Validation

How we tested it — and where it breaks.

Every U.S.-listed company, every month, for 28 years. We checked whether our turbulence signal actually sorted the companies that later failed from the ones that survived — point-in-time, with no peeking at the future.

Educational research about historical patterns. Not a trading recommendation. Reference run 2026-08-06.

The idea, and the test

A company's stock gets choppier in a characteristic way before it gets into real trouble — the same critical slowing down that shows up before tipping points in ecosystems, climate and other complex systems. We measure that choppiness as the rolling variance of daily returns.

The test is deliberately unflattering. We take the entire universe of U.S.-listed companies — not a hand-picked set — and at every point in time ask one question: did the companies that went bankrupt in the next 12 months rank as riskier on this signal than the companies that survived? Every measurement uses only data available at that moment. Acquisitions count as survivals, not wins. Companies that died are kept in at −100%, so nothing is quietly dropped to flatter the numbers.

Why we don't show you a 0.96

You will see other early-warning work quote accuracy scores up around 0.96. We could too — on a small, curated list of famous collapses this same signal scores that high. We don't lead with it, because a curated list is the easy version of the test. The honest number is what the signal does across every company, including the thousands of unglamorous ones, with no cherry-picking.

An AUC of 0.801 means: pick a company that later failed and one that survived at random, and the failing one ranked riskier about 80% of the time. 0.5 is a coin flip; 1.0 is perfect. This is a ranking statistic about the past, not a prediction about any single company's future.

It sizes the door — it doesn't point you through it

Here is the single most important thing the data says. Sort every company into four turbulence bands, then look at the actual range of returns over the next 24 months. As turbulence rises the typical outcome drifts down — but the spread of outcomes explodes. The Extreme band holds the worst wipeouts and the biggest moonshots, and they look identical at the start.

This is why the product is descriptive rather than a buy/sell call: a high band tells you the door is wide — the range of what could happen is large — not which way you will go through it. A second company health axis sharpens the downside within each band. It sorts the trapdoors, not the jackpots, which is why we read the two together rather than turbulence alone.

Why the door widens

That widening is not a claim of ours, and it is not a surprise. It is a known property of compound returns. Multiplicative compounding pushes long-horizon outcomes into strong positive skew even when the individual periods are well-behaved, and the size of a single period's swings governs how fast that happens — the mathematics is set out in Farago and Hjalmarsson, Long-Horizon Stock Returns Are Positively Skewed (Review of Finance, 27(2), 2023).

Which is to say the shape has a name: many poor outcomes and a few very large ones. Our panel shows exactly that, and it strengthens as turbulence rises. The figures below are read straight off the chart above — the same deciles, expressed as a quantile skew, where 0 is a symmetric spread and 1 is entirely one-sided.

Every step up the bands is more skewed than the one below it. By the Extreme band the upper half of the range is more than twice the lower half — and the median asset is down. Both things at once is the whole point: the typical name in that band loses money, and the band's reputation for spectacular winners is earned by a minority of its members.

What this does and does not establish

It gives the widening a mechanism rather than leaving it an unexplained regularity, and the mechanism is somebody else's, published and peer-reviewed. That is worth more than novelty.

It also means we should not overclaim. If skew emerges from compounding at a rate set by volatility, then high-turbulence names showing more 24-month skew is partly what compounding predicts anyway. We are not presenting this as a discovery. What the panel adds is the measurement — across every U.S.-listed company rather than a model or a sample — and the product decision that follows from it: describe how wide the range is, never which way it will go.

The bankruptcy result higher up the page is a separate matter. No property of compounding tells you which companies later failed; that has to be measured, and it was.

It lights up at every crisis

The share of the market sitting in the two highest turbulence bands, month by month, for 28 years. The fast (90-day) lens spikes early and sharply; the slow (trailing-year) lens lags and smooths into a regime read. Shaded bands are the 2001, 2008–09 and 2020 recessions.

COVID is the clearest example: the fast lens hit roughly 97% of the market in weeks while the slow lens barely moved — a true shock that never became a sustained regime. Both lenses ship, because they answer different questions.

Strongest on the biggest companies

A fair worry about any signal like this: maybe it only works on tiny, illiquid junk stocks. The opposite is true. Split the market into ten size buckets and the signal is weakest on the smallest names and strongest on the largest — the reverse of a penny-stock artifact.

Where it doesn't work

Most validation pages skip this part. It is the part that matters most. The signal is not uniform across sectors. In most, turbulence and company strength score best together, so we blend them. In three we don't — and the table says exactly where. In each row the highlighted value is the basis we actually use.

Industrials: turbulence alone scores below a coin flip (0.445) here, so we don't use the price signal at all. We read company fundamentals instead (0.818).

Consumer Defensive and Energy: turbulence alone already scores highest, so blending in the company axis doesn't help. We keep it simple.

We also tested bank- and REIT-specific fundamental recipes to sharpen the weaker sectors. None beat the generic company-strength measure, so we use the simple one everywhere and don't pretend otherwise.

It is not a timing trigger. The signal characterises risk; it does not tell you when something will happen. Both axes tend to light up together, late, rather than one leading the other. Useful for describing how exposed a company is — not for trading a date.

What this doesn't prove

• This is a retrospective study of historical patterns, not a live prospective forecast. The signal was measured against the past, not deployed in real time.
• Results may not generalise beyond 1998–2026 U.S.-listed companies. Different markets or regimes could behave differently.
• Walk-forward testing respects time direction but cannot test against a future regime with no historical analog.
• Nothing here is a trading recommendation. Historical ranking accuracy does not imply forward predictive validity for any individual security.

Method

A point-in-time discrimination study over the full US-listed panel: 1.27M company-months and 9,411 bankruptcies (1998–2026), survivorship-honest and walk-forward. The headline 0.801 is measured on the top-3,000 product universe (958,912 name-months, 3,853 bankruptcies). Fundamentals use as-reported trailing figures joined as-of their filing date — no restatement lookahead.

The Filter Lab is a research program; conclusions follow the data. Analysis by Ryan W. Malone (independent researcher). AI was used as a drafting and engineering aid, with all results computed and reviewed by the author.

If you publish fundamentals yourself

The same discipline is available as a paid engagement. Send a sample of your published output and get back a written report naming each disagreement with the source filing, the arithmetic behind it, and the field-level pattern that produced it. Private diligence, delivered to you, never published.

Data quality review →

See it on live companies How the analysis works