E-10. Cascade proxy — rejected (2026-07-13)
- 1m data over 4 years, 3 conditions (volume spike + sharp drop + recovery wick) → long the rebound. Grid of 12 combinations.
- Result: all 12 combinations had negative Sharpe in train (-0.72 ~ -2.75). Val was also negative throughout. Clearly rejected.
- Analysis: not overfitting but a dead hypothesis. By the time the "recovery wick" condition holds, much of the rebound is already spent — the residual return from entering on the next candle < the 0.2% round-trip cost. The ordering across combinations is also consistent (looser conditions are worse = more noise trading) — a cost problem, not a signal.
- Follow-up hypothesis (untested): the real edge is in collecting limit fills during the crash — not testable with candle backtests; real liquidation data are needed. Can be revisited once the recorder (record_liquidations.py) has accumulated data.
Source and migration evidence
Source: backtest/docs/experiment-log.md#E-10
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.