Overfitting is one of the central risks in quantitative research. It happens when a strategy is tuned so closely to historical data that it captures noise rather than a repeatable market behavior. The result is a backtest that looks impressive and a live strategy that disappoints.
Overfitting can be obvious, such as a strategy with many parameters selected to maximize past return. It can also be subtle. Trying many ideas and only publishing the winner is a form of selection bias. Changing the universe, date range, rebalance frequency, and filters until the result looks good can create the same problem.
Warning signs include extreme sensitivity to small parameter changes, performance concentrated in one short period, very high turnover with small assumed costs, and rules with no economic rationale. A model that requires dozens of fragile conditions may be explaining the past rather than discovering structure.
Reducing overfitting starts with simplicity. Fewer parameters are easier to understand and harder to abuse. Out-of-sample testing helps, but only if the out-of-sample period is not repeatedly used for tuning. Walk-forward validation, cross-market tests, and cost stress tests can provide additional evidence.
Researchers should record failed tests. A clean research log makes it harder to forget how many attempts were made before the attractive result appeared. The more trials conducted, the more skeptical one should be of the best result.
No method eliminates overfitting completely. Markets change, and evidence is always incomplete. The goal is to build a process that rewards robust ideas and makes fragile ideas uncomfortable to defend.
In live trading, humility is a risk control. A strategy that knows it might be wrong is designed differently from a strategy that mistakes a perfect backtest for truth.