ID: S046
Slug: spx-0dte-vol-conditioned-condor-v1
Failed at: Pre-freeze deflation gate (Stage 1–2, historical backtest)
Run date: 2026-07-12
Outcome: FAILED (FAILED)
Headline metric: DSR 0.02 (VIX≥17) / 0.22 (VIX≥20) @ N=30 trials — both below the noise benchmark SR₀ = 0.286; PBO = 0.43 (bar < 0.30)
Fail reason: The in-sample per-session Sharpe of the vol-gated book (0.138 / 0.212) is below the Sharpe you'd expect to find by luck alone from ~30 searched trials (SR₀ = 0.286). The edge sits under the overfitting waterline; PBO 0.43 confirms the threshold selection is a coin-flip. Caught before freezing, so no forward slot was spent.
Verdict
Failed the deflation gate on a historical backtest — a searched vol-deployment threshold that looked strong in-sample collapses once it is charged for the search (DSR below the noise benchmark, PBO 0.43). Overfit, killed before going live.
What we tested (backtest, not forward)
A locked configuration — no re-tuning: managed near-ATM SPX 0DTE iron condor, symmetric 0.2% shorts, 0.35% wing, parity-centered at the ~09:35 entry, uniform 25% profit-take, held to cash settlement, no stop or time-exit (both tested and ruled out — the losers mean-revert and the wing is already the stop). Deployment gate: trade only when VIX-open ≥ {17, 20} (a two-cell grid). All P&L is on the assembled historical intraday 0DTE surface (Databento OPRA 2022→mid-2024 + Massive, 2022-06 → 2026-05, ~546 fillable sessions) with a fair, moneyness- and vol-scaled cost model — not forward/live data (S046 never reached forward paper). Deflation charged at N = 30 trials, PBO < 0.30 (both owner-set). Validators: the lab's own López-de-Prado deflated_sharpe and pbo_cscv.
What we found
| Deflation gate | Required | Observed (VIX≥17 / VIX≥20) | Verdict |
|---|---|---|---|
| Deflated Sharpe (DSR) @ N=30 | > 0.95 | 0.02 / 0.22 | FAIL |
| — observed per-session Sharpe vs SR₀ | > SR₀ = 0.286 | 0.138 / 0.212 (both below) | FAIL |
| PBO (CSCV, 10 splits, 10 configs) | < 0.30 | 0.43 | FAIL |
| Deployed CVaR₅ vs always-on benchmark (−$1,251) | ≤ benchmark | −$1,050 / −$905 | pass (necessary, not sufficient) |
| skew / kurtosis (per session) | — | −1.84 / 6.4 | (fat left tail — deflates hard) |
The one thing that passes (tail improves vs always-on) is necessary but nowhere near sufficient: dropping the worst regime almost always improves a high-win-rate book's average tail — that isn't evidence of a real edge once the search is charged for.
The backtest record — $25k account, 1 contract per deployed session
Historical backtest, full 0DTE surface, fair cost. The gated books look great — which is exactly why the search was tempting — but that separation is search luck, not a regime (see the deflation gate above):
| book | end value | max drawdown |
|---|---|---|
| always-on (trade every session) | $35,122 | $11,210 |
| deploy VIX ≥ 17 (searched) | $41,232 | $2,642 |
| deploy VIX ≥ 20 (searched) | $40,927 | $2,642 |
Start $25,000; avg defined risk ≈ $939/contract (~3.8% of the account); worst single session −$1,542 (~6%).
The per-session P&L distribution of the VIX-elevated book (below) is the classic left-skewed short-premium shape: a cluster of small winners banked by the profit-take, an 84% win rate, and a thin deep-loss tail — the negative skew (−1.84) / fat kurtosis (6.4) that the Deflated Sharpe charges hardest.
Why it failed, in one line
A searched, non-monotone threshold is the classic overfit — a plausible mechanism, an in-sample Pareto win, and even a second-measure replication were not enough to survive an honest N=30 + PBO charge on a backtest.
What we learned
Run the deflation gate before you freeze: it turned an 8–10-month forward-paper commitment into a same-day null. A plausible mechanism, an in-sample Pareto win, and even a second-measure replication were not enough — a searched, non-monotone threshold is the classic overfit, and only an honest N=30 + PBO charge exposes it.
What this doesn't tell us yet
The danger-zone U-shape is real in-sample and even replicated by an independent 1-day-vol measure — so this does not say a conditional variance-risk-premium effect is absent, only that a threshold searched on this curve is indistinguishable from luck once charged for the search. Whether a threshold fixed a priori would survive on fresh data is untested here. And it is a backtest on an assembled surface, not forward marks — live fills and the real intraday tail remain unmeasured.
What happens next
No forward-paper slot was spent — the deflation gate turned an 8–10-month forward commitment into a same-day null. If the conditional-VRP idea is pursued, the only honest route is a fresh setup with the wing and threshold fixed a priori and confirmed on a genuine holdout — never a re-search of this same curve.
Slice & dice
For the specialist — methodology details (click to expand)
The vol-edge curve (fair cost, VIX-open, 444 non-vault train sessions): calm (<13) ≈ +$21/session, mid-vol 15–17 = −$183 (the danger zone), high vol 25+ = +$126. VIX1D prior-close (independent 1-day vol) replicates the danger zone (13–15 = −$229). Deploying ≥17 lifts ann Sharpe 0.25→1.12 and tail-aware Kelly φ* 0.19→0.34; ≥20 → 1.39 / 0.38. All in-sample. The deflation benchmark SR₀ = 0.286 is the expected-maximum per-session Sharpe from 30 trials under the null (via the trial-Sharpe dispersion across the searched wing × threshold grid); the observed 0.138 / 0.212 fall below it, so Deflated Sharpe Ratio (DSR) < 0.5 and the edge does not clear significance. Engines: the exploration workspace. Validators: the engine. Spec: the internal file.