Processing

Overfitting-fix experiment batch (see NEXT_STEPS.md) -- one row per experiment, updated live as each one trains and is evaluated on the holdout.

GA Parameter-Search Experiments

Mean-reversion strategy-parameter GAs (BTC-MeanReversion-Trader) run directly on this server -- a different project and schema from the batch above (no crash/bull/flat regime split; each row is one resolution/strategy config, trained on its own set of blocks with a held-out one). Not the turtle-trading system -- those GA experiments were run locally and are already fully written up on Turtle Report; this table is reserved for runs that actually execute on this server, currently mean-reversion only. All runs use a 2× leverage cap (Plus500's retail crypto CFD tier) and fitness is computed from realized-only P&L -- a position still open at a block's end doesn't count until it actually closes (see caveat column). Two further fixes, added after a random-genome control found most evolved genomes were statistically indistinguishable from -- or worse than -- a random draw from the same search space: fitness is now margin-over-random (every genome is scored against a fixed baseline of 20 random genomes on the same blocks, not in absolute terms -- the "vs random genomes" row below), and any block with fewer than 3 trades gets a fixed penalty instead of its raw excess return, closing the loophole where a genome could post a good score just by sitting out a falling market.

GA Momentum, Full-History Experiments

BTC-Momentum-Trader -- structurally different from the mean-reversion GA above: each genome trades the entire historical price series in one continuous pass (not isolated blocks) before being scored, closer to how the main live population works. Population is culled 50% every generation (elitism 4) and refilled by breeding survivors. Genuine wipe-out: no floor stops a trader from sizing new positions once equity is low -- a trader whose mark-to-market equity crosses -100% of starting equity (matching the main project's own WIPE_BELOW) is permanently dead for the rest of that run and can never breed again. Three experiments, one per hold-time ceiling, with candle resolution and 200-bar lookback context scaled to match (5min hold → 1-minute candles, 4h hold → 5-minute candles, 8h hold → 15-minute candles). Checkpointed to disk after every generation, so progress below is live, not just a final result.