which is importing it. A dedicated view puts New York and Mumbai side by side, because that is the strait our two products share.

public central bank prints end to end · open the terminal →