Sixteen million grievances filed with a federal agency form one of the best behavioral datasets in consumer finance. They contain a textbook natural experiment — and a warning about what happens when filing becomes frictionless.
On September 7, 2017, Equifax disclosed that criminals had spent months inside its systems and walked out with personal records of roughly 147 million Americans. For most datasets, an event like this is a nightmare of confounds. For the CFPB's complaint database it is a gift: three near-identical companies — Equifax, Experian, TransUnion — doing the same business under the same regulations, and a shock that hits exactly one of them on a known date.
Here is what that looks like, week by week.
The clean lines make one quieter point and one honest limit. The quiet point: the control group matters — without Experian and TransUnion flat underneath in the breach week, the Equifax spike could be read as seasonal noise or a portal change. The limit: by 2018 all three bureaus have drifted onto a similarly higher plateau (Equifax ~577/week, Experian ~574, TransUnion ~544) — and all three were already climbing before the breach, roughly 35–50% a year through 2016–17. A control group that rises in step with the treated firm can't support a claim that the breach alone caused the plateau; the identified effect here is the spike, not the level it settled at.
Zoom out from twelve weeks to fourteen years and the story changes scale entirely. In 2019 the CFPB received about 277,000 complaints. In 2025 it received 5.4 million. Nothing about American banking got twenty times worse in six years — but the cost of filing collapsed. Credit-repair services and template generators industrialized the dispute letter, and the credit bureaus (the same three companies from Figure 1) became the target of nearly nine in ten filings.
Whether this is a crisis of credit reporting or a crisis of measurement is the researcher's dilemma, and the honest answer is “both.” Identity theft and post-pandemic fraud are real; so is automated filing. Either way, the lesson generalizes: any complaint channel that gets cheap enough stops measuring harm and starts measuring the channel itself.
“Template-assisted” is usually an assertion. Here it is measurable, because millions of complainants consent to publishing their narrative text. Normalize each published narrative and ask a blunt question: does this exact text appear in the database more than once?
The templates themselves are legible artifacts. The single most-filed text — 30,108 complaints, 2023–2026 — opens “In accordance with the Fair Credit Reporting act. The List of accounts below has violated my federally protected consumer rights…” and recites 15 U.S.C. 1681 section by section. Another, filed 12,335 times in 2025 alone, cites Engelhardt v. Gravens and the “doctrine of estoppel by silence” — a 1925 Missouri case that circulates through credit-repair kits. Nobody writes that sentence twice independently.
Normalize by population and the geography is stark. Georgia leads the nation at roughly 1,900 complaints per 100,000 residents in 2024 — about 3× the rate in southern New England (Connecticut, Rhode Island, Massachusetts) and 6–19× the rate across Wisconsin, Minnesota, and northern New England. Florida and DC are close behind Georgia; Delaware and Nevada round out the top five. The pattern tracks the geography of thin credit files, subprime lending, and the credit-repair industry itself, though the article does not test that mechanism directly.
Source: CFPB Consumer Complaint Database (US public domain), full public CSV extracted 2026-07-09: 16,681,759 complaints with unique IDs, December 2011 – July 2026. Product families harmonize the CFPB's changing taxonomy (e.g. all “Credit reporting…” variants merged). Per-capita rates use 2020 Census resident population. Complaints are published only after the company responds or after 15 days.
Template share (Figure 3): narratives of ≥100 characters are lowercased and whitespace-collapsed, then hashed; a narrative counts as duplicated if its exact normalized text appears ≥2 (or ≥10) times anywhere in the 2015–2026 corpus. Narratives exist only where the consumer opted into publication (2015 onward). 2026 is omitted from the chart: narrative publication lags filing, so the partial year is not comparable. Computed in DuckDB 1.5.4 (scripts/47_text_enrichment.py); the top recurring templates ship in 1data_aggregates/top_templates.json.
Complainants are self-selected; recent volume is heavily template- and service-assisted (Figure 3 puts numbers on this, and its exact-match method is a floor — merge-filled templates count as unique text). Duplicate text does not mean a complaint is meritless: the same lender can genuinely wrong thousands of people the same way; what it measures is industrialized filing. The “consumer disputed?” field was retired in April 2017 and is absent from the current export. 2011 and 2026 are partial years. Complaints about depository institutions under $10B in assets are not published. A proper choropleth of Figure 4 (and county-level drill-down) is planned with MapLibre + PMTiles.
Data: US public domain. Article text and figures: CC BY 4.0.