Do not begin by deleting the losses
The fastest way to overfit a strategy is to discover a losing cluster and immediately create a filter that removes it. First determine whether the same condition also produces important winners and whether the relationship persists outside the development sample.
Classify the losses
- Immediate failure: little or no favorable excursion after entry.
- Adverse-then-recovery: meaningful MAE before the trade improves.
- Opportunity giveback: substantial MFE that ultimately becomes a loss or weak exit.
- Session cluster: repeated losses during a specific time window.
- Regime cluster: losses concentrated in a measurable market state.
Compare against winners
A feature is not useful merely because it appears in losing trades. Compare its frequency and distribution in winners. The question is whether it meaningfully changes the probability or expected value of an outcome.
Separate entry diagnosis from exit diagnosis
If losses show almost no MFE, the entry model deserves scrutiny. If trades regularly travel well into profit before failing, the entry may be acceptable and the exit or risk process may be the more useful research target.
Turn patterns into controlled tests
Once a repeatable pattern is identified, formulate a narrow hypothesis and test it without rewriting the entire strategy. That preserves causal clarity and reduces the temptation to optimize every historical weakness.
Ask better questions about your strategy.
EdgeQuery is being built to analyze automated trading-system history beyond headline statistics. The founding beta focuses on trade-level behavior, market context, failure patterns and research priorities.