For AI agents: a documentation index is available at https://docs.coverbase.com/llms.txt. This page is also available in markdown by appending .md to the URL.
The Coverbase Financial Health Score answers one question about a vendor without asking the vendor anything: is this company financially sound enough to depend on? It runs 0-100, where higher is healthier, and it is computed from regulatory filings where a company files them and from funding, headcount, and corporate-standing signals where it does not.
What makes it usable in practice is how it’s acquired: no supplier outreach, no questionnaire, no waiting on a vendor to opt in. Coverage is portfolio-wide from day one, and every vendor is rescored on each Vendor Intelligence run.
This is an optional feature currently in beta. Let your Coverbase representative know if you’d like it turned on for your organization.
What it does
A score with its evidence class attached
Every score is published with the confidence tier behind it, so an inferred number is never presented with the authority of an audited one. The tier is derived from what actually resolved, never declared.
Ratios ranked against real peers
A ratio only means something against a comparable set. Coverbase ranks each vendor inside its own industry cohort, drawn from thousands of filers that reported the same concept in the same period.
Every input is inspectable
Open the rationale on any pillar and you get the inputs behind it: each one’s measured value, what it scored, how heavily it counted, and whether it was ranked against peers.
Distress caps the score outright
A bankruptcy filing or a dissolved registration doesn’t get averaged into a weighted mean. It pins the score, and the card says which rule fired.
Confidence tiers
This is the most important thing on the card. A Tier C 72 and a Tier A 72 are different claims about different evidence. The tier travels with the score everywhere the score goes: into the API, into MCP answers, into exports.
If a vendor’s filings stop resolving, it drops to a lower tier and the card says so. It does not keep presenting a statement-grade number built on estimates.
Scores are held to a common scale across tiers by the stability class below, so ranking a mixed portfolio on the raw number is reasonable. The tier still tells you how much evidence stands behind any individual result, and it belongs in any conversation about a surprising one.
Stability classes
A single number cannot order companies that survive by different mechanisms. An entrenched, cash-generative incumbent and a venture-funded startup are not two points on one axis, and treating them as one produces a ranking that is hard to defend in front of a customer. Every vendor is therefore assigned a stability class, derived from observable scale, operating history, cash generation and dependence on outside capital. Never from a name, a ticker, or an opinion.
Each class occupies a score range, and the ranges are ordered so an early-stage company’s ceiling never reaches an established company’s floor. A vendor that just raised cannot out-rank a profitable incumbent on inference alone.
Two limits on that. Evidence outranks inference: the ranges constrain what may be inferred, and a bankruptcy filing or going-concern opinion drives a score straight through any floor. And every adjustment is published. Where a floor or ceiling moved a score, the card shows the rule, the reason, and the before-and-after figures, so a corrected number does not read as a computed one.
Runway and burn for private companies
For the majority of a portfolio that files nothing, the score models how long the company’s cash lasts. Burn is derived from headcount rather than guessed. Fully-loaded cost per employee is the operating figure that behaves most predictably across software companies, adjusted for stage and grossed up from personnel to total operating cost. Revenue is netted off at gross margin, using a conservative stage estimate where no figure is available. Capital is the last round plus a reserve, less what has been consumed since. The output is a range, along with two readings that often matter more than the midpoint: whether the company is self-sustaining (revenue covers burn, so no cliff applies), and its capital efficiency (revenue per dollar raised). Every assumption used is recorded on the scorecard. Most private rounds are reported with a size but no closing date. Crediting the vendor with a round closed yesterday flatters it, and charging a fixed number of months of burn penalises the companies with a large payroll. So an undated round is treated as half spent: observing a company at an unknown point in its funding cycle is observing it, on average, halfway through. These are bounded estimates, labelled as such. They separate a vendor with two years of cover from one with two quarters, which is the distinction that affects a buyer. They are not a guess at a bank balance.Funding and headcount without a data subscription
The model above needs a headcount and, ideally, a last round. Crunchbase and PitchBook supply both, for the customers who license them and for the vendors those providers happen to cover. Coverbase’s own enrichment also researches each vendor’s employee count, last round size, round type and founding year from public sources, and the score reads those directly. For a portfolio with no financial-data subscription behind it, that is the difference between a cash model that runs and one that never had a headcount to price. Two rules apply. Researched figures are cited as Web research rather than attributed to a provider Coverbase never called, and they never raise a vendor’s confidence tier, because a search result is not a filing however precise it looks. Where a licensed provider does answer, its figure wins and the researched one fills gaps only.What the score is built from
Five pillars. Which ones carry the score depends on the tier, because a tier with no filed statements has nothing to put in the financial core:
The weights renormalize over the pillars that actually exist for a vendor, which is why a Tier C score is not simply a Tier A score with holes in it.
How a score is computed
1
Each raw value is trimmed at its cohort's extremes
A single outlier filer shouldn’t set the top or bottom of the scale for everyone else.
2
Values are percentile-ranked inside the peer cohort
This is the step that does the real work, and it matters more than any weight. It’s what makes 4x debt/EBITDA at a utility and at a software company comparable. Ratios where lower is better (leverage, days sales outstanding, burn) are inverted so that higher always means healthier.
3
Level is blended with direction
Where the trend matters as much as the level, the input score is weighted toward the current level but carries a meaningful share of the direction it’s moving.
4
Events accumulate by severity and decay by age
The events pillar doesn’t percentile-rank. It sums decayed severity, so a two-year-old event does not weigh the same as last month’s. Positive events carry negative severity and pull the burden back toward zero.
5
Pillars aggregate, then the composite renormalizes
Each pillar is a weighted mean over the features present; the composite is a tier-weighted mean over the pillars present.
6
The result is smoothed month over month
Damps routine churn so the score doesn’t jitter on noise.
7
Caps and floors apply last, unblended
An override is not averaged in. See caps and floors below.
8
The score maps to a risk band
0-19 Very High Risk · 20-39 High · 40-59 Medium · 60-79 Low · 80-100 Very Low.
Two rules that shape everything
A vendor is never penalized for something we could not compute
A vendor is never penalized for something we could not compute
Missing features drop out and the remaining weights renormalize. The cost of sparse data is paid by the confidence tier, not by the score. Otherwise a poorly-covered vendor would look distressed purely for being poorly covered, which is a very expensive kind of wrong.The same applies at the ratio level: a ratio with a zero, missing, or wrong-signed denominator is treated as absent, not as zero.
But it is never rewarded for it either
But it is never rewarded for it either
The rule above says sparse data must not read as distress. It does not license sparse data reading as strength.An observation we never made is capped rather than scored perfectly. For a company whose filings we read, an empty adverse-event stream really does mean nothing was disclosed; for a private company it means nothing was observed, and it is scored well but not perfectly. Likewise a score resting on neither statements nor a defensible cash estimate is held below the top band.Read together, the two rules are what keep the least visible vendors in a portfolio out of the top of the ranking.
Scores decay slowly upward and move fast downward
Scores decay slowly upward and move fast downward
Smoothing damps routine movement in both directions, but severe events and every override bypass it entirely. A company recovering from distress climbs gradually as the evidence accumulates; a company entering distress drops immediately.
Peer comparison
A ratio in isolation is not a judgement. 4x debt/EBITDA is unremarkable at a utility and alarming at a software company, so every Tier A ratio is scored as a percentile inside a cohort rather than against an absolute threshold. The cohort is real, and it is large. It’s built from the population of filers that reported the same concept in the same period: typically thousands of companies per ratio, not a hand-picked comparison set. Industry narrows it. Ranking a software vendor against every filer (banks, REITs, miners) produces a percentile that is arithmetically correct and analytically meaningless, so cohorts are scoped to the vendor’s own industry classification. When the cohort is too small, the card says so. The score falls back to a wider population and discloses that it did. A disclosed wide cohort is honest; an undisclosed one is misleading.Tier B and C signal features (funding, headcount) are scored against documented absolute curves rather than peer percentiles, because no equivalent free population exists for them. Any input scored that way is flagged as not peer-ranked, both on the card and in the API.
Caps and floors
Some facts should not be averaged. A bankruptcy filing is not a modest deduction to be offset by a healthy current ratio, so these rules apply after the weighted score and override it outright:
Every override that fires is recorded with its reason and shown on the card, so a pinned score explains itself rather than looking like an unusually harsh weighted result.
Adverse events
Events that bear on financial health (bankruptcy filings, going-concern doubt, mass layoffs, auditor changes, delistings, CFO departures, down rounds, material impairments, regulatory actions, late filings) are weighted by severity and decay over time. Events awaiting human confirmation are labelled Needs review, and they count against the score until someone confirms or dismisses them. Dismissed events stay on the record for audit but stop affecting the number.What the card shows
- A one-to-two line summary covering what the score was built on, the stability class, and how much runway the company has
- Score, risk band, and confidence tier, with 3-month and 12-month movement
- Pillar breakdown, each with the weight it carried and the inputs behind it
- Financial vitals: revenue and growth, operating margin, current ratio, debt/EBITDA, interest coverage, free cash flow margin, estimated runway, total raised, headcount, last funding round
- Registered entity standing: status, jurisdiction, and filing compliance
- Peer context: the cohort the vendor was ranked inside and its percentile within it, with any fallback disclosed
- Score history, plotted as discrete restatements rather than a smooth line, because each point is a separate computation
- Adverse events with date, review status, and a link to the source that reported each one
- Caps and floors applied, with the reason
- Sources, each with the date its data was true, plus the sources that were asked for and unavailable
Why this score?
Open the rationale on any pillar and the card shows the inputs that produced it, each with the measured value, what it scored, and how heavily it counted. Inputs are ordered by how much they cost the pillar, so the reason for a low score is at the top. A capital position of 50 might resolve to a current ratio of 1.8x scoring 62 at half the weight, and a registration status Coverbase could not determine. That second one matters more than it looks: an undeterminable input is labelled as such rather than shown as a bare number, because “50” reads as measured mediocrity when it actually means nobody knows. Two things travel with every input:- Whether it was peer-ranked. An input scored against the vendor’s own cohort is a much stronger claim than one scored against an absolute curve.
- The date the value was true. Filed figures lag. An input from a filing two quarters old carries its own as-of date, not the date of the refresh.
Submitting a vendor’s own financial statements
Most vendors are private companies, so most score Tier C however sound they are: there is nothing filed to read. If you already hold that vendor’s financials — and in a diligence relationship you usually do — you can give them to Coverbase and the score is rebuilt on them.1
Open the vendor's Financial health tab
Vendor Intelligence > Financial health. The overflow menu at the top right carries Upload financial statements and Upload supporting information; both open the same dialog, and there is an Upload financial statements button beside the export for the same thing.
2
Drop the documents in
Audited financials, tax returns, SEC filings, management accounts. Supporting evidence belongs here too — a Dun & Bradstreet report, a Better Business Bureau profile, a consumer report. Coverbase reads those for the entity’s identity and for concerns, and will not invent fiscal periods from them.
3
Generate
Generate analysis runs it. Uploading and generating are separate on purpose — you can add files across several sessions and run the analysis when the set is complete. One pass reads every document together, because a statement set only makes sense as a set. It takes a few minutes; the tab shows progress and says plainly if a document could not be read.
Exporting the analytics workbook
Once an analysis exists, Export analytics workbook in the tab’s overflow menu downloads the full credit spreadsheet: the memo and review summaries, every metric beside the two figures it was computed from, the raw line items with the page each was read from, and the reference sheets that define what Low, Medium and High mean. It is offered only where an analysis exists — the workbook is a rendering of one, and there is no such thing as an empty one. Nothing in it is inferred: the credit agency and reputational rows say Not established rather than deriving an agency rating from a statement set.Changing what the analysis reads
Both the upload dialog and the Documents used in this analysis list on the tab let you take a document out of consideration. The file stays on the vendor’s Documents tab — excluding it from a financial analysis is not a reason to delete it. Removing or adding a document does not blank the figures already on screen. They are still true of the documents they were computed from, so Coverbase keeps them and marks the analysis stale: the tab says the document set has changed and offers Regenerate. Nothing is recomputed, and nothing is charged, until you ask for it.What you get back
The classification and audit opinion, the periods covered, the issuing legal entity and its auditor, and the full metric battery per fiscal year with the change between years:- Balance sheet — working capital, current ratio, quick ratio, cash ratio, total funded debt, debt to equity, debt to capitalization, net worth, asset turnover, return on assets, equity multiplier, return on equity
- Income statement — revenue, EBIT, EBITDA, degree of financial leverage, debt to EBITDA, gross and operating margins, net income, net income margin, earnings per share, inventory turnover, AR turnover
- Cash flow — free cash flow, change in cash
How the figures are produced
The model reads the statements and reports line items — figures that appear on a balance sheet, an income statement or a cash flow statement, each with the page it came from. Every ratio, every year-over-year change and every risk rating is then computed by Coverbase, not by the model. A credit memo’s arithmetic has to be reproducible from its inputs, and a model asked to divide two numbers is occasionally wrong and always sounds certain. A figure the statements do not report comes back as not reported, never as zero. A zero is a measured result and reads as one.Where it appears
Vendor Intelligence
The Coverbase Financial Health Score card, refreshed on the regular Vendor Intelligence cadence. Its Financial health page opens the same data full width, with score history, every pillar expanded to its inputs, the full vitals set, and the event timeline.
Assessments
Financial facts are supplied to control evaluations as authoritative evidence, ranked ahead of web results.
Intake and inherent risk
A new vendor request is scored against its registered identity and financial standing before the web is consulted.
API and MCP
Available on the vendor fact-sheet endpoint as
financials.financial_health, and conversationally through MCP: “what’s the financial health score for Acme?”Coverage and limits
Coverage is honest about its edges, and these are the ones worth knowing before you build a process on the score.- SEC EDGAR covers roughly 7,000 US filers, often 15-25% of a typical vendor portfolio, on a quarterly cadence with a 40-90 day filing lag. Many private vendors will sit at Tier C, and that is the system working as designed rather than failing.
- Tier B and C scores are inferred. They rank vendors sensibly against each other. They do not carry statement-level precision, and the confidence tier is how you tell.
- Runway is a bound, not a measurement. The priors behind it are reasoned from what is typical of software companies, not fitted to observed outcomes. Use it to rank and to prompt a question, never as a figure to quote back to the vendor.
- A stability floor can mask weakness. An incumbent whose fundamentals are deteriorating, but which has triggered no hard rule, will not read worse than its class floor. That is a deliberate trade against the opposite failure, and it is why the class and any adjustment are published beside the score rather than applied silently.
- The score is a ranking, not yet a probability. The current weights are heuristic, chosen to order vendors sensibly rather than fitted against observed failures. Treat the score as “who needs attention first”, not as an estimated probability of default. Calibration against a failure label set is the next step, and until it lands, no probability language belongs on a slide.
- Headcount and funding data lag and skew toward venture-backed technology companies. A bootstrapped logistics vendor may barely register in them.
- A scorecard built on one or two signals is capped mid-scale. Most of such a score comes from an adverse-events pillar that found nothing because nothing was watched, so it reports what it found without reading as reassurance. The cap is published beside the score.
- Trend features need history. Where a source publishes only current values, trend inputs are omitted until Coverbase has two snapshots of its own. They are never assumed flat.
- The score describes the entity that was matched. If a vendor is a subsidiary and the filings belong to the parent, the score describes the parent’s balance sheet. The Corporate Registration card shows which legal entity was bound. See Vendor Intelligence.
Two expansions are in progress: UK Companies House, which extends Tier A past US SEC registrants and is the single highest-value coverage improvement available; and a news and event feed, which will widen the adverse-events pillar beyond what regulatory filings disclose. Until the latter lands, events are limited to what filings reveal.
Availability
The Financial Health Score is an optional feature in beta, enabled per organization. Contact your Coverbase representative to have it turned on. Nothing is computed or stored for an organization that doesn’t have it enabled.It pairs with Vendor Intelligence, which validates the corporate identity the score is computed against, and Security Intelligence, which rates a vendor’s external security posture. For reading the card day to day, see the financial health and security intelligence guide.