Report · Early warning
What shows up before a nonprofit fails
The year a nonprofit files its last return, the return itself usually reads as ordinary. We worked backwards from 29,581 organizations whose last filing was followed by an IRS revocation, and 32,509 filer-years where revenue then fell by half, and asked what the public record already showed. A watch list built on the standard financial ratios catches 18% of the coming collapses. Built on the ratios plus how the organization and its funders were behaving, it catches 31%.
Published 2026-10-06
Cite this: Plinth. (2026). What shows up before a nonprofit fails: early-warning signals in US nonprofits. data.useplinth.com/reports/early-warning.
How we counted
The usual way to judge a nonprofit's health is a set of ratios off its balance sheet: months of liquid assets, months of reserves, operating margin, whether net assets are above zero. Anyone who has watched an organization fail from close up knows the ratios can read as fine while things underneath are not. Since the IRS began publishing nearly every return as data, that intuition can be tested: take the organizations whose ending is on the record, wind the clock back one filing, and see what was visible.
An ending has to be unambiguous to count. 1,088,844 observation points are 990 and 990-EZ filers at their third-or-later return, 2019 to 2021. A death is a final filing followed by an automatic IRS revocation that was never reinstated; organizations that simply stop filing with no revocation (mergers, group returns, church exemptions) are excluded from both sides. A collapse is the next return showing revenue under half of the lower of the last two years, for organizations over $100k. The lower of two matters: measured against a single year, the "collapses" are dominated by one-off windfalls returning to normal, and the strongest predictor becomes a large surplus, which is a spike artifact and not distress.
The signals
Some of the strongest signals are not financial. They are about people and funders. An organization that cut its officer pay in half was 4.5x as likely to be filing its last return (1.4% against 0.3%); one whose paid officers went to zero, 3.8x; one whose treasurer or CFO left the roster, 2.1x. Filing behavior carries signal of its own: dropping from the full Form 990 to the shorter EZ ran at 2.7x, and a skipped filing year at 2.5x.
Before a death
Share of organizations filing their last return, with and without each signal, all sizes. Each row's comparison covers the organizations the signal can be judged for.
The funder side is where this data can see what no single return shows. An organization that lost every one of its institutional funders was 3.1x as likely to see revenue halve (12.7% against 4.0%, interval 10.3% to 15.4%). Losing a single funder worth a tenth or more of revenue ran at 1.9x for collapse and 2.6x for death, and it is the version of the signal a reader can act on, because it names a relationship rather than a portfolio. What happens after that particular loss has a report of its own: what happens when the money stops.
Before a collapse
Share of organizations whose revenue then fell by half, with and without each signal, organizations over $100k.
How much the ratios miss
The fair test has three steps. First, the classic screen: the ratios, in a logistic model, which is how the research literature has run them since the 1990s. Second, the same ratios through a modern gradient-boosted model, so the estimator improves but the information does not. Third, everything: ratios, trajectory, funding mix, funder behavior, governance, government money. Each model ranks all organizations, and the score is how many of the coming failures land in the top tenth of the ranking, a watch list of fixed size.
Share of coming failures caught in a top-decile watch list
Five-fold cross-validation, 2019 to 2021. Hollow dot: the ratio screen. Gray dot: the same ratios, gradient boosting. Filled dot: everything.
For a death, the ratios carry real signal and the gap is meaningful: the full model lifts the catch from 29% to 46% for organizations over $100k, with roughly half of the lift coming from the better estimator and half from the added information. For a collapse the ratios genuinely fail: an area under the curve of 0.65, barely a screen, against 0.80 with everything in. Put in review-load terms: to catch what the full model catches in its top 10%, the ratio screen has to flag 18% of all organizations for a collapse and 19% for a death.
Government money is stable, until it stops
Organizations with a government-grants line on their return collapse at less than half the base rate (2.2% against 4.9% across 155,862 organization-years). The money is steady while it flows. When it stops, the picture inverts: an organization whose federal awards went to zero was 2.5x as likely to collapse, and one whose government-grants line fell to under a quarter, 2.2x. In a model of the whole population these events barely move the score, because few organizations experience them in any year. For the organizations that do, they are among the sharpest signals on this page.
It shows a filing early
A fair objection: the final filing arrives 12 to 24 months late, so a signal read from it may arrive with the obituary. So we reran the death model for organizations over $100k using only what was knowable at the filing before, one more year removed from the end. It keeps most of the discrimination: an area under the curve of 0.75 against 0.81, and a top-decile catch of 33% against 46%. The record does not merely describe the end. It sees it coming.
What does not predict
Three negative results, reported because a list of signals is only credible next to the signals that failed. A changed mission statement predicts nothing (0.9% against 0.9%). An organization whose key funder shrank its own grantmaking by 40% or more shows no elevated risk (0.5% against 0.5%): funder retrenchment does not detectably propagate to its anchor grantees. And for a revenue collapse, the two most classic distress markers are flat: negative net assets (4.5% against 4.9%) and two deficits in a row (4.9% against 4.9%). The balance sheet the year before a collapse is, on average, unremarkable.
What to do with this
For funders: the strongest per-organization signals are events you can watch for by name. A grantee whose treasurer left, whose officer pay halved, or whose other institutional funders walked away is telling you something its ratios are not. And if your grant is the one worth a tenth of the budget, the exit you design is part of the risk: the grant-cliffs report is about that decision.
For organizations: every signal on this page can be read off your own filings before anyone else reads it. The list is triage, not a determination: a top-decile score means heightened odds across a population, and 46% of coming deaths in the top tenth also means the rest were not there.
Does it hold up? Models, holdouts and what carries them
Area under the ROC curve and average precision, five-fold cross-validation on a sample of all cases plus 150,000 controls; the top-decile catch is on the same sample. Base rates: death 0.9% of 1,002,654 observation points, death over $100k 0.5%, collapse 4.9%. The unit is organizations throughout.
| Outcome · model | AUC | Avg. precision | Top-decile catch |
|---|---|---|---|
| Death, all organizations · ratio screen (logistic) | 0.762 | 0.174 | 36.0% |
| · same ratios, gradient boosting | 0.812 | 0.266 | 43.3% |
| · everything, gradient boosting | 0.829 | 0.296 | 45.7% |
| Death, over $100k · ratio screen (logistic) | 0.719 | 0.047 | 29.1% |
| · same ratios, gradient boosting | 0.778 | 0.109 | 40.1% |
| · everything, gradient boosting | 0.811 | 0.132 | 46.0% |
| Collapse · ratio screen (logistic) | 0.647 | 0.275 | 18.3% |
| · same ratios, gradient boosting | 0.721 | 0.360 | 24.0% |
| · everything, gradient boosting | 0.802 | 0.477 | 30.9% |
Temporal holdouts, full model, with the calendar year excluded from every fit (a tree cannot extrapolate to a year it never saw). 2019 is the only pre-pandemic year the design allows, so the splits are: forward across the boundary (fit 2019, test 2020 and 2021), inside the later regime (fit 2020, test 2021), and reverse. Out-of-year performance runs a few points below cross-validation on both outcomes, and the drop does not concentrate at the pandemic boundary; read the headline figures as in-window, and any deployed version of this as needing yearly refits.
| Split | Death over $100k | Collapse |
|---|---|---|
| fit 2019, test 2020 and 2021 | 0.78 | 0.75 |
| fit 2020, test 2021 | 0.78 | 0.73 |
| fit 2020 and 2021, test 2019 | 0.83 | 0.74 |
What carries the death model, by permutation importance: revenue size, board size, net-asset drawdown, months of liquid assets, months of reserves, investment-income share. For collapse: the calendar year, months of reserves, officer compensation, investment-income share, revenue size, operating margin. The calendar year leading the collapse list is the pandemic cohort effect, and is one reason the holdout caveat above exists.
More from the funding graph
IRS Form 990 and 990-EZ e-file returns; observation years 2019 to 2021, outcomes read from filings through FY2023; institutional funder relationships from grant lines resolved to recipient EINs; federal awards from USAspending.gov resolved to EINs. Deaths are unreinstated automatic revocations, which follow three consecutive missed filings, so quiet dissolutions dominate and mergers are excluded rather than resolved. Every rate carries a Wilson 95% interval and its support; the unit is organizations. Everything here is association: what preceded failure across a population, not what caused it, and a score is triage, never a determination about a specific organization. Form 990 filings lag by 12 to 24 months and figures are dated to their filing; the corpus's most recent complete fiscal year is FY2023, and this design deliberately stops observation at 2021 so both outcomes have room to be observed. Method, leak audit and scripts: docs/analysis-early-warning.md. Source: IRS Form 990 downloads and USAspending.gov.