This isn't a machine trying to guess where the market goes next. It's a disciplined insurance desk: it sells the promise that a price floor won't break, collects the premium, and lets time do the work. Here is the exact chain of reasoning behind every trade — replayed against real ones.
No trading background needed. Everything after this point is built from these.
Open to close, every 15 seconds, the same fixed checks run. The overwhelming majority end in "nothing to do." Restraint is the first feature, not an idle system.
Only two underlyings are ever eligible — SPY and QQQ — because their options markets stay liquid even 13 days out and far from the money.
SPY's own price action sorts the moment into one of three regimes, recomputed every 30 minutes so noise can't jerk it around.
Replayed from a real cycle: SPY, August 4, 10:00 ET. One failure anywhere below and there is no trade.
Now — and only now — the bot opens the live put chain, 7 to 14 days out, hunting the strike closest to 0.30 delta.
Both legs leave together as a single multi-leg limit order at the estimated midpoint — never two market orders, which donate money to the bid-ask on every fill. If it doesn't fill, the bot offers 5¢ worse once, then waits. It doesn't chase.
Every 15 seconds the open spread is re-tested against all five, top to bottom, until one fires.
| Spread | Opened | Closed | Credit | Debit | What ended it | Result |
|---|---|---|---|---|---|---|
| QQQ 681/676 | Aug 3 | Aug 4 | $1.05 | $0.34 | 60% captured | +$71 |
| SPY 746/741 | Aug 3 | Aug 4 | $0.86 | $0.41 | 60% captured | +$45 |
| SPY 731/726 | Jul 31 | Aug 3 | $0.97 | $0.42 | 60% captured | +$55 |
| QQQ 679/674 | Jul 23 | Jul 27 | $1.12 | $2.12 | Broke the short strike | −$100 |
| SPY 728/723 | Jul 23 | Jul 30 | $0.97 | $0.45 | 60% captured | +$52 |
| SPY 736/731 | Jul 21 | Jul 23 | $0.92 | $1.82 | Broke the short strike | −$90 |
Every win ended the same way — bought back around 35–60¢ on a roughly dollar credit, well before expiration. Every loss ended the same way too: not a slow bleed, but the underlying actually trading through the floor being defended.
64 symbols, every 15 seconds, a full page of live technical analysis on each — trend crossover, RSI, VWAP position, volume confirmation, opening-range bias, sector-relative strength. All of it real, all of it published continuously.
But none of these 64 names can currently generate a trade. The scanner was built for two strategies that have since been switched off, and the wiring from "NVDA looks bullish" to an actual order is disconnected.
One wire is still live: SPY and QQQ's own VWAP reading feeds gate 08 of the spread funnel. That's the only room in the building with its hand on a trigger.
A trained pair of gradient-boosted (XGBoost) models — one scoring near-term up-moves, one down — built on this bot's own history from about a dozen engineered features: RSI, moving-average spread, VWAP distance, volume ratio, momentum over several windows, the day's return, encoded time-of-day and day-of-week, and VIX. A newer version adds pattern features: consecutive same-direction candles, range compression.
A working SHAP layer can decompose any single prediction, feature by feature. It's loaded, functioning, and ready — and it has no live trades to inform. It runs offline overnight, backtesting candidate versions of itself against each other.
Illustrative — research, not why this tradedNot everything this system knows how to do is something it's currently doing. Both lists are real code.
A system that revises itself, rather than one running on rails.
Eight gates, five ways out, two eligible underlyings, a $5-wide bet with its worst case known in advance. No real money — every trade runs on a paper account, fully logged and replayable. This same reasoning also runs a focused, SPY-only version elsewhere on the site.