Put-selling on capitulation in an uptrend

ID: S008 Slug: put-selling-capitulation-uptrend-v1 Failed at: Stage 1 (Quick-screen) Fail reason: magnitude Date: 2026-05-30

What we tested

Two distinct "capitulation" triggers, both requiring the stock to be trading above its 200-day moving average (uptrend filter) and at elevated realized volatility (≥ 30th percentile of its own past year):

• Variant A — intraday capitulation: a single down day of 2%+ that's also larger than 1.5× the stock's typical daily range (statistically unusual single-session flush).
• Variant B — multi-day capitulation: a cumulative 8-15% decline over the trailing 15 trading days, with the rate of decline slowing in the last 3 days (suggesting the slide is exhausting itself).

For every signal, we simulated selling one 45-day put at strike roughly 7% out-of-the-money. The position was held with three exit rules in priority order: take 25% of the premium within the first week, take 50% any day after that, or force-close at day 21. We priced the put using Black-Scholes with the underlying's recent realized volatility (× 1.15 as a standard adjustment for OTM put skew) and subtracted realistic trading costs ($1.30 commission + 5% slippage on the close).

The pre-set criteria — frozen before the test ran — required a positive expected value per trade, high win rate, statistical significance, and crucially no single year contributing more than 30% of total profit (the catastrophic-loss check). For 11 years of data across ~2,100 tickers (top-1,000 US equities monthly PIT + 54 fixed ETFs), the test produced 14,467 Variant-A trades and 23,915 Variant-B trades.

What we found

Both variants failed, but in importantly different ways.

Variant A — intraday capitulation

Variant A is unambiguously cut. The signal is statistically distinguishable from zero, but in the wrong direction — the strategy loses money on average. The two halves of the sample produce opposite signs (early years negative, later years positive). 2020 alone lost $4.6M in simulated capital, which is more than 3× the total net loss across all 11 years.

Variant B — multi-day capitulation

Variant B is the interesting cut: 6 of 7 pre-set criteria pass. Per-trade expected value is positive and statistically significant with overwhelming sample size. Both halves of the sample are positive. The only failure is the year-concentration check — and that one failure is decisive.

Why the concentration check is the binding criterion

The dollar P/L tells the story the per-trade stats hide:

Variant B year-by-year net realized profit
  2016:  +$45,600
  2017: +$120,586
  2018: +$141,248
  2019:  +$74,194
  2020:  +$40,453
  2021: +$185,588
  2022:  −$2,837,344  ← rate-shock year, bonds + cyclicals capitulated
  2023: +$153,621
  2024: +$421,052
  2025: +$119,463
  2026: +$186,009
  ─────────────────
  TOTAL: −$1,349,530

Ten of eleven years are positive. One year (2022) loses more than twice the absolute value of the eleven-year net. The trader running this strategy on real capital ends 2026 with less money than they started with, despite a per-trade win rate of 87%.

This is the textbook short-volatility fat-left-tail signature. Selling out-of-the-money puts is selling insurance against downside; in most years that insurance is overpriced and you collect the premium. In years when capitulations cluster (2020 COVID, 2022 rate-shock), the same trade structure means your losses pile up faster than your winners can compound. The strategy generates positive expected value per trade and a negative realized P/L over time.

Why this matters / what surprised us

The pre-registered gates worked exactly as designed. A naive backtest reporting "87% win rate, +0.88% mean per trade, statistically significant" would have called this a winner and shipped it. The 30%-concentration check is the difference between "looks good" and "actually makes money over a decade including fat-tail years". Future short-vol setups in this lab will inherit this check as a default rather than something we have to remember to add.

Variant B's signal might still have an exploitable form — but not in the structure we tested. Three honest reframings worth a new H-NNN pre-registration:

1. Sized for tail-risk: same signal but with position sizing that targets a fixed worst-case loss (e.g. notional sized so a 100% gap to zero only loses 1% of capital). The per-trade EV becomes smaller in absolute dollars but the catastrophic year stops being fatal.
2. Defined-risk structure: replace naked short puts with put credit spreads (sell the 93%-OTM put, buy the 85%-OTM put). The premium drops by maybe a third but the worst-case loss is bounded. Setup becomes: does the residual edge after buying tail protection still clear a tightened gate?
3. Vol-of-vol gate: only fire the signal when SPY's VIX is below its own trailing median (i.e. avoid selling vol when vol-of-vol is rising). The 2022 disaster occurred during a period of rising vol-of-vol; this gate might filter the worst subset.

These would each be a new pre-registered setup (S009 / S010 / ...), not a modification of S008. The current cut stands.

The RV-rank proxy was probably not the failure point. When we considered Stage 2, we noted that the realized-vol-rank used as a proxy for true IV-rank could itself be the reason a Stage-1 result disappoints. The cross-check at RV-rank ≥ 0.50 (informational only, not pass/fail) produced essentially identical numbers (Variant A mean -0.12% vs -0.15%; Variant B mean +0.98% vs +0.88%). The catastrophic-tail problem is structural to short-vol-into- capitulation, not an artifact of the proxy. A true IV-rank reconstruction would not save this setup.

What this doesn't tell us yet

1. A defined-risk version isn't ruled out by this test. A put credit spread or a put butterfly caps the worst-case loss and may avoid the 2022-style year-concentration. Worth pre- registering as a separate setup with its own gates.
2. A position-sized version isn't ruled out either. The simulation here used uniform 1-contract sizing per signal, which means a $30 stock and a $300 stock contribute equally to per-trade EV but very differently to dollar P/L. A notional-weighted or risk-weighted version would change the year-concentration dynamics.
3. The signal itself (capitulation in an uptrend) is not debunked. The directional intuition — that prices recover after panic flushes in stocks that remain in longer-term uptrends — is supported by 87% win rate and statistically significant positive per-trade EV in Variant B. The cut is on the trade structure, not on the underlying observation.

What happens next

S008 is closed. The strategy is failed at Stage 1.

If you want to pursue the underlying observation further, the suggested follow-ups (in rough order of "most likely to survive the same gates"):

• S009 candidate: put-credit-spread-capitulation-uptrend-v1 — same signal, defined-risk structure (spread instead of naked short put). Pre-register before running.
• S010 candidate: put-selling-capitulation-vol-of-vol-filter-v1 — same naked-put structure but add a "VIX not rising" gate to filter out the 2022-style environment. Tests whether the bad years can be identified ex-ante.
• S011 candidate: put-selling-capitulation-risk-weighted-v1 — same signal and structure but with notional sized to cap per-trade worst-case loss. Tests whether the edge survives at more realistic position sizing.

Each would need its own pre-registered gates and a fresh H-NNN.

VARIANT A:
  mean_ev_pct_notional   ≥ +0.30%
  win_rate               ≥ 80%
  welch_p_value          < 0.05 AND mean > 0 (direction-aware)
  n_signals              ≥ 500
  half1_mean_ev > 0 AND half2_mean_ev > 0 (split at median date)
  max_year_share         ≤ 30%
  max_ticker_share       ≤ 5%

VARIANT B:
  mean_ev_pct_notional   ≥ +0.40%
  win_rate               ≥ 82%
  welch_p_value          < 0.05 AND mean > 0
  n_signals              ≥ 400
  half1_mean_ev > 0 AND half2_mean_ev > 0
  max_year_share         ≤ 30%
  max_ticker_share       ≤ 5%

CROSS-CHECK (informational only, does NOT affect pass/fail):
  Same metrics with RV_rank ≥ 0.50 threshold