Among the most turbulent companies with the weakest balance sheets, about one in ten filed for bankruptcy within two years. Among companies that were moving just as violently but carried the strongest books, it was about one in seventy-five.
That is across 1998 to 2024, and it includes three crises.
Everything else in this piece is an argument about when that sentence is true and when it is not, because the honest answer is: not always, and the exception is bigger than we expected.
This one is a headcount
Most of what we publish is about returns, and returns are slippery. A two-year return is a chain of small movements multiplied together, and multiplication does strange things. A company whose share price swings more will tend to show a worse typical outcome even if nothing about the business is worse, purely because of how compounding works on a volatile path. We published a piece that leant on exactly that effect, had it pointed out to us by the researcher whose work we had cited, and withdrew the headline.
This measurement has none of that in it. It is a count of companies that filed. A company either went bankrupt in the following two years or it did not, and no amount of compounding arithmetic changes the answer.
| Balance sheet | Filed within 24 months | 95% interval | Company-months |
|---|---|---|---|
| Strongest third | 0.570% | 0.374% to 0.677% | 233,365 |
| Middle third | 1.249% | 0.720% to 1.707% | 741,554 |
| Weakest third | 5.467% | 3.446% to 7.123% | 233,198 |
Nine and a half times, and the intervals do not touch. The bottom of the weak range sits five times above the top of the strong range. There is no reading of this table where the two groups are the same.
That makes it the most robust thing we have measured.
Where the balance sheet does its work
Now split the same table by how much the share price had been moving beforehand. We sort companies into four turbulence bands, from Calm to Extreme, using only price history available at the time.
| Turbulence band | Strongest third | Middle | Weakest third | Weak ÷ Strong |
|---|---|---|---|---|
| Calm | 0.43% | 0.34% | 0.52% | 1.2× |
| Elevated | 0.35% | 0.41% | 1.73% | 4.9× |
| High | 0.51% | 1.15% | 4.38% | 8.6× |
| Extreme | 1.33% | 3.53% | 10.02% | 7.5× |
Read the weak column downward: 0.52%, 1.73%, 4.38%, 10.02%. That is a nineteen-fold climb from the calmest companies to the wildest. Every one of those is an average over 1998 to 2024 including three crises, and the interval on that last cell alone runs from 6.26% to 13.04%.
Now read the strong column downward: 0.43%, 0.35%, 0.51%, 1.33%. Barely threefold, and it does not really begin until the last row.
So the two things are not independent risks that add up. A weak balance sheet in a company whose price is already moving violently is a different object from a weak balance sheet in a quiet one. The turbulence is what turns the balance sheet into a live question.
And where it does nothing at all
Look at the top row again.
0.43%, 0.34%, 0.52%. Strongest, middle, weakest. It is not a small effect. It is not an effect. The order is wrong — the strongest books have a higher rate than the middle — and the intervals sit directly on top of one another, 0.15% to 0.86% against 0.27% to 0.80%.
This is not an empty cell being coy. There are 254 bankruptcies in one of those groups and 191 in the other. There were enough companies and enough failures for a real difference to show up, and none did.
Among companies whose share price has been calm, we cannot tell you anything about bankruptcy risk from the balance sheet. Not “a weak signal”. Nothing.
We are saying so here because the alternative is to publish a tool that shows a confident balance-sheet risk reading on a calm company, which would be showing you something the data does not support. The reading is being greyed out in the product, with the reason attached, and this is the measurement that forced it.
The one caution in the other direction: the intervals in that row are wide enough that a modest effect could still be hiding. “We found nothing” and “there is nothing” are not the same claim, and we have only earned the first one.
This is what the textbook says should happen, which we did not notice until after we measured it. In a structural default model, the distance to default is roughly the gap between what a company owns and what it owes, divided by how much its value moves. Divide by a small enough number and the gap is enormous no matter how thin it looks on paper. A company whose price barely moves is a long way from default almost by construction, and the balance sheet has nothing left to say about it. The null is not an anomaly. It is the first-order prediction, and we appear to have walked into a clean measurement of it.
The obvious objections
Isn’t this already known? Partly, and the part that is known deserves stating before anything else. Shumway (2001) found that about half the accounting ratios used in earlier bankruptcy models are not statistically significant once market variables are included, and that idiosyncratic return variability is strongly related to bankruptcy. Chava and Jarrow (2004) concluded that accounting variables add little predictive value once market variables are in the model. So the field has held for two decades that volatility beats accounting ratios on average.
What we think is different here is the shape rather than the direction. Those results shrink a coefficient across the whole panel. This one says that in a specific and large part of the cross-section the accounting signal has no content — a flat, unordered null with overlapping intervals on 445 bankruptcies, sitting next to a 7.5x difference in the same table. Average effects and conditional nulls are different objects, and the second is the one that tells you where a model should not be applied.
We have not found this stated this way, and we would rather be told it is old news than keep reinventing it badly. If you know the paper, we would genuinely like the citation.
Is this just small companies? Partly, and we have now measured how much. Company size runs through everything in this panel — we published a piece showing that most of what looks like “strong companies have steadier stocks” is size wearing a costume.
So we ran the same bankruptcy table inside size deciles, formed within each month, so that a $500m company in 1999 and a $500m company in 2024 are not treated as the same size. About a third of the raw gap was size. Holding it fixed turns 9.6× into 6.2×, and three estimators agree: Mantel-Haenszel gives 6.24×, direct standardisation 6.31×, and dropping the two thinnest deciles 6.03×. 6.2× is the number we carry from here. It is smaller than the one we led with, and it is the one that cannot be taken away.
The part we did not expect is that the ratio does not shrink as companies get bigger. It grows. Among the smallest tenth it is 4.1×; by the eighth decile it is 19.6×, and it climbs across every decile thick enough to read. If the effect were size in disguise it would thin out once size was held fixed. It intensifies. The mechanism is not mysterious: among small companies a strong balance sheet buys much less protection, and solid microcaps still fail at 2.19%, four times the rate of solid names overall. Among billion-dollar companies a solid balance sheet is close to absolute at 0.10%, so a weak one stands out by a factor of twenty. Weak beats solid in 10 of 10 deciles, and in all eight thick enough to carry an interval, that interval sits strictly above 1.0×.
Two things travel with that. The top two deciles rest on fifteen solid-company failures each, so the direction there is real but the magnitude is arithmetic and we do not quote it — the same treatment we give the sector table below. And “one in ten” is itself largely a statement about small companies: 27,517 of the 92,631 turbulent-and-weak company-months sit in the smallest decile against 1,073 in the largest. No decile rejects one in ten — every interval covers 10% — but it reads highest among the smallest names, and the honest sentence is “about one in ten, and strongest among the smallest.”
Are the intervals suspiciously wide? They are wide on purpose, and it is the most important technical decision in the piece. Bankruptcies do not arrive evenly. They arrive in waves: 2001, 2008-09, 2020. If you treat each company-month as an independent observation you get beautifully tight intervals that are simply wrong, because the observations are not independent — they are the same three storms counted many times over. Resampling in blocks of calendar time instead gives standard errors roughly seven to eight times wider. Those are the honest ones, and they are the ones above.
Why is there no sector table? There is one. It is not in this piece because it has no intervals yet, and without them one sector reads at 69.6× on the strength of a numerator of two-tenths of a percent, and another divides by zero. A ratio computed from almost no deaths is arithmetic, not a finding. It gets published when it has been bootstrapped, like everything above.
Does the two-year window flatter the result? It is the window we fixed in advance and have used across the study, so it is not chosen to suit this answer. A shorter window would show fewer failures everywhere; a longer one, more. We have not tested whether the ratio holds at other horizons.
What this does not say
It is not a prediction, and it is not advice.
“One in ten” is a twenty-six-year average that includes three crisis periods, not a standing hazard rate and not a forecast for this year. Draw a different set of two-year windows and you get a different number — the interval on that cell runs from 6.26% to 13.04%, and that spread is the honest expression of how much the answer depends on which years you happen to live through.
It says nothing about any individual company. Ninety percent of the companies in the worst cell did not file. A measurement across 1.2 million company-months describes a population; it does not know anything about the name you are looking at.
And it is a measurement of bankruptcy, not of returns. A company can avoid filing and still lose you most of your money. It can also file and have been a fine investment for someone who sold. This piece answers one narrow question with a yes or no, which is exactly why we trust it more than the ones that answer wider questions with a number.
Method
1,208,117 company-months of US-listed companies, 1998 to 2024, each carrying a point-in-time turbulence reading and a sector-relative balance-sheet ranking computed from data available at the time, joined to whether the company filed for bankruptcy within the following 24 months. 23,337 filings. Company-months inside 24 months of the end of the panel are excluded, since their outcome is not yet observable.
Balance-sheet strength is ranked within sector, so “weakest third” means weakest against its own industry rather than against the whole market. Turbulence bands are cut on realised price variance, not on any forecast.
Confidence intervals are block-bootstrapped over calendar time. Naive intervals that treat each company-month as independent are roughly seven to eight times too narrow on this panel and should not be used.
The underlying table is available to anyone who wants to break it. Ask.