0DTE Bot — Full entry-simulation backtest
The ENTIRE strategy simulated from raw data — GEX score, permission gate, contract
selection, entries, exits — nothing taken from the bot's ledger · worst-case fills headline ·
generated 2026-07-30
Verdict. Simulated end-to-end over the 10 faithful GEX days (Jul 16-29,
2026; 12 OPRA days minus 2 with no SPX/NDX GEX recordings), the full strategy under
worst-case fills (buy at the real ask, sell at the real bid, $1.32 fees/round-trip)
lost -10,075$ across
1114 simulated entries. The split is the whole story:
SPX made +42,589$ (732 trades, +58$/trade) while
NDX lost -52,664$ (382 trades, -138$/trade) —
NDX median entry spread is $3.60 vs $0.10 on SPX; the strategy's edge does not survive the
NDX spread. In the bot's own optimistic mid-fill accounting the same simulation shows
+137,232$ — the gap to worst-case is pure spread cost.
Peak simultaneous premium at risk: ~64,080$.
Validation A — is my score engine the bot's score engine?
At every recorded live decision timestamp I recomputed score and permission from point-in-time snapshots only, then compared with what the bot recorded (
Jul 16-23 recorded decisions used an OLDER rubric (
At every recorded live decision timestamp I recomputed score and permission from point-in-time snapshots only, then compared with what the bot recorded (
trade_permissions, 2988 rank-1 decisions).
On the three days where the DB stores the bot's exact scan inputs AND the current
rubric was live (Jul 27-29, gex-rubric-2026-w31-v1, n=449):
score match 95.3%, permission match
98.4%, technicals-adjustment match 100.0%
(gate: ≥90% — PASSED).
Contract Pick #1 match at the same timestamps: 42% overall —
SPX 65%, NDX 20%; mismatches are the same
side and typically 1-2 strikes away (SPX median miss 5 pts, score delta ~0.5) — the
cost of 1-minute NBBO vs the bot's 1-second stream.Jul 16-23 recorded decisions used an OLDER rubric (
legacy-unversioned) and their snapshots were later overwritten by
GEXBot day-blob rebuilds (different sub-minute sampling), so recorded-score agreement
there (12%) measures rubric drift + vendor resampling, not
engine error; those days are still simulated — with today's rubric, as a backtest
should.
Validation B — do my simulated entries look like the bot's real entries?
On Jul 27-29 the bot really traded: 245 live SPX/NDX entries. My simulation, run blind on the same days, produced 313 entries and matched 116 of them on exact (contract, 30-min block). It covered 86% of the bot's SPX entries but only 24% of NDX ones — same-side, near-strike differences driven by 1-minute quotes and (on Jul 27) a denser decision grid than the bot's actual ~2-minute live scan rate. The simulation trades MORE than the live bot (313 vs 245): it never sleeps, while the live tab scanned irregularly.
On Jul 27-29 the bot really traded: 245 live SPX/NDX entries. My simulation, run blind on the same days, produced 313 entries and matched 116 of them on exact (contract, 30-min block). It covered 86% of the bot's SPX entries but only 24% of NDX ones — same-side, near-strike differences driven by 1-minute quotes and (on Jul 27) a denser decision grid than the bot's actual ~2-minute live scan rate. The simulation trades MORE than the live bot (313 vs 245): it never sleeps, while the live tab scanned irregularly.
1 · Equity curve — the full simulated strategy
2 · Exit policies on the same simulated entries (worst-case fills)
| Exit policy | Net P/L (ALL) | SPX | NDX | Win rate | Profit factor | Max DD $ | Avg hold |
|---|---|---|---|---|---|---|---|
| time_10m | +3,020 | +37,624 | -34,604 | 46% | 1.01 | 77,268 | 10m |
| fixed_t30_s30 | -2,020 | -28,256 | +26,236 | 46% | 1.00 | 108,077 | 22m |
| baseline_full | -10,075 | +42,589 | -52,664 | 51% | 0.96 | 46,074 | 4m |
| fixed_t30_s40 | -13,030 | -45,246 | +32,216 | 50% | 0.98 | 123,087 | 28m |
| fixed_t20_s40 | -15,170 | -29,596 | +14,426 | 58% | 0.97 | 124,316 | 22m |
| time_15m | -18,065 | +41,704 | -59,769 | 45% | 0.97 | 106,997 | 15m |
| pl_150_80_120 | -43,640 | -46,621 | +2,981 | 67% | 0.91 | 144,693 | 41m |
| time_30m | -98,495 | +2,349 | -100,844 | 40% | 0.87 | 188,294 | 30m |
| hold_to_end | -1,474,965 | -411,716 | -1,063,249 | 24% | 0.27 | 1,523,030 | 191m |
baseline_full = the bot's actual rule set: $150/80%/$120 trailing profit lock on every minute mark + exit after 2 consecutive same-key adverse GEX/permission/zero-gamma/VWAP decision readings + labeled 15:55 ET force close (the live bot has NO forced close — audit finding). pl_150_80_120 = profit lock alone (no GEX exits): it holds losers to the close and does far worse. Exit reasons in baseline_full: gex_exit:permission 716, profit_lock 321, gex_exit:bias 38, gex_exit:zero-gamma 19, gex_exit:vwap 19, eod_force_1555 1.
3 · Day by day (baseline_full, worst-case)
| Day | SPX trades | SPX P/L | NDX trades | NDX P/L | Total | Decision data |
|---|---|---|---|---|---|---|
| 2026-07-16 | 44 | +4,182 | 0 | — | +4,182 | blob rebuild (15 s grid) |
| 2026-07-17 | 92 | +7,009 | 38 | -5,120 | +1,888 | blob rebuild (15 s grid) |
| 2026-07-20 | 96 | +2,983 | 32 | -6,282 | -3,299 | blob rebuild (15 s grid) |
| 2026-07-21 | 85 | +3,803 | 43 | -16,327 | -12,524 | blob rebuild (15 s grid) |
| 2026-07-22 | 78 | -153 | 42 | -20,885 | -21,038 | blob rebuild (15 s grid) |
| 2026-07-23 | 84 | +4,339 | 42 | +2,065 | +6,404 | blob rebuild (15 s grid) |
| 2026-07-24 | 80 | +9,504 | 45 | -9,469 | +35 | blob rebuild (15 s grid) |
| 2026-07-27 | 89 | +8,828 | 38 | -6,480 | +2,347 | live+blob (15 s grid) |
| 2026-07-28 | 39 | -506 | 49 | +9,185 | +8,679 | live scan cadence (~1-2 min) |
| 2026-07-29 | 45 | +2,601 | 53 | +650 | +3,251 | live scan cadence (~1-2 min) |
4 · What was simulated, exactly
- Decisions: 23,681 decision timestamps across 10 days x 2 tickers — the recorded GEX snapshot cadence, downsampled to the 15 s Auto-Live minimum. All 19 score components recomputed per timestamp (GEX majors/zero-gamma/max-change from production snapshots; VWAP/SMA50/SMA200/Fib from the same Alpaca IEX SPY/QQQ bars the live bot uses — including its daily-bar lookahead of the running session close, reproduced deliberately and verified against recorded SMA values; vanna/charm flow from stored profiles; VIX from stored VIX snapshots where present, else the recorded-score residual inference described in the caveats).
- Permission gate: blockers exactly as coded, incl. "too close": |spot − level| ≤ max($3.00, 0.05% of spot) against classic major +/− GEX (volume).
- Contract selection: same-day-expiry (0DTE ENFORCED — the live code does not enforce it; labeled deviation), strike band [96%, 104%] of spot-to-target, sorted by distance-to-GEX-target, first 20 priced from real OPRA cbbo-1m NBBO (staleness ≤120 s, crossed/zero/sentinel quotes rejected), Black-Scholes greeks from the mid at r=4.5% exactly like the bot's databento path, and the bot's 12-step score. OI is absent on the databento path (bonus never applies) — live parity.
- Entries: permission = "possible trade after confirmation", Pick #1 only, score ≥100, bullish→calls / bearish→puts, one entry per contract per 30-min block, ≤10 entries per block (shared across tickers; the cap never bound), entry at the real ask (worst) / mid (optimistic).
- Exits: replayed on real per-minute NBBO; every worst-case sell uses the actual bid of that minute, never a bar high.
Caveats — read before believing.
- 1-minute resolution: the bot lives on 1 s quotes; this simulation prices decisions and exits on 1-minute NBBO. Sub-minute spikes are invisible: profit locks/stops trigger on minute marks. Direction of bias is mixed — peaks are understated (fewer lock arms) but so are intraminute stop-outs.
- 10 days, one regime: Jul 16-29, 2026 was a drifting-then-selling market with VIX 15-21. Nothing here generalizes beyond that regime; a 60/20/20 walk-forward is impossible on 10 days.
- Entries are simulated, not recorded: on Jul 16-24 the bot's own live scans are unrecoverable (snapshots overwritten by vendor day-blobs; recorded decisions used an older rubric), so the entry stream is my reconstruction of what THIS rubric would have done on that data at 15 s cadence — 801 of the 1114 trades. The three live days validate the machinery (Validation A/B above).
- VIX component: no licensed VIX feed. Jul 14-16 have recorded GEXBot VIX (15.7-16.8). Jul 27-29 use a per-timestamp adjustment inferred from the bot's own recorded scores (disclosed, point-in-time). Jul 17-24 assume adjustment 0 (VIX 15-20) — if VIX actually crossed 20 on those days, scores there are up to 0.5 too high, which can only REDUCE the number of qualifying entries.
- Labeled deviations from live code: 0DTE-only universe enforced; 15:55 ET force close added; GEX-confirmation exits evaluated at decision cadence (≥15 s) not the browser's irregular refresh; decision cadence on blob days (15 s) is denser than the bot's actual erratic scan rate — more entries than the live bot would have taken, each still individually rule-faithful.
- Sequencing: block caps/dedup applied per ticker sequentially (SPX before NDX within a day); the 10-per-block cap never triggered, so ordering had no effect.
- Fees $0.66/contract/side; no assignment/exercise modeling (positions closed by 15:55 ET).
Code:
backtests/src/entry_sim.py ·
results: backtests/results/entrysim_*.csv|parquet ·
strategy audit: backtests/STRATEGY_AUDIT.md ·
exit study on recorded entries: index.html ·
6-month long-history test: longhist.html