What 116 Dead Companies Looked Like 18 Months Before the End
A blind backtest on 116 dead companies, judged only on facts knowable 18 months out, versus matched survivors. What the data shows (July 2026).
A company's terminal event rarely comes out of nowhere. To test whether the warning signs are actually visible in advance, we ran a blind backtest as part of the WillItSurvive assessment of 10,525 companies assessed in July 2026. We took 116 dead companies, masked their identities, and asked our production judges to assess them using only facts knowable 18 months before each company's terminal event. Then we did the same for matched surviving peers at the exact same cutoff dates, with the same masking, and ran the survivor arm twice with independent replicates.
We never claim we predicted these deaths.
The headline numbers
Of the 116 dead companies in the blind backtest (assessed against facts knowable 18 months before each terminal event, results verified July 2026), 90% received a negative verdict and 0% received a positive one, with zero safe-to-join calls across the entire dead arm. The remaining 11 companies, about 9%, landed in the middle bucket.
The survivor arm looked very different. Of the 58 matched survivors judged blind at the same cutoff dates in replicate one, only 17% drew a negative verdict while 24% drew a positive one, against 90% negative on the dead arm (July 2026 backtest). The second replicate on 54 survivors came in at 20% negative and 29% positive. Replicate agreement on the survivor arm was 49 of 53, or 92%, across the two independent runs (July 2026 backtest).
On the 66 dead companies that had a directly paired survivor, the pattern held: 89% negative, 10% middle, 0% positive.
We ran an identifiability probe that tried to name each masked company. On the 17 dead companies the probe could not identify, the blind judges still returned 88% negative verdicts and zero safe calls (July 2026 backtest). That clean subset is small, and the clean survivor subset is smaller still (n=5, mostly middle-bucket verdicts), so we claim recognition-independence only for the dead arm. But the signal on unrecognized dead companies matches the signal on recognized ones (90% negative on the 99 recognized dead companies), which suggests the judgments track fundamentals, not fame.
Who the pairs were
The pairing design put each dead company next to a plausible peer at the same moment in time. Aeropostale was judged at a November 2014 cutoff against American Eagle Outfitters. Argo AI at April 2021 against Waymo. Quibi at April 2019 against Netflix. Yellow Corporation at February 2022 against Old Dominion Freight Line. BlockFi at May 2021 against Coinbase, and Voyager Digital at January 2021 against Gemini. Katerra at December 2019 against Skanska, Convoy at April 2022 against Uber Freight, Jawbone at December 2015 against Fitbit, Juicero at March 2016 against Breville, Rite Aid at November 2023 against CVS Health, Forever 21 at September 2023 against H&M, and InVision at July 2022 against Figma.
The list runs across retail (Bon-Ton vs. Kohl's, Nine West vs. Steve Madden, Stein Mart vs. Burlington, Zulily vs. Wayfair, HHGregg vs. Best Buy, Century 21 Stores vs. Ross Stores, J.C. Penney vs. Macy's), quick commerce (Gorillas vs. Flink, JOKR vs. Getir, Buyk vs. Gopuff, Getir's US operations vs. Instacart), mobility and EVs (Sidecar vs. Lyft, Bird Global vs. Voi, Canoo vs. Fisker, Fisker vs. Tesla, Lordstown vs. Nikola, Drive.ai vs. Cruise, Pearl Automation vs. Mobileye, Local Motors vs. Divergent 3D), health (Babylon Health vs. Teladoc, Cue Health vs. Abbott, Olive AI vs. Epic Systems, Invitae vs. Myriad Genetics, SmileDirectClub vs. Align Technology, Forward vs. One Medical), edtech (Byju's vs. Unacademy, Udacity vs. Coursera, Chegg vs. Course Hero and the reverse pair), fintech (Powa vs. Klarna, Plastiq vs. Melio, Fast vs. Bolt, Daylight vs. Revolut, RealtyShares vs. Fundrise, ScaleFactor vs. Pilot, Reali vs. Opendoor), consumer software and hardware (Rdio vs. Deezer, Hipmunk vs. Expedia, IRL vs. Discord, EyeEm vs. Shutterstock, Humane vs. Rabbit, Quirky vs. Etsy, Jasper AI vs. Copy.ai, Kite vs. Tabnine, Busuu vs. Duolingo, Homejoy vs. TaskRabbit, Atrium vs. Clio), food delivery (Munchery vs. Sun Basket, Sprig vs. DoorDash), plus RadioShack vs. Micro Center, Stack Overflow vs. GitHub, Beepi vs. Carvana, and One Hour Translation/Blend vs. TransPerfect.
The seeding errors that turned out to be signal
The most interesting result came from the pairs where the design went wrong. Ten companies seeded as healthy survivor twins drew negative verdicts from the blind judges in replicate one, and every one of the ten turned out on review to be a genuinely failed or gutted company that had been mis-selected as a healthy peer (July 2026 backtest). The list: Chegg, Course Hero, Fisker Inc., Fitbit, Getir, Flink, Nikola Corporation, Rabbit Inc., Sun Basket, and Unacademy. Both replicates flagged all ten.
In other words, when the blind assessment disagreed with our own seeding, the assessment was right ten out of ten times.
What the misses looked like
The dead-arm miss list, meaning positive verdicts on companies that were dead within 18 months, is empty. The 11 dead companies not flagged negative landed in the middle bucket, our too-close-to-call zone, not the safe zone. That asymmetry matters more than raw accuracy: for a reader deciding whether to join or bet on a company, the costly error is a confident safe call on a dying business, and the backtest recorded zero of those across 116 dead companies (July 2026).
What this means for readers
1,571 contested records carry an independent second opinion. The backtest tells us the machinery can distinguish, at an 18-month horizon and blind to identity, companies that later ended from matched peers that did not. It does not tell us any specific company's future, and we never declare an outcome. Forward-looking material appears only as dated watch-items you can verify yourself.
Questions this data answers
How were the 116 dead companies judged? Each was masked and assessed on facts knowable 18 months before its terminal event, using the same production judges as the live assessment. Matched surviving peers were judged at the same cutoff dates with the same masking, and the survivor arm was run twice.
Did the assessment ever call a dying company safe? No.
An identifiability probe found 17 dead companies it could not name. Those still drew 88% negative verdicts with zero safe calls, so on the dead arm the result does not depend on recognition. The clean survivor subset (n=5) is too small to make the same claim there.
Were survivors treated differently from dead companies? Yes, sharply. Survivors judged blind at the same cutoffs drew 17% negative and 24% positive in replicate one (20% and 29% in replicate two), versus 90% negative and 0% positive on the dead arm.
What about survivors flagged negative? All 10 flagged survivor twins, including Chegg, Fisker, Nikola, and Getir, turned out to be genuinely failed or gutted companies mis-selected as healthy peers. Both replicates caught every one.
Assessments reflect evidence available July 2026. Companies change; treat watch-items as things to verify, not conclusions.
Figures are as of each company's assessment date unless dated otherwise.