Methodology

Overhang reads the quarterly regulatory call report of every FDIC-insured bank under $100 billion in assets, measures each one against the two commercial real estate concentration criteria in the 2006 interagency guidance, allocates each bank to the counties where it takes deposits, and publishes the result for the nine Northeast states. Everything on this page is the method. The data behind it is not published in full, but the method is, including the parts that did not work.

Universe and the $100B ceiling

An institution enters the universe in a quarter when it is FDIC-insured and reporting, holds less than $100 billion in total assets, carries a loan book of at least 20% of assets, and is not a foreign bank agency or an uninsured institution. Inactive institutions stay in the history so that a merger does not open a hole in a county's series.

The 20% loan test removes custodians and trust banks, which take deposits but barely lend. In the Northeast it excludes Bessemer, Mizuho USA, Sumitomo Mitsui Trust, Wilmington Trust, Bank of Utica, Stafford Savings and Monet. Their deposits would otherwise sit in the denominator of every county they touch and depress the exposed share for reasons that have nothing to do with real estate.

The ceiling is about business model, not size. A regional under $100 billion — Webster, Valley National, Flagstar — takes deposits and lends in the same territory, so the Summary of Deposits distributes its exposure to roughly the right places. Above that line the super-regionals fund themselves nationally, the consolidated head office deposit dominates, and the proxy stops describing geography at all.

Four ceilings were tested against Q2 2026 data. The table is recomputed every quarter alongside the figures it sits next to.

At $10 billion, half the region loses its grade, including the whole Boston-to-Philadelphia corridor, and the ranking ends up talking about rural Pennsylvania to a reader in Greenwich. At $250 billion the coverage looks better and the allocation gets worse, because the banks added are exactly the ones whose deposits do not sit where their lending does.

In public text the universe is always written as “deposits held by banks under $100B in assets”, never “community bank deposits”. A bank with $80 billion in assets is not a community bank.

Metrics per institution

From Schedules RC-C, RC-R, RC and RC-N of the call report, per bank per quarter:

The guidance asks for nonfarm nonresidential not occupied by the owner, which is a different field from the nonfarm nonresidential total. Using the total instead moves a single Pennsylvania bank from 394% to 671%, so the distinction is not cosmetic. Where the owner-occupied split is not reported for an early quarter, the total is used and the row is marked.

Noncurrent ratio, annualised net charge-off rate and allowance coverage are computed the same way and feed the deterioration term of the score.

The two criteria

Criterion 2 is conjunctive. Both halves must hold. A bank at 525% of capital in commercial real estate that has stopped growing does not meet criterion 2, and screeners that treat the 300% line on its own report it as flagged when the regulator would not. Overhang reproduces the criteria as written and handles the stalled-but-concentrated case with a separate state rather than by bending the test.

A bank with fewer than twelve quarters of history has no three-year growth figure. Criterion 2 is not evaluated for it, and its pages say so.

The four bank states

“Concentrated” exists because the guidance on its own hides half the picture. Mature concentration that has stopped growing is a different kind of exposure from concentration that is still building, not an absence of exposure, and the product publishes both. A bank can meet more than one of these conditions at once; county counts count it under each, while a bank row shows the strongest one.

Denominator and the CBLR

The denominator is total risk-based capital when the bank reports it. Banks that have elected the Community Bank Leverage Ratio stop reporting risk-based capital entirely; for those the denominator is tier 1 capital plus the allowance for credit losses.

That is not a workaround. It is the definition the call report itself uses: Memorandum item 13 of Schedule RC-C asks CBLR banks to test their construction concentration against tier 1 plus the allowance. The choice was checked against Meridian Bank's Q2 2026 10-Q, where it lands within 0.7% of the capital the bank computes internally.

Treating a missing risk-based capital figure as zero, or silently substituting equity, produces ratios that are wrong by a wide margin for roughly a third of the banks in the region. The CBLR requirement dropped from 9% to 8% of leverage ratio on 1 July 2026, so more banks will elect it from Q3 2026 onward. When a bank switches regime its denominator changes and its ratio moves without its exposure moving. Those quarters are flagged in the data and the affected rows say so for subscribers.

Allocating a bank to a county

Deposits come from the FDIC Summary of Deposits, an annual branch-level census taken on 30 June. The denominator is national: a Pennsylvania bank with a Florida branch has that Florida deposit in the divisor, so its Northeast weights sum to less than one. The Q2 2026 figures use the 2025 Summary of Deposits.

The head office deposit cap

The Summary of Deposits books deposits to the branch that holds the account, and a bank's head office ends up holding everything that never walked into a branch: online deposits, brokered funds, corporate and national accounts. Left alone, that turns a head office county into a false concentration.

For a bank with ten or more branches, when the head office reports more than five times the median of its other branches, the head office figure is replaced by that median for the purpose of weights and coverage. The original figure is kept beside it and the row is marked. In the 2025 Summary of Deposits this affects 590 banks and removes $2.16 trillion from the national weighting.

Connecticut and the Stamford case

The Census replaced Connecticut's eight counties with nine planning regions in 2022, and the FDIC switched conventions between the 2022 and 2023 Summary of Deposits. Overhang uses the planning region names throughout, never “county”.

Connecticut is allocated by branch coordinate in every year, by point-in-polygon against the nine planning regions, with a town lookup for branches without a coordinate. The reason is specific: the FDIC's own county field is wrong in blocks there. Tested against 880 Connecticut branches with Census ground truth, point-in-polygon matched 880 of 880 while the 2023 Summary of Deposits missed 43 — all 36 Stamford branches assigned to Greater Bridgeport when Stamford is in Western Connecticut, and seven New Milford branches assigned to Northwest Hills when New Milford is also Western Connecticut. The FDIC corrected both in the 2025 file. The 2023 and 2024 files are still wrong at the source, and the coordinate layer repairs them.

National geographic validation

Outside Connecticut the FDIC county field is used, after being checked against Census geography. Of 76,084 branches in the 2025 Summary of Deposits tested by point-in-polygon against the full-resolution TIGER 2024 boundaries, 75,890 agreed — 99.75%.

Of the 194 disagreements, about half sit within 500 metres of a county line, inside the error of the geocode itself. The rest are towns split by a county line that the FDIC assigns whole to the dominant county: Milford in Delaware, Katy and Frisco in Texas. Thirteen branches have coordinates that fall in no county at all, several with latitude and longitude transposed and one New Jersey branch geocoded in England; those are marked and never used for point-in-polygon. Retired county codes still alive in the file — Naugatuck, Saipan, Valdez-Cordova, the old Miami-Dade — are resolved by coordinate and each use is logged.

Exposed deposit share, strict and broad

The strict share is the fraction of a county's in-universe deposits that sit in banks meeting either criterion. It is the headline and the only exposure term in the score, because it is the test the regulators themselves use to decide where to look, and anyone can check it against the call report.

The broad share adds banks carrying commercial real estate above 300% of capital that do not meet the growth test. The guidance does not mark them, because the book stopped growing. They are published because mature concentration is a different exposure, not an absent one: a bank with four or five times its capital in stabilised commercial real estate does not have the recent underwriting problem criterion 2 hunts, it has a balance sheet problem that surfaces when the market reprices.

The broad share carries its own one-quarter and four-quarter changes and stays out of the score. It appears on a county page only when it differs from the strict share by more than five points, which is when it says something the strict number does not.

Grade, ranking and the three coverage bands

Percentiles are taken within the region and within the quarter, among counties with coverage. The grade is the quintile of that score, labelled credit pressure. E is the highest pressure and A the lowest. A high grade is bad news for the county and useful news for the reader, so the scale is an intensity scale, not a warning scale.

Any grade also requires at least three in-universe banks in the county. A county that clears the coverage floor but has one or two banks gets the share without a grade, and says which of the two reasons applies.

The low band is not a soft version of the full band. It is the zone of lower confidence in the method, and it is declared rather than hidden, because moving the ruler to make Western Connecticut and Nassau fit would have been the dishonest fix. In Q2 2026: 160 counties at full coverage, 36 low, 21 not rated, 180 carrying a grade.

Ties are broken deterministically by score, then exposed deposit share, then county code. About one county-quarter in ten ties on the raw score, because so many counties sit at zero exposed share, and without a stable rule the grade moved between runs and poisoned the quarterly change.

Reproducibility and checksum

Every published figure carries a calculation version. Rerunning the same version on the same inputs has to reproduce the same ranking, byte for byte. A new version never overwrites an old one and never mixes with it: a change in method triggers a full recomputation of the series under the new version, because a delta measured across two versions is meaningless. This page and every data page state the version in force. These figures are calculation v1, built 2026-09-15.

Backtest: what we tested and did not find

One question, tested twice: did Northeast counties in the top quartile of the score go on to deteriorate more than the other rated counties over the following eight quarters? Cohorts were formed at Q4 2019, into the pandemic, and Q4 2022, into the rate shock.

They did not. Three of the four outcome measures show no separation between the top quartile and the rest, in either cohort. The one large and statistically significant result runs backwards, and for a mechanical reason worth stating plainly: counties enter the top quartile because they have flagged banks, most other counties have none, and zero has nowhere to fall. When a flagged bank's growth slows, the conjunctive criterion unwinds and the count drops. That is mean reversion in the flag, not improvement in the county.

Two things the test did show. In the 2019 cohort the noncurrent trend in top-quartile counties worsened relative to the rest, at p = 0.011 — weak, one window, but in the direction the score claims. And bank failures were useless as an outcome in the Northeast: the region produced too few to measure anything.

This is why Overhang is a map of exposure and not a forecast of deterioration. Eight quarters may be the wrong horizon for commercial real estate, where losses surface on refinancing dates rather than calendar quarters. The test is rerun every four quarters with a longer horizon, and the result is published here whichever way it points.

Manual reconciliation at the FFIEC

Before any of this was published, three institutions were recomputed by hand from their filings in the FFIEC Central Data Repository and compared field by field with what the pipeline produced, including one bank operating under the CBLR. All fields matched exactly. Twenty banks were pulled for the wider smoke test; the three hand-checked cases are the ones that close the loop between the API and the filed call report.

Limitations

• The backtest found no predictive power. This is a map of exposure.
• Deposits are a proxy for where lending happens. That is why the universe stops at $100 billion and excludes institutions without a loan book.
• Counties dominated by the four largest banks in the country — Westchester, Manhattan, Brooklyn, Philadelphia, Suffolk in Massachusetts — do not reach sufficient coverage at any ceiling tested and receive no grade. Commercial real estate exposure there is not attributable by deposit.
• The Summary of Deposits is annual, a photograph taken on 30 June.
• The call report arrives 30 to 45 days after quarter end.
• Concentration is exposure, not loss.
• None of this is investment advice.

Calculation changelog

Sources

• FDIC BankFind API — institutions, quarterly financials, Summary of Deposits, failures.
• Interagency Guidance on Concentrations in Commercial Real Estate Lending, Sound Risk Management Practices (2006).
• Call report Schedule RC-C Memorandum item 13, for the CBLR construction concentration test.
• US Census Bureau TIGER 2024 county and Connecticut planning region boundaries.
• FFIEC Central Data Repository, for the manual reconciliation.

Questions about the method go to data@getoverhang.com. Corrections are published here with the version that carries them.

Back to the report