The bottom is usually not the bottom.

The rule was frozen before the run and hasn't been tuned since. No fundamentals, no household filter — the chart rule alone:

1. A 30%+ decline over 3+ months. The peak is the highest monthly high in the 36 months up to and including the trough.
2. A confirmed monthly swing low. A later monthly bar trades above the trough month's high while the low holds. A confirmation that takes several months counts the same as a fast one.
3. Two entries. The close of the confirming month, and — what the live scanner actually flags — the close of the first weekly pullback low that confirms after it (median wait 3.0 weeks). They came out the same, so the tables below use the first.

Detection runs on the same engine as the live scanner, and it confirms a low from later bars only, so the replay is causal. One event per ticker per twelve months, so no two windows overlap (20,289 clustered events dropped). Split-adjusted Alpaca daily bars, aggregated to weekly and monthly; the still-forming period is excluded.

Read the per-year counts before any headline number. Ten years of daily history covers four real declines — late 2018, 2020, 2022 and 2025 — and 2020 and 2022 dominate. One blended rate hides that.

A confirmed monthly low is a signal, not a floor. Of 14,343 events old enough to judge, 8,116 (56.6%) were undercut within twelve months of the entry, and 68.4% at some point in the remaining tape. 1,014 more are too recent for a twelve-month window to have elapsed and are left out of the denominator rather than counted as clean.

That's the honest denominator for the whole false-bottom story, and it's the number to hold onto: the trap is the norm, not the exception. It's why a name that has already faked one bottom is worth more attention than a first-time low (below).

Everything else on this page blends a universe far wider than the one the tool runs on. The live scanner only looks at names above $10B, and it treats a prior false bottom as a booster. This is that cell, and it's the one to quote.

By entry year. The best cell in the study, and 362 of its 1,970 events (18.4%) are 2020 entries. One blended number would hide how much of it is one recovery.

• A prior false bottom is the strongest signal in the study. Twelve-month median +8.3% with one vs +0.0% without; undercut 52.1% vs 61.6%. A low that comes after the first buyers were already stopped out holds better. The core thesis survives.
• Late confirmation is safer, not worse. Identical returns, undercut 50.3% for a late confirmation vs 58.7% for a next-bar one. A low that consolidates for a few months before it clears the trigger is a better low.
• Deeper isn't cheaper. Declines of 50%+ recovered no better than declines of 30–50% (twelve-month median +4.5% vs +4.6%), and with far wider dispersion. "It's down 60%, it must be a bargain" isn't supported.
• The balance sheet didn't predict the rebound. A 0–100 health score computed from SEC filings as of the entry date (nothing filed later is used) sorted large caps into Strong / Solid / Mixed / Weak. The bands returned within a few points of each other, in no useful order. It's a sanity check on the business, not a return predictor.

The S&P is the wrong control for the rule itself. The alternative facing a Phoenix candidate isn't the index — it's buying the fallen name without waiting for the low to confirm. So we ran that too: buy the close of the first month a stock is 30%+ below its 36-month high, three months on, knowing nothing more. Its own population — 19,445 entries, including every decline that never confirmed a low because it kept falling or the tape stopped.

• Undercut within twelve months: 60.2% buying naively vs 56.6% waiting for confirmation (53.8% vs 47.9% in large caps).
• Twelve-month median: +4.6% naive vs +4.5% confirmed. In large caps the naive entry did slightly better (+16.1% vs +14.2%), because it bought lower.
• The price of waiting: the confirmed entry sits a median +10.6% above the trough month's close (mean +23.4%). Confirmation trades roughly that much entry price for a few points less chance of a lower low, and the same return a year out.
• Even hindsight only gets you to even. Buying the trough month's own close — which nobody can do on purpose — reaches 48.1% beat-the-S&P. The ceiling on this whole idea is the index.

Everything above is a buy-and-sit-for-a-year measurement, which nobody trades. So the same events were traded the way a swing trader would: a stop under the low, a reversal exit, a hard 24-month time stop, each trade benchmarked to the S&P over its own holding window. The stop distance is treated as a gate — a candidate whose invalidation sits too far away is rejected, not taken at bad odds.

Every management rule made it worse. Mean R runs 1.12 with a stop under the pullback low and no gate, 0.84 at an 8% gate, 0.76 at 5%: the tighter the risk control, the less the setup returns per unit of risk. A stop within 5% of entry on a stock that just fell 40% is hit 87.5% of the time — it sits inside the noise band. The reversal exit is the single most expensive rule. Every gated variant is still positive-expectancy, and none of them beats the index.

The entry at the confirming month's close isn't a tradable entry at all: the stop sits a median 18.3% away, so almost any sane risk filter rejects it. Only the weekly-pullback entry offers a stop close enough to gate on.

The recovery story is a large-cap story. Below $2B the median event was down a year later (-7.2%), 31.5% beat the index, and 63.9% of confirmed lows were undercut. A broken small cap mostly stays broken — which is the Broken List's whole premise, measured from the other side.

The universe is today's Nasdaq screener above $2B union every ticker we've ever recorded. 7,622 of 15,357 events (49.6%) come from names not on today's screener — the ones a today's-list-only backtest silently drops, and dropping them flatters every rebound number. A company that delisted before we began recording is still missing; this mitigates the bias, it doesn't remove it.

• A window that runs past the end of the data is censored — excluded, never counted as flat. A stopped tape is a realized outcome at the last traded close. Re-pricing every delisting at −100% moves the twelve-month median by a tenth of a point.
• 859 events have their peak on the first monthly bar we hold, so their real decline may have begun before the tape; their depth is a floor, not a measure.
• Forward horizons are calendar months from the entry date to the last trading day on or before the target. Dividends are ignored on both sides.
• This is a descriptive measurement of a past sample over four regimes in ten years. It isn't a forward claim, and the sample is far too small per cohort to be one.