Overfitting occurs when a model is calibrated so closely to historical data that it captures the specific random fluctuations of that period rather than the underlying patterns that would persist in new data. The more parameters a model has relative to the amount of data, the higher the overfitting risk. If you test 1,000 variations of a strategy (different parameter combinations, different signals, different exit rules) and select the best performer, the winner is almost certainly the one that got lucky on that historical sample — not the one with the most robust underlying logic.

The multiple comparison problem is the mathematical driver of overfitting: if you test 100 strategies with no predictive power, you expect about 5 to show statistically significant positive returns at the 5% significance threshold by pure chance. If you then pick the best of those 100 strategies, you are selecting the luckiest random performer, not a genuine winner. Harvey, Liu, and Zhu estimated that most published factor research fails to account for the multiple testing problem — a substantial fraction of published 'factors' may be overfitted to the specific historical samples studied.

Lookahead bias occurs when a backtest inadvertently uses information that would not have been available at the time of the simulated trade. Common examples: using end-of-year financial statement data to make trading decisions dated to February, when the statements weren't publicly filed until March (the strategy would have been impossible to execute); using a stock's eventual bankruptcy in 2010 to exclude it from a 2005 portfolio (survivorship bias, a form of lookahead); using daily closing prices that are only available after the market closes to generate intraday signals.

Point-in-time data eliminates lookahead bias for fundamental strategies. Point-in-time databases record exactly what information was publicly available as of each historical date — including revision history of economic data (GDP, employment figures are revised multiple times), the filing dates of earnings announcements (not the reported quarter), and the analyst consensus estimates available before (not after) each announcement. Using these datasets substantially reduces backtest overstatement compared to using as-reported current data.

The primary prevention for overfitting is strict separation between optimization data and evaluation data. Optimize strategy parameters on an early historical window, then evaluate on a later window that was never used in development. Any strategy improvement made after seeing the evaluation data invalidates the separation — the 'test set' becomes part of the training process. Walk-forward testing (repeatedly training on expanding windows and testing on subsequent periods) is the most rigorous implementation of this separation.

Parsimony — using the simplest model that explains the data — is the practitioner's discipline against overfitting. Occam's razor in quantitative finance: a strategy with 2 parameters that generates a Sharpe of 0.8 is more credible than a strategy with 15 parameters that generates a Sharpe of 1.2 on the same data. Additional parameters should only be added when they are theoretically motivated and demonstrate consistent improvement across multiple independent out-of-sample periods, not just when they happen to improve the in-sample fit.