The 18-Month Window: How Early Trouble Shows Up in Public Facts

A company rarely ends suddenly. Bankruptcies, shutdowns, and fire sales are usually preceded by a long trail of public facts: layoffs, going-concern language, executive departures, delisting notices, funding rounds that never close. The question we set out to test is simple. If you froze the clock 18 months before a company's terminal event and looked only at what was publicly knowable then, would the trouble already be visible?

To find out, we ran a blind backtest against the WillItSurvive assessment of 10,525 companies assessed in July 2026. This article explains what the test was, what it showed, and what it did not show.

How the blind backtest worked

We took 116 dead companies from our graveyard set and rebuilt each one's record using only facts knowable 18 months before its terminal event.

Judging dead companies alone is not enough. A pessimistic judge could flag everything and look accurate. So each dead company was paired where possible with a matched surviving peer, judged blind at the same cutoff date under the same masking, with two independent replicates on the survivor arm. Aeropostale was paired with American Eagle Outfitters at a November 2014 cutoff, Yellow Corporation with Old Dominion Freight Line at February 2022, Quibi with Netflix at April 2019, Convoy with Uber Freight, BlockFi with Coinbase, Katerra with Skanska, Jawbone with Fitbit, and so on across 66 pairs.

What the dead arm showed

In the blind backtest assessed July 2026, 105 of 116 dead companies (90%) received a negative verdict on facts knowable 18 months before their terminal event, and zero received a positive one. Not a single company in the dead arm was rated called safe. The remaining 11 landed in the middle bucket. In the paired subset of 66 dead companies, results were nearly identical: 89% negative, zero positive.

This holds across very different failure modes. Retail chains with visible same-store decline (Bon-Ton, Stein Mart, Nine West), venture-funded burn stories (Katerra, Fast, Quirky), crypto platforms judged before the 2022 crash reached them (BlockFi at a May 2021 cutoff, Voyager Digital at January 2021), and hardware bets like Juicero and Humane's AI Pin all showed enough public strain 18 months out for blind judges to flag them.

The survivor control: distress is not a death sentence

The survivor arm is what makes the result meaningful. In the first replicate, only 17% of 58 matched survivors received a negative verdict and 24% received a positive one, judged blind at the same cutoff dates as their dead twins. The second replicate on 54 survivors was similar: 20% negative, 29% positive. Across the survivor arm, the two independent replicates agreed on 49 of 53 companies (92%) in the July 2026 backtest.

So the pipeline is not simply pessimistic. It separated companies that were dead within 18 months from matched peers facing the same market at the same moment, and it never once issued a safe rating on a company that was dead within 18 months. We publish this carefully: the supported claim is separation on pre-collapse facts, not prophecy. We never declare an outcome for a living company.

The most interesting errors were not errors

Ten "survivors" were flagged negative by the blind judges in both replicates: Chegg, Course Hero, Fisker, Fitbit, Getir, Flink, Nikola, Rabbit Inc., Sun Basket, and Unacademy. On review, each of these ten turned out to be a genuinely failed or gutted company that had been mis-selected as a healthy peer when the pairs were built. The blind assessment caught all ten seeding errors, in both replicates, before any human noticed the mismatch. Fisker, for example, had been proposed as the "survivor" twin for Canoo at a July 2023 cutoff; it was already on the same trajectory.

A fair objection: maybe the models recognized Quibi or RadioShack through the masking and recalled the ending. We ran an identifiability probe that tried to name each masked company. On the unrecognized "clean" subset of 17 dead companies, results barely moved: 88% negative verdicts and zero safe ratings in the July 2026 blind backtest. That supports recognition-independence on the dead arm. The clean survivor subset was only 5 companies, too small to claim the same for survivors, and we say so plainly.

What this means for readers

The practical lesson is that the 18-month window is usually legible. Layoff cadence, funding gaps, customer concentration, auditor language, and leadership churn are public, dated, checkable facts. That is why every one of the roughly 10,800 answer cards in the assessment carries a WATCH section: dated items a reader can verify themselves rather than a forecast from us. Of the 10,525 companies assessed in July 2026, 1,571 contested records also carry an independent second opinion from a separate model, because close calls deserve a second look.

We do not publish predictions. What the backtest supports is narrower and, we think, more useful: given only pre-collapse public facts, a disciplined assessment separated the doomed from matched survivors, and the fatal cases were visible well before the end.

Questions this data answers

How accurate was the blind backtest on companies that later died? Of 116 dead companies judged blind on facts knowable 18 months before their terminal event, 90% received a negative verdict and zero received a positive one (July 2026 backtest). Not one was rated called safe.

Was the judging pipeline just negative about everything? No. Matched survivors judged blind at the same cutoff dates were rated negative only 17% of the time in replicate one and 20% in replicate two, with roughly a quarter rated positive. The separation between the two arms is the result, not the raw negativity rate.

Could the models have recognized the masked companies and recalled their fate? An identifiability probe found a subset of 17 dead companies the models could not name. That clean subset was still 88% negative with zero safe calls, supporting recognition-independence on the dead arm. The clean survivor subset (n=5) is too small to extend the claim.

What were the survivor "false negatives"? Ten survivor twins flagged negative in both replicates (including Chegg, Fisker, Nikola, and Getir) each turned out to be genuinely failed or gutted companies mis-selected as healthy peers. The blind judges identified all ten seeding errors.

Does this mean WillItSurvive predicts company deaths? No. We never declare an outcome; forward-looking material is published only as dated watch-items with instructions to verify. The backtest shows past separation on historical facts, which is evidence about the method, not a forecast about any current company.

Assessments reflect evidence available July 2026. Companies change; treat watch-items as things to verify, not conclusions.