The bot generated solid signal volume (+419% funnel ingest WoW) but suffered severe conversion collapse: only 5% of actionable signals became open positions, and only 34% of those closed profitably. At $2.65 per win with 43% ghost position accuracy, the system is burning capital on low-conviction trades and poor exit discipline.
The funnel leaks catastrophically at actionable_pass → position_open (130 → 97, 75% pass-through). Of 2,614 fanout signals, only 130 (5%) reached actionable threshold, indicating either overly permissive scoring thresholds upstream or a disconnect between scoring logic and real execution risk. The 44 graded positions represent only 45% of closed positions, suggesting incomplete post-trade analysis feedback.
At $0.601 per position opened and $2.65 per win, the cost-per-win is prohibitive. 180 closed positions with 44 graded implies 136 positions lack post-trade instrumentation. The 43% ghost win rate (0.2% avg) vs. actual closed position performance gap suggests the bot is opening lower-quality positions than its scoring model predicts. AI spending ($43.10/3,075 calls = $0.014/call) is efficient, but applied to weak signals.
- News-driven signals (56% of winners, n=57) show the strongest conviction—prioritize news ingest quality and speed.
- Trailing stop exits on extended moves (46 trades, 0% slippage) executed perfectly—this is the cleanest exit pattern.
- Signal volume growth (+419% WoW ingest) proves pipeline capacity is not the constraint.
- Conviction_drop exits (n=632, +1.0% slip) dominate but indicate premature position sizing—bot is over-committing to weak signals.
- Stale signal exits (n=174, +1.9% slip) and conflicting_signal exits (n=35, +3.2% slip) reveal poor signal decay modeling and real-time contradiction detection.
- Ghost positions show 43% win rate vs. <34% actual closes—bot is systematically opening positions weaker than its scoring predicts.
217 ghost positions at 43% win rate (+0.2% avg) suggest the scoring model correctly identifies winning setups ~40% of the time, but actual opened positions convert at only ~22% (44 wins from 180 closed). This 21-point gap indicates systematic position-sizing bias toward weaker signals or execution timing failures. The +0.2% average gain on ghosts is also anemic—targets are likely too tight or conviction thresholds too low.
- Raise actionable_pass threshold by 2-3 sigma: currently 5% of fanout signals pass; reduce to top 1-2% to eliminate low-conviction noise. Ghost data proves 43% of medium-conviction trades are losers.
- Implement real-time signal decay: Stale exits (+1.9% slip, n=174) and conflicting_signal exits (+3.2% slip, n=35) indicate poor time-decay and contradiction detection. Add intraday signal freshness scoring and conflict veto logic.
- Audit position-sizing logic: 180 closed vs. 44 graded is a red flag. Enforce post-trade grading on 100% of closes, and correlate position size with ghost vs. actual P&L to identify which signal classes are being oversized relative to conviction.