E-18. [gpt 5.6 sol] Strategy 8 — 5m regime RSI/BB with a real limit-order path (2026-07-15)
- Preregistered D-022: 96 combinations, EMA/ATR up/down/sideways regimes, RSI/BB signals, 20-minute TTL, maker entry/TP + taker SL/time, same-bar SL first, 1bp trade-through stress.
- train: 253,525 bars / 880 days. Top combinations filled 1.9~2.6 times per day, fill 84~91%. The desired activity was met, but 0 were Stage 1 eligible.
- Best: slow144 / RSI30 / BB2.5 / offset0.25ATR / stop1.5ATR / 3R. 1,880 orders, 1,695 trades, fill 90.2%, Sharpe -9.99, PF 0.31, -74.04%.
- Failure anatomy: stop 558 + entry-bar stop 534 / time 521 / take 82. Even at 0 cost: -43.69%, Sharpe -4.31, PF 0.62. With 1bp trade-through: -74.74%.
- Fills by regime: sideways 999, up 408, down 288. The regime filter limited direction but could not tell whether extreme prices would persist.
- Inference: fill selectivity matters more than limit-order cost savings. A limit placed under a falling knife fills with high probability only when bad information arrives, not at a good price.
- Verdict D-023: class rejected. The next separate hypothesis should be re-entry on a pullback after a confirmed recovery, not an order placed immediately on touch.
Source and migration evidence
Source: backtest/docs/experiment-log.md#E-18
Original SHA-256: 069c4aaaf53a20930ca8d94ba6365b4ce2b6bc560171d662a4c179ea55a6857c
Raw data and account information are not part of the public package. The original results do not replace statistical or scientific revalidation. The English text is a translation; numbers, tables and charts are carried over unchanged.