METHODOLOGY · hi-1.0.0
How the verdicts are computed
Every brand gets a categorical verdict — Proven & strong · Proven & steady · Showing strain · Distressed · Too new to judge · Not enough disclosure — computed from its own FDD Item 20 filing. Two rules make it honest. First, an evidence gate: a positive verdict must be earned with at least 25 franchised units and 3 disclosed years, because failures take years to surface (median SBA time-to-default is about 61 months) — a young system's clean record is arithmetic, not achievement, so it reads "Too new to judge," never "healthy." Second, hard flags outrank fundamentals: going-concern audit language, disclosed bankruptcies, and live distress signals cap or override the tier. Underneath, a numeric engine built from the factors below orders brands within tiers; it is internal and never displayed. We foreground exit quality, the signal most rankings underweight.
The factors under the verdict
Item 20 Table 1 + status tables
The core signal. Two or more consecutive years of net decline is a serious warning and caps this factor.
Item 20 terminations vs. transfers
Our signature metric. Terminations + distressed “ceased operations” as a share of total exits — catches rot that healthy-looking transfer volume hides.
Item 20 Table 3
High churn means franchisees want out, even when sales look “healthy.”
Projected vs. actual openings (Table 5)
Chronic over-promising is a tell. Undisclosed projections render “Not Disclosed” — never imputed.
Item 19 disclosure + completeness
Whether the brand discloses financial performance and files complete tables.
Citability rules
- ▸ Every score links to its source filing and the specific Item it draws from.
- ▸ When a brand doesn't disclose (e.g. no Item 19), we say “Not Disclosed” — we never estimate or impute.
- ▸ Extractions below 70% confidence are flagged and held back, not published as fact.
- ▸ The formula is stable and versioned. When it changes, the version increments and history is retained.
Sources & records
Every dataset feeding a number on this site, with its citation, refresh cadence, and known limitation:
cards.web.commerce.state.mn.us, Franchise Registrations / Clean FDD, 2016–2026 · Rolling crawl; each brand's latest filing plus up to 3 spaced filings for trendlines
Registration-state visibility only: franchisors selling solely in non-registration states never appear in any public portal.
data.sba.gov FOIA 7(a) loan file (franchise-coded loans) · Quarterly file refresh
Charge-off rates computed over resolved loans only (charged-off + paid-in-full); current loans excluded.
CourtListener / RECAP (PACER mirrors) · Continuous monitoring
Only cases mirrored into RECAP are visible; absence of a docket is not absence of litigation.
California EDD WARN Act filings · As filed
CA-only feed today; other states' WARN feeds not yet ingested.
Named outlet cited on every event · Continuous
Events carry their outlet as the source; news reports are treated as signals, never as filings.
Coverage — stated honestly
The corpus currently holds real FDD extractions for 776 franchise systems (220,712 franchised units) — roughly 25.9% of the ~3,000 systems visible across US registration states. Nearly all of it is sourced from Minnesota's registry today, which skews the sample toward brands that register there; we publish that bias rather than paper over it. Coverage grows with every crawl, and cross-corpus findings live on the research page with per-study sample sizes attached.
Distress Index
A separate fast-layer composite (di-1.0.0) of live signals — bankruptcy filings, mass-closure announcements, franchisee litigation, SBA default rates, and foot-traffic decay — each weighted by severity and recency. It is independent of the verdict: a brand can rate well on fundamentals yet throw a distress flag. That contrast is the point. The Distress Index is live and explicitly not a citable annual figure.
FDD Risk Score
A fractional logit scorecard (unweighted), FDD + SBA loan-record features (v2), trained on the 309 systems with at least 5 resolved SBA 7(a)/504 loans. The label is each brand's observed charge-off rate over resolved loans; the inputs are figures the FDD itself discloses plus three drawn from the federal loan record behind its franchisees. The trained scorecard then scores all 569 systems with a real filing — including brands whose franchisees never borrow through SBA, which is the point: the FDD is public years before loan outcomes are.
Accuracy is reported cross-validated, never in-sample. Under 5-fold cross-validation, measured against the 136 brands whose own charge-off rate is itself reliable (≥30 resolved loans), Spearman rank correlation is 0.59 and mean absolute error is 4.8 percentage points. Across all 309 labels including very thin cohorts those figures are 0.37 and 8.8pp — lower mostly because a six-loan brand's own rate is itself uncertain by ±15pp, so it is a noisy yardstick rather than evidence the model is worse. Quintile calibration (modeled → observed): 5.7% → 6.9% · 9.6% → 11.7% · 12% → 14.2% · 15% → 18.4% · 23.3% → 27.4%.
Each brand also carries a confidence tier reflecting the evidence behind its score: 155 corroborated · 163 insufficient · 52 measured · 199 directional. Below the corroborated tier we publish a range rather than a number, because a thin loan cohort cannot support a two-digit claim.
| Feature | Coefficient (standardized) |
|---|---|
| Item 20 exit rate | +0.102 |
| Net unit growth | -0.133 |
| System size (log units) | -0.212 |
| No Item 19 disclosure | +0.003 |
| Royalty rate | -0.027 |
| Item 3 litigation (log) | -0.062 |
| Non-clean audit opinion | +0.089 |
| Investment ceiling (log) | -0.139 |
| Share financed by high-loss lenders | +0.348 |
| Single-lender dependence | -0.134 |
| Share of buyers who are first-time operators | +0.015 |
The score shown on brand pages is the percentile of the modeled charge-off rate across all scored systems, always displayed with its top drivers and, when a resolved cohort exists, the observed rate beside it. It is deliberately a linear scorecard, not a black box — every score decomposes into the disclosed figures that produced it. A score is context for reading the filing, not a substitute for it.
Editorial
Rankings are algorithmic and sourced. They are not for sale and are firewalled from any future referral or sponsorship layer. Corrections: every page links to its underlying filing; if a number is wrong, the filing is the arbiter.