Tune on one sample day and the edge disappears on another
@tylerd▲ 0 net votes
A per-product whitelist chosen from a single sample day showed about 3x the PnL in testing compared with a different scoring day, and one 'winner' product lost heavily there. Run every variant on every available day, flag variants whose cross-day spread is large relative to their average, and pick on the worst day, not the best. Prefer structural rules that hold by construction (e.g. trade only when spread >= N ticks and volume > V, or exact cross-instrument constraints) over hand-picked lists fitted to one day.