Skip to content
FinObservatory

County banking markets / methodology

County banking markets: methodology

What the 3,237 pages in this module are built from, how the two concentration indices are constructed, which one leads and why, and what the survey cannot answer.

Source and unit

The FDIC Summary of Deposits (SOD) is an annual census of the branch offices of every FDIC-insured institution, taken as of June 30. FinObservatory holds it as two parquets, both from the verified BankFind bulk pull recorded in docs/data_provenance.md: sod_branches (the branch rows) and sod_market_structure (state and county rollups derived from them). SOD dollar fields are in thousands of US dollars; pages display them as $M, $B and $T on that basis. Coverage runs 1994–2025.

Why the county

Because it is close to the geography that bank merger review actually uses, and the rest of this site does not have it. The Federal Reserve, on the geographic markets it screens:

“Many geographic markets follow Metropolitan Statistical Area (MSA) definitions or rural county lines, but some markets comprise multiple MSAs/counties or parts of MSAs/counties, reflecting that economic activity does not always track political boundaries.”

A county is a proxy for a Fed banking market, not one of them: the second clause of that sentence rules out treating the two as equal. The Fed’s actual definitions live in CASSIDI and are drawn from commuting and shopping patterns. See Competitive Effects of Mergers and Acquisitions FAQs, questions 10 and 13.

The county panel is rebuilt from the branch rows, not read off the rollup

Every county number on these pages is derived from sod_branches, keyed on lpad(stcntybr, 5, ‘0’) = county_fips. That re-derivation reproduces sod_market_structure exactly for all 3,194 county rows of the 2025 survey: branch count, institution count, total deposits and HHI all match, apart from the 1 county whose HHI is null (below). The branch counts also close three ways:

county rows, 2025   sum(branches) = 76,120
state rows,  2025   sum(branches) = 76,120
raw branch file, 2025      rows = 76,120

All three are equal, so no branch is unassigned to a county and no branch is counted twice.

The re-derivation is not ceremony. The rollup’s county rows are keyed on (year, state, county_fips), not (year, county_fips), so when a single branch carries a state code that disagrees with its own FIPS county the rollup splits that county-year into two rows and neither row is the county’s total. Six county-years are split this way: Camden, New Jersey (34007) in 1999 and 2000, where one branch is filed under DE, and York, Pennsylvania (42133) from 2007 to 2010, where one branch is filed under NJ. Camden in 1999 is stored as a 1-branch row with an HHI of 10,000 and a 143-branch row with an HHI of 1,619; the branch file says the county has 144 branches across 19 institutions. Reading those rows as counties would invent a phantom single-bank county in each of the six years and put a phantom point on two county pages. Keying on (year, county_fips) removes them. On the 102,544 county-years the split does not touch, the rebuilt panel and the rollup agree exactly on branches, institution count and total deposits.

Two HHIs, and why only one is a headline

The Herfindahl-Hirschman index is the sum of squared percent market shares, on a 0–10,000 scale. This survey supports two of them per county, and they measure different things.

Branch-count HHI squares each institution’s share of the county’s branch offices. It is what every headline, ranking and chart on these pages leads with.

Deposit HHI squares each institution’s share of the deposits booked in the county. It is the column already carried in sod_market_structure, and it is the one the national and state pages use. At county level it is not reliable, for a reason the Fed states plainly:

“Central booking occurs when a bank records deposits at a central office and does not book the deposits to a branch that is connected to the location of the depositor. As a result, centrally booked deposits of an applicant can overstate the presence of an institution in the local banking market.”

SOD books a deposit at the branch of record. It therefore contains the distortion raw. The sharpest case in the 2025 survey is Lincoln, South Dakota: Citibank, National Association books 98.2% of the county’s deposits from 2 of its 37 branch offices, that is 5.4% of them. The deposit HHI reads 9,650, highly concentrated on the 2010 bands. The branch-count HHI reads 489, unconcentrated on the same bands. The deposit index is not describing local competition there; it is describing where one bank keeps its ledger.

Deposit HHI is therefore an upper bound in counties where institutions book centrally, and a lower bound in every county the money actually came from. It is still shown, on every page, beside the branch measure, and labelled.

Neither index is the Fed’s screen HHI

The Fed’s delegation screen fires when a deal “would raise the HHI by 200 points or more to a level of 1800 or higher in any local banking market”, or would take the acquirer above a 35% market share. The 1,800 line appears on the charts as a dashed reference. But the index it applies to is not the one computed here. In the Fed’s calculation, commercial-bank deposits get 100% weight and thrift deposits 50%; credit-card banks, trust companies and internet banks are generally dropped from local market shares altogether; and central booking and large government deposits are adjusted for case by case. The indices on this site are unweighted and include every reporting institution. A number here is a descriptive statistic, never a merger-review outcome.

The shaded bands are the 2010 Horizontal Merger Guidelines classification (section 5.3): unconcentrated below 1,500, moderately concentrated 1,500–2,500, highly concentrated above 2,500. The 2023 Merger Guidelines moved the highly-concentrated line to 1,800, which coincides with the Fed’s banking screen level.

Null HHI is null, not zero

In 20 county-years the county’s total SOD deposits are zero, because its branches report zero deposits, and the deposit HHI is undefined. The longest-running case is Arthur, Nebraska: null in 5 surveys, 2021–2025. In 2025 its 1 branch office books no deposits at all, which leaves a branch-count HHI of 10,000 and no deposit HHI. These are carried as null and displayed as n/a, never coalesced to a number. A zero here would be a fabricated reading of an unconcentrated market.

The median is across counties, not people

Every median on the index page is unweighted across the counties in that survey, and the survey does not carry the same number of them every year: the count runs from 3,194 to 3,210, and is 3,194 in 2025. A median across counties weights a county of 10 million and a county of 500 equally. This estate holds no county population data, so nothing here says how many people live in the counties whose institution counts fell, and the flat median is not a claim about the typical depositor.

Branch-count HHI in a small market also takes discrete values (six branches split 2-1-1-1-1 gives exactly 2,222.2), so the median across counties sits on a lattice point and steps rather than glides. The flat median between 1994 (2,222) and 2025 (2,222) is therefore checked against the mean, which is not confined to that lattice: it reads 2,993 in 1994, 2,690 in 2009 and 2,988 in 2025. Both fall and return. The other two series on the index page do not move together and are not offered as corroboration: counties above 1,800 end below where they started (2,045 to 1,998), while single-institution counties end above (220 to 236).

FIPS codes that leave the file

Of the 3,237 county FIPS codes in the panel, 43 are absent from the 2025 survey. Treating all of them as counties that lost their last bank would fabricate a finding. 16 are not counties in the 2025 geography at all: geography changes, cross-checked against the Census record of substantial changes to counties and, for Connecticut, the Federal Register notice of 6 June 2022 approving the nine Council of Governments planning regions as county equivalents, and, for the freely associated states, the Compacts themselves:

  • Connecticut’s 8 counties (09001–09015) last appear in the 2022 survey; the 9 planning regions (09110–09190) start in 2023. No county code and no planning-region code appear in the same survey, so the two sets do not overlap and nothing is double-counted across the change.
  • 4 Alaska codes were retired as the state's census areas were reorganised: Valdez-Cordova (02261), Wrangell-Petersburg (02280), Prince of Wales-Outer Ketchikan (02201) and Skagway-Hoonah-Angoon (02232). This register records WHEN each code stops appearing in the survey, which is what the deposit file can show. It deliberately does not map each retired code onto its successors: those reorganisations were not all clean splits (parts of Prince of Wales-Outer Ketchikan went to Ketchikan Gateway, 02130, which reports in all 32 surveys), and the successor geography is a Census fact that this deposit file cannot establish.
  • Bedford, Virginia (51515) reverted from an independent city to a town inside Bedford County (51019) in 2013.
  • Florida 12025 is the retired code for Dade County, renamed Miami-Dade and recoded 12086 in 1997. It nonetheless carries branch rows long after the recode: 2 TD Bank, National Association branches, 2019; 1 Wells Fargo Bank, National Association branch, 2021-2024. Their branch addresses (addresbr, citybr) are in Palmetto Bay, Miami and Coral Gables, Florida, all inside Miami-Dade, so these are branches filed under a retired code rather than a county that came back.
  • Northern Mariana Islands 69010 is not one of the four Census county equivalents (69085 Northern Islands, 69100 Rota, 69110 Saipan, 69120 Tinian). It last reports in 2024, with 1 branch office. That branch is First Hawaiian Bank (cert 17985), and cert 17985 files under Saipan (69110) in the 2025 survey, so Saipan is where the code’s last branch went. Tinian (69120) first appears in the county file in the 2025 survey, which makes it the tempting successor and the wrong one: its 2025 branch is Bank of Guam (cert 20884), a different institution. This register matches successors by FDIC cert for that reason.
  • Kwajalein (68150) sits in a freely associated state: it is not a recode and not a Census county equivalent. See the section below.

The switch is not clean. 2 of the 4 retired Alaska codes keep receiving branches after their successors appear: 02261 for 2 surveys and 02280 for 5. A raw count of distinct county_fips therefore double-counts those geographies in the overlap years. No code listed above appears in the 2025 survey, which is why every cross-section on these pages is taken there.

The remaining 27 are codes this register does not flag, with no SOD branch reported after the year shown on the index. That sentence is the whole claim. Calling them banking deserts would go beyond the survey, which counts branches of FDIC-insured institutions and nothing else. Departure is not always permanent either: 19 FIPS codes in the panel drop out for one or more surveys and come back.

The county file is not confined to the United States

The FDIC surveys the branches of every insured institution, including those in the three Compact of Free Association states (Federated States of Micronesia, Marshall Islands, Palau). These are sovereign countries in free association with the United States, not US territory, and Census does not treat their FIPS codes as county equivalents. 6 such codes are in the 2025 survey, carrying 12 branches between them. 1 more code has left it: Kwajalein (68150). They are kept, because dropping them would silently edit the survey, and they are labelled on their own pages. Two consequences bind every sentence in this module: no page may claim to cover “every US county”, and no page may call every code in the file a county equivalent. Where these pages say “county”, read it as a row of the SOD county file.

What this module cannot tell you

  • Whether a county is a banking desert. Credit unions file with the NCUA, not the FDIC, and are absent from this survey. Branchless and internet banks hold deposits that appear in no county. A resident of a county with no branch may bank one county over. SOD measures branch presence, not banking access.
  • The full set of counties without a branch. There is no Census county master list in this data estate. Every code here is one that reported at least one branch at some point since 1994. A county that never reported one is invisible to this module, so the absent-code tables are a lower bound.
  • Anything about lending. SOD is a deposit survey. It says nothing about who lends in a county, at what price, or to whom.
  • Whether concentration is harmful. An HHI is a structural statistic. Nothing on these pages estimates an effect on rates, fees or credit availability.
  • A merger-review answer. See the screen-HHI section above. Wrong geography, wrong weights, wrong institution set.

Back to county banking markets, or to the national deposits methodology for the survey definition and the DEPDOM reconciliation anchor.