Overfitting
Tuning a strategy so closely to past data that it captures noise instead of a repeatable pattern — great in the backtest, poor in live trading.
Overfitting happens when a strategy has enough adjustable parts — parameters, filters, exceptions — to explain the random quirks of one historical sample. Each extra rule improves the backtest a little; together they describe the past perfectly and the future not at all.
Warning signs: many parameters relative to the number of trades, results that collapse when a parameter moves one step (a 14-period lookback works, 13 and 15 do not), and rules added to remove specific losing trades. The more variations tested, the more likely the best one won by luck.
The defences are simplicity and unseen data. Keep the rule count low, prefer parameters that work across a broad range, hold back an out-of-sample period you never optimise on, and judge the strategy on that period alone.
On the desk
A strategy with 9 tuned parameters shows a Sharpe ratio of 2.8 on 2015–2021 data. On the untouched 2022–2023 data it posts 0.2. A 2-parameter version scored only 1.1 in-sample but held 0.9 out-of-sample — the simpler rules captured something real.