The Bankruptcy Trio as a Short Strategy — A 2012-2026 Backtest

Our bankruptcy-warning study found a three-signal combination sharp enough to earn its own scanner: a company tells the SEC its own published figures can no longer be relied upon, its auditor states substantial doubt about its ability to continue as a going concern, and its short-term debt exceeds everything it holds in current assets. History says roughly half of the stocks that ever lit up all three signals at once fell more than 30 percent within the next three months. The obvious next question is whether a private investor could have turned that into money — not by reading the signal, but by shorting it on a fixed, mechanical rule set anyone could follow.

We built that rule set as literally as a retail account could trade it and ran it survivorship-free from 2012 through 2026, including every stock that later delisted. In the universe of stocks a retail account could realistically borrow and short, it loses money before a single dollar of borrowing fee, and a real-world stock-loan rate finishes the job.

What we built, and what the timeline actually allows

The test reruns the historical scanner logic day by day rather than reading it off today's list: every trading day since the signal became technically possible, we check whether all three conditions were newly true, and treat that day as a trigger. Two versions of the trigger definition run side by side — the narrow definition matching the live scanner exactly, and the broad definition used in the original 792-bankruptcy study, run here as a robustness check. Both reproduce that original study's own published counts exactly: 98 episodes at a 53.06% three-month decline rate under the narrow definition, 115 episodes at 52.17% under the broad one, with zero deviations in either reconstruction.

Three dates matter for reading the results honestly. The going-concern-style balance-sheet check behind this signal only became computable from 29 January 2010 onward, so no Trio trigger can exist before then. The first trigger — under either the broad or the narrow, live-scanner definition — falls on 13 January 2012, which makes the honest test period 2012-2026, not "since 2005" and not "since 2010." (Restricted to the universe of stocks a retail account could realistically borrow and short, that first trigger moves to 30 March 2012 — a liquidity and price filter, not the trigger definition, accounts for the later date, and it is where the primary variant's 61 trades begin.) And triggers are cut off on 31 March 2026, the hard edge of currently available regulatory filings; the stocks that were triggered are then followed forward on price data through 3 August 2026.

Every assumption in the trade construction leans against the strategy rather than for it. Positions enter at the opening print of the first trading day with actual volume after the trigger — never at a theoretical same-day price. A stock can be rolled forward at most 10 trading days (an eleventh day is still allowed) and 21 calendar days before the trigger is marked "not entered" rather than silently dropped. Only one open position per company is allowed at a time; a second trigger on a company that already has an open short is recorded as blocked, not opened. Every cover carries a 0.5% execution cost against the position, and prices are computed in total-return (dividend-adjusted) space, which is the economically correct choice for a short seller, since the short pays the dividend. Splits are corrected only where a data provider actually documents one — a manual check found that 900 of 902 price jumps that an older, cruder detection method had flagged as splits were not splits at all, just real market moves that would otherwise have been erased from the price history.

The trading rules, exactly as tested

Four exit rules run against the same entries and the same 15% stop:

The staggered variant does not move its stop to breakeven after the first tranche closes — the specification called only for the fixed 15% stop, and adding a trailing rule would have been an invented improvement, not a tested one.

Each trading day is resolved in a strict order, and where two outcomes could apply on the same day, the rule favors the less generous reading:

"Open first, then the trading range. If the stop and a profit target are hit on the same day, the stop applies first."

— execution rule, backtest results documentation, 4 August 2026.

Fills happen at the threshold price itself, never at the day's best or worst print; only a price gap through the level fills at the open. A second version of every run repeats the same logic using only the closing price to check the stop, as a check on whether intraday price spikes were doing unfair work — more on that below.

What 61 real trades produced

The pre-registered primary variant — chosen before any run was produced, specifically to avoid picking the best-looking result after the fact — is the -30% target, on the narrow (live-scanner) trigger definition, in the universe of stocks a retail account could actually have borrowed and shorted (last traded price at or above $1, and at least $250,000 in median daily dollar volume over the prior 20 trading days), with the stop checked against the full intraday trading range:

Every one of the four tradeable variants under the narrow definition loses money on average. After 61 trades at one notional unit of exposure each, the primary variant's equity curve sits at -0.945 units across the full 14-year window — spread across 57 different companies, with at most three trades from any single one, so this is not one bad name dragging down the average. Add a matching 0.5% cost on the entry side as well as the exit, and the primary variant's average falls further, to -2.05% per trade.

Widen the universe to every trigger, including the illiquid and effectively unshortable penny stocks a retail account could never actually borrow, and the picture flips on paper: the same -30% target variant shows a capped average of +1.48% per trade across 387 trades. That number is not a hidden opportunity — it is exactly the population this study deliberately excludes from the "tradeable" universe, and it includes trades that lost more than 100% of the position before being capped for the average (a forced buy-in when the loss exceeds the capital at risk) — six to seven such trades per run, seven different companies in total, the worst of them GWTI at -814.55%.

We also reran every combination using only the closing price to trigger the stop, to check whether short intraday price spikes were doing the damage. They were not: the closing-price version of the primary variant raises the hit rate to 37.7%, but the average return gets worse, not better — -2.66% instead of -1.55% — because the losses that do get through the looser stop are larger (one trade in that version lost 103.74%). The stop's timing is not the cause of the weak result.

Why it fails: the arithmetic behind the loss

The primary variant's average return decomposes cleanly: a 31.15% win rate at an average gain of +28.63%, against a 68.85% loss rate at an average loss of -15.20%, nets out to -1.55% per trade. The payoff ratio (a win roughly 1.9 times the size of a loss) sounds attractive, but it sets a hurdle: breaking even needed a hit rate near 34.7%. The strategy achieved 29.51% — short of the number it needed, not by a small margin most people would call noise, but short nonetheless.

The stop-loss decided 37 of the 61 trades, 60.66% of the total. Six more trades — 9.84% — closed only because the company was delisted while the position was open, and even those were not a bonus: their returns ranged from +1.2% to -11.1%, averaging about -2.9%, a softer outcome than the standard stop but still, on balance, a small loss. Only the remaining 18 trades that reached the -30% profit target carried the average, and there were not enough of them.

Two additional facts belong in any honest reading of this result. First, 59.02% of the primary variant's 61 trades — 36 of them — cluster in 2021 and 2022 alone, the post-pandemic speculative unwind. That is not fourteen years of independent evidence; it is mostly one market phase, and the hit rate inside that phase (28.6% in 2021, 18.2% in 2022) ran below the 14-year average, not above it. Second, 6 of the 61 entries (9.84%) would in practice have fallen under a short-sale price restriction the day after a sharp prior decline — a real-world friction this backtest, which has no order book, does not model. Both facts bias the result in the strategy's favor, not against it, which makes the -1.55% average the more, not less, credible number.

The number that finishes it off

None of the above prices in the cost of actually borrowing the stock to short it. We modeled that cost the simple way brokers do: an annual rate times days held, divided by 365, subtracted from the trade's return.

Six of the eight tradeable rule variants are already negative with no borrowing fee at all. The two that were not — a +40%-target variant at +2.13% per trade and a staggered-exit variant at +0.27% per trade, both measured under the broader trigger definition — turn negative at borrowing rates of just 7.4% and 0.9% per year respectively. That is not a stress scenario; it is close to the floor of what a lender would actually charge for a normal, liquid short.

And these are not normal, liquid shorts. The companies in this test are overwhelmingly small and already in financial distress — exactly the group where stock loan is expensive, thin, or simply unavailable, and where a lender can force a buy-in in the middle of a trade. This backtest has no borrow-availability data and cannot measure that risk directly. What it can show is the sensitivity: the calculated edge is thin enough to disappear inside the first few percentage points of a real borrowing fee, which means that for any actual implementation, borrowing cost would not be one line item among several — it would be the deciding one.

What this means for the original signal

This result does not walk back the original bankruptcy-warning study. As a warning signal, the Trio combination remains one of the sharpest we measured: historically, roughly 52 to 53 percent of its triggers were followed by a 30-percent-or-worse price decline within three months, several times the base rate of a matched control stock with the same price, the same trading volume, and the same distance from its 52-week high but no active signal. That finding stands untouched.

What this backtest adds is a separate, harder question: does a fixed, mechanical trading recipe built on top of that signal — a specific entry timing, specific profit targets, a specific stop — actually convert the edge into money once real trading frictions are included? The answer here is no, at least not for this recipe. The signal is doing its job: stocks that trigger it really do fall more often and harder than similar stocks that don't. What defeats the strategy is the arithmetic of turning a probabilistic edge into a fixed-rule trade — a stop that closes 60.66% of positions is not automatically compatible with a signal that is "only" right roughly half the time on the underlying price move, and the mismatch between those two numbers is the whole story.

"The Trio was selected using the same data this backtest measures it against. The result is confirmatory, not evidence out of sample, and not a forecast."

— in-sample disclosure, backtest results documentation, 4 August 2026.

What this does, and does not, prove

It does not prove that these stocks can never be shorted profitably. It proves that this specific, fixed mechanical recipe — next-open entry, fixed 20/30/40% targets, a fixed 15% stop, no position sizing by signal strength, no trailing stop, no discretion — loses money in the tradeable universe once realistic borrowing costs are added, over the full 2012-2026 period this signal has existed. A different exit design, a time-based exit, position sizing tied to signal strength, or an options-based structure were not tested here and might behave differently; we make no claim about them either way. And as stated above, this is a confirmatory, in-sample test of a rule built on the same data it is measured against, not an independent replication.

The three underlying warning signals continue to run live in our Bankruptcy Study: The Sharpest Three-Signal Combination scanner, unchanged by this result — it remains a measured warning list, not a trading system. The original signal study, with the full breakdown of lead times, false-alarm rates, and the matched-control comparison behind the 52-53% figure above, is here: The Warning Signs Before Bankruptcy — 792 U.S. Cases Since 2005. Both studies, and everything that follows in this series, live in Studies.

This article is a historical analysis and not investment advice. It contains no buy, sell, or short recommendation, no forecast, and no statement about any individual company listed today. Short selling carries risks beyond those of a long position, including potentially unlimited loss and the possibility of a forced buy-in regardless of price. Anyone making investment decisions should assess their own situation and risks — if in doubt, with professional advice.