0DTE Bot — Did the exit rule make money?
Exit-strategy study on the bot's own recorded trades · worst-case fills headline ·
generated 2026-07-30
New: the full strategy — entries included — has now been simulated from raw
GEX + OPRA data over 10 trading days (this page only re-plays exits on recorded entries).
Full entry simulation →
Newer: did July's candidate winners survive SIX MONTHS of history (Jan 2 – Jul 13, 131 days, Tier B mechanical entries)? Short answer: no. 6-month long-history test →
Newer: did July's candidate winners survive SIX MONTHS of history (Jan 2 – Jul 13, 131 days, Tier B mechanical entries)? Short answer: no. 6-month long-history test →
Is there any hope? — the bottom line across all three studies
One real hope, surrounded by dead ends — and it's testable.
Dead (no ambiguity left):
- NDX — killed by chronic $1.30–$3.50 spreads, in every test, every month of 2026. Not fixable by better exits. (6-month test, entry sim, exit study all agree.)
- The strategy without its GEX brain — six months of history says the mechanical parts (timing, contract quality, exits) have zero edge on their own.
- The current profit-lock exit — sells winners so early it captures essentially none of their peak potential.
Alive:
- The full strategy, on SPX only: the one test that included the bot's actual GEX intelligence showed +$58/trade across 732 trades under worst-case fills (entry simulation) — the only green number an honest test produced in this whole project. SPX's penny-wide spreads are the reason it is even possible.
Why this is not yet proof: it is 10 days, in one calm market regime. Plenty of losing strategies look great for 10 days. The only part of the strategy showing promise is the part history cannot verify — point-in-time GEX data for the past does not exist to buy.
The plan that costs nothing: keep the bot recording on paper, SPX-only, and let the faithful window grow — every trading day adds evidence for free. At ~3 months of recordings, re-run the identical entry simulation (one command). If +$58/trade survives 60 days the way it survived 10, that is a real signal. If it evaporates, these backtests saved tens of thousands of real dollars. Until then: no real money, no NDX.
Verdict. Under honest, worst-case fills (buy at the ask, sell at the bid), the bot's
actual exit rule lost −$34,053
across its 251 closed live trades — about −$136 per trade.
The same trades scored +$17,069 in the bot's own optimistic mid-price
accounting: the entire gap is the bid/ask spread, roughly $51,388 of round-trip
spread cost across all trades. The bot's edge, if any, is currently smaller than the spread it
pays — mostly on wide-spread NDX contracts (NDX: −$284/trade; SPX: about breakeven).
Best alternative exit on the exact same entries: simply selling after 15 minutes (+$49,377 worst-case). But read the fine print: that profit comes from 5 huge NDX winners on two trending days — without them it's negative. The solid takeaway is directional, not a new rule: the current profit lock sells winners far too early (it captured -6% of peak potential — i.e. nothing), while every dollar crosses a spread that eats it. No profit-lock setting in the whole grid was profitable under worst-case fills.
Best alternative exit on the exact same entries: simply selling after 15 minutes (+$49,377 worst-case). But read the fine print: that profit comes from 5 huge NDX winners on two trending days — without them it's negative. The solid takeaway is directional, not a new rule: the current profit lock sells winners far too early (it captured -6% of peak potential — i.e. nothing), while every dollar crosses a spread that eats it. No profit-lock setting in the whole grid was profitable under worst-case fills.
Data honesty — read before believing any number above.
- Tiny window: the ledger holds 716 rows, but 451 are simulator ("replay") trades and 14 never closed. The real live, closed sample is 251 trades on just 3 trading days (Jul 27–29, 2026) — 13 more live trades from Jul 15 were left open and are excluded. Three days cannot prove or disprove a strategy; they can only measure exit mechanics.
- Worst-case fill definition: entry at the recorded ask; every sell at an estimated executable bid (mark − half the entry spread); fees $0.66/contract/side. Optimistic numbers (mid-price fills, what the bot's paper ledger itself assumes) are shown greyed for reference.
- Engine validation passed: replaying the bot's own rule over the surviving native
mark sequences reproduced the recorded exits in 55/55
(100%) of validatable trades; the bot's own shadow profit-lock agreed with its real
exits in 59/60 clean cases. Full detail:
backtests/results/validation.csv. - Path caveat: the stored per-trade price paths are a 120-slot ring buffer that keeps
overwriting itself after the trade closes — only 55 trades kept a fully replayable native path
(55 usable_full / 23 usable_partial / 173 unusable). Alternative policies are therefore replayed on the
bot's cached 1-minute bars (full coverage, real timestamps). The bot's own result uses
its recorded 1-second exits — it is never measured at degraded resolution. One-minute replay
understates intra-minute peaks, so tight stops/locks are approximate; time-based exits are robust.
Forensics:
backtests/PATH_FORENSICS.md. - Scope: this is an exit study on recorded entries — it holds the bot's entry decisions fixed and only re-plays exits. It is NOT the full entry-simulation backtest (that needs the 11-day OPRA options-data pull).
1 · The bot's rule vs alternatives (worst-case fills)
| Exit policy | Net P/L | Win rate | Avg/trade | Profit factor | Max drawdown | Capture % | Optimistic ref. |
|---|---|---|---|---|---|---|---|
| Sell after 15m | +$49,377 | 47% | +$197 | 1.32 | −$46,637 | 8.1% | +$100,664 |
| Sell after 30m | +$46,691 | 45% | +$186 | 1.23 | −$65,129 | 7.7% | +$97,288 |
| Sell after 10m | +$35,882 | 46% | +$143 | 1.27 | −$32,956 | 5.9% | +$87,270 |
| Fixed: +30% target / −30% stop | +$25,562 | 51% | +$102 | 1.16 | −$41,044 | 4.2% | +$52,637 |
| Fixed: +30% target / −40% stop | +$19,242 | 55% | +$77 | 1.11 | −$46,209 | 3.2% | +$45,721 |
| Fixed: +30% target / −25% stop | +$15,796 | 47% | +$63 | 1.11 | −$36,303 | 2.6% | +$47,243 |
| Fixed: +30% target / −20% stop | +$13,679 | 45% | +$54 | 1.10 | −$35,276 | 2.3% | +$54,819 |
| Fixed: +20% target / −40% stop | +$8,985 | 62% | +$36 | 1.06 | −$42,629 | 1.5% | +$45,096 |
| Fixed: +20% target / −30% stop | +$8,135 | 58% | +$32 | 1.06 | −$37,850 | 1.4% | +$51,830 |
| Fixed: +50% target / −20% stop | +$4,665 | 35% | +$19 | 1.03 | −$48,250 | 0.8% | +$73,185 |
| Hybrid: −40% stop until lock arms ($150/80%/$120) | +$4,557 | 57% | +$18 | 1.03 | −$37,966 | 0.8% | +$47,300 |
| Fixed: +30% target / −15% stop | +$2,987 | 38% | +$12 | 1.02 | −$37,218 | 0.5% | +$57,870 |
| Fixed: +50% target / −25% stop | +$2,539 | 37% | +$10 | 1.01 | −$49,203 | 0.5% | +$57,820 |
| Sell after 5m | +$1,331 | 43% | +$5 | 1.01 | −$34,416 | 0.3% | +$52,719 |
| Fixed: +20% target / −25% stop | +$52 | 53% | +$0 | 1.00 | −$38,958 | 0.1% | +$46,703 |
| BOT'S ACTUAL RULE (profit lock $150/80%/$120 + GEX exits) | −$34,053 | 53% | −$136 | 0.57 | −$39,577 | -5.5% | +$17,069 |
Top 15 of 318 policies tested (profit-lock grid 252, fixed
target/stop 49, time stops 8, hybrids 7, hold-to-end), plus the bot's actual rule (highlighted).
"Capture %" = share of each trade's best-possible profit actually banked. Full grid:
backtests/results/sweep_results.csv.
2 · Charts
Every cell of the profit-lock grid is negative under worst-case fills — tightening or loosening the current lock does not fix it. (Cells in $000; min-lock fixed at $120; other min-lock values look the same, see CSV.)
3 · Day by day (worst-case)
| Day | Trades | Bot's actual rule | Sell-after-15-min |
|---|---|---|---|
| 2026-07-27 | 70 | −$497 | +$29,936 |
| 2026-07-28 | 78 | −$15,148 | −$3,486 |
| 2026-07-29 | 103 | −$18,408 | +$22,927 |
4 · Caveats & next steps
- 3 days, one regime. Jul 27–29 were strong trend days; time-stops shine in trends and bleed in chop. Keep the bot recording — every extra week makes this study stronger.
- Spread is the enemy #1. Before touching exit logic: avoid wide-spread contracts (NDX spreads averaged ~$2 = $200/trade round trip) or demand a bigger expected move per trade.
- Next milestone: the full entry-simulation backtest over the 11 recorded GEX days needs the OPRA minute-quote pull (Databento key required; ~$0 at current entitlement). That also upgrades this exit study from 1-minute to real NBBO for every trade.
- Not investment advice to switch to time stops — their profit here is 5 outlier trades. The defensible change list: cut spread cost, stop selling winners in the first 2 minutes, and re-test on more days.