Genetic algorithm strategy evolution
Population of factor combinations evolves across generations against walk-forward validated holdouts. The GA mutates and crosses factors to find non-obvious combinations.
All use cases
Multi-factor screensLive in portal
Stock Screening
GA-evolved screening strategies for US equities — smart-money flow, earnings catalysts, multi-factor combos with walk-forward validation.
Most stock screens are built backwards: someone picks factors that explain past returns, optimizes them on the last 5 years, and then sells it as a forward-looking edge. Our screener uses a genetic algorithm to evolve factor combinations against walk-forward validated holdouts — so the strategies you see are the ones that survived being tested on data they never saw during training. Out-of-sample Sharpe is reported per strategy, and overfit strategies are visibly flagged, not hidden.
What makes it different: Out-of-sample Sharpe ratios reported per strategy. Strategies that overfit on training data are visibly flagged, not hidden.
What it does
Capabilities
Smart-money flow + earnings catalysts
Pre-built factors include 13F position changes (smart money), earnings revisions, surprise scores, and analyst movement — composable in the GA search space.
Walk-forward validation
Every strategy is scored on data the GA never saw. Out-of-sample Sharpe > 1 is good; > 2 is exceptional. Strategies with high IS Sharpe but low OOS Sharpe are flagged as overfit.
Target leakage protections
Multiple safeguards against the GA accidentally discovering forward-looking features (e.g. earnings revisions that 'predict' the print they're part of). Documented C1/C2/C3/C6 fixes.
How it works
Input → output
Input
Universe: Russell 1000 · Window: 2018-2024 · Holdout: 2024-2026
- 1GA initializes a population of 100 factor combinations across the available factor space
- 2Each generation: score on training window (2018-2024), select top 20 by Sharpe, mutate + cross to repopulate
- 3After N generations, the surviving strategies are scored on the holdout window (2024-2026)
- 4Strategies ranked by OOS Sharpe — overfit ones (high IS, low OOS) flagged with a warning badge
- 5User can clone any strategy, edit the factors, re-run with their own universe + window
Output
Ranked table: strategy ID, factors, IS Sharpe, OOS Sharpe, overfit flag, top-10 current picks. Click any strategy for a forward-validated backtest report.
Output preview
portal · live
Russell 1000
Universe · holdout: 2024-2026
Top 4 GA-evolved strategies
| Strategy | IS Sharpe | OOS Sharpe | Flag |
|---|---|---|---|
| ev-001 | 2.4 | 2.1 | strong |
| ev-014 | 1.9 | 1.6 | strong |
| ev-022 | 2.8 | 0.4 | overfit |
| ev-031 | 1.3 | 1.2 | good |
Overfit flagged, not hidden.Strategy ev-022 has high IS Sharpe but low OOS → don't trust it.
Source: Walk-forward 2018-2024 → 2024-2026
Why this isn't a wrapper
Why most screens are subtly broken
The vast majority of public stock screens are built on factors selected because they explained past returns — and that's exactly the problem.
- GA discovers factor combinations a human wouldn't think to try — but only the ones that survive holdout testing are surfaced.
- Target leakage is the silent killer of backtests. We have explicit C1/C2/C3/C6 fixes documented; most screeners have none.
- We report OOS Sharpe and IS Sharpe separately so you can see whether a strategy is genuinely forward-looking or just curve-fit.
- Strategies flagged as overfit aren't hidden — they're shown with a warning, so you can see what NOT to trust.
Want to try it?
Book a 30-minute walk-through. We'll demo a real run on a ticker you pick.