Recency bias: judging a system by its last ten trades

[Fairman]

Recency bias: judging a system by its last ten trades

Ten trades is nothing for a statistical judgment, yet it dominates emotional perception. A cluster of losses feels like the system has stopped working. A cluster of wins feels like mastery.

For scalpers who see many trades a day, the recent window shrinks even further. Two hours of poor results can trigger parameter changes, while two hours of good results can trigger overconfidence.

The remedy is a predefined evaluation window. Decide in advance how many trades or how many days are needed before any change is considered. For many systems, that means at least a few hundred trades.

Track confidence intervals. With a win rate near 55 percent, the standard error over 100 trades is about 5 percentage points, so results between 45 and 65 are statistically indistinguishable.

Discipline here protects against two errors: abandoning a good system too early and clinging to a bad one too long.

What sample size must a change in performance reach before any adjustments are made?