ID: S044
Slug: spx-0dte-managed-condor-v1
Failed at: Stage 3 (forward paper) — invalidated by the same settlement-proxy bug as S045
Run date: audit 2026-07-20 · forward paper 2026-07-13 → 2026-09-11 (44 sessions, run to full term)
Outcome: FAILED
Headline metric: corrected backtest Sharpe −1.2, mean −$41/session, 76% win — and still negative
Fail reason: Identical to S045. The backtest settled each 0DTE expiry off spx_close = SPY × 10, which runs 9–20 SPX points below the real index. Rallies into the upper wing were booked as wins instead of max-losses. Priced off the options' parity close, the condor loses at every short-offset and wing tested, before costs.
What we tested
A near-ATM short condor entered 09:35 ET, managed with a 25% profit-take and a time-stop, no active price stop (max loss = the wings). One contract. Stages 1–2 on the unified four-year SPX 0DTE surface including the 2022 bear market; Stage 3 forward on live quotes with real frictions.
What we found
Corrected backtest. Negative at every short-offset and wing width, before costs. The 76% win rate is what made the uncorrected version look strong, and it is precisely what a tail-aware gate is built to discount.
Forward paper — positive, and uninformative.
| forward paper (44 sessions) | |
|---|---|
| total | +$5,009 |
| mean per session | +$114 |
| win rate | 38 / 44 (86%) |
| standard deviation | $685 |
| standard error of the mean | $103 |
| 95% interval for the true mean | −$93 … +$320 |
| corrected backtest expectation | −$41 — inside that interval |
| worst session | −$1,976 |
| CVaR₂₀ | −$954 |
The forward mean is 1.5 standard errors from the backtest expectation — well inside noise. And the tail is the tell: one session lost $1,976, roughly seventeen times the average session's gain. Eighteen average winners are undone by a single bad day, and 44 sessions is nowhere near enough to see how often those days come.
What we learned
86% winners is the warning, not the evidence. This setup produced the highest win rate anywhere in the 0DTE family and the deepest single-session loss in its forward book. Read on win rate it passes comfortably; read on conditional value-at-risk (CVaR) it does not. The gate was tail-aware, and that is the only reason this record is honest.
One bug, one family. The condor and the fly are different trades that shared a settlement helper. Correcting one corrected both — and invalidated a third study built on top of them (S049). A shared utility inherits its error to everything downstream, which is an argument for auditing the boring shared code first, not the strategy logic.
Scheduled stops need to run where they can reach the machine. The plan was to stop and evaluate this book on ~2026-09-07, delegated to a cloud routine scheduled for 2026-09-08. That routine had to SSH to the VPS, which the cloud egress gateway blocks, so the stop never happened and the book kept trading until 2026-09-11. The same blocker had already broken the weekly report and the site-health scan. Automation belongs on the machine it needs to touch.
What this doesn't tell us yet
Whether the condor's wider short offset would have held up over a longer forward run. The corrected backtest says no at every offset tested; the 44-session forward book says nothing either way, at ±$103 per session.
Whether an 86% win rate with a −$954 CVaR₂₀ could ever be sized into something acceptable. That is a position- sizing question we did not pursue, because the corrected expectation is negative — sizing cannot rescue a negative mean, it only changes how fast it arrives.
What happens next
Nothing. Closed, flag off as of 2026-09-12, scheduler jobs being removed. The captured chain data stays: it feeds the positioning-timing study, which asks a different question.
For the specialist — methodology details (click to expand)
Shared root cause. The condor and the fly are different trades that called the same settlement helper. See the S045 report for the mechanism — SPY × 10 running 9–20 SPX points below the true PM settle, moving outcomes across the payoff kink rather than adding symmetric noise.
The forward statistics. n = 44, mean +$113.8, population sd $685, standard error $103.2, 95% interval −$93 … +$320. The corrected expectation of −$41 sits 1.5 standard errors away and inside the interval. Tail: worst −$1,976, CVaR₂₀ −$954, 38/44 winning. The worst single session is ~17× the mean session gain, so roughly eighteen average winners are erased by one bad day — and 44 sessions cannot establish how often such days arrive.
On the missed stop. The scheduled stop (~2026-09-07, delegated to a cloud routine on 2026-09-08) required
SSH access to the production host, which the cloud egress gateway blocks. The book therefore traded four
sessions past its pre-registered term, through 2026-09-11. Those sessions are included above; excluding them
does not change the conclusion. The operational lesson is recorded in lab/candidates/ and in the ops notes:
a scheduled stop must run on the machine it needs to touch.
Related
S045 (ATM fly, same root cause), S049 (dual-fly diversification, built on S045 and void with it), S048 (return-stacking — its 0DTE sleeve is this family and fails with it).