Definition
Overfitting
Overfitting, or curve-fitting, occurs when a strategy is tuned so closely to historical data that it captures random noise rather than a genuine pattern, and consequently fails on new data.
Overfitting is the single biggest reason backtested Indian strategies disappoint live. It happens when too many parameters are optimised, too many strategy variants are tried, or the same dataset is reused until something looks good purely by chance, an effect called multiple testing or data snooping.
Defences include keeping models simple, limiting the number of parameters, reserving genuine out-of-sample data, using walk-forward analysis, and stress-testing with Monte Carlo simulation. A useful instinct: if a result seems too good, assume it is overfit until proven otherwise.
Related terms
- BacktestingBacktesting is the process of simulating a trading strategy on historical data to estimate how it would have performed, including returns, drawdowns and risk, before committing real capital.
- Walk-Forward AnalysisWalk-forward analysis is a backtesting technique that repeatedly optimises a strategy on one window of historical data and tests it on the immediately following out-of-sample window, rolling forward through time.
- Out-of-Sample TestingOut-of-sample testing evaluates a strategy on data that was deliberately withheld during model development, providing an unbiased check of whether the discovered edge generalises beyond the fitting period.
- Data Snooping BiasData snooping bias is the statistical distortion that arises when many strategies or parameters are tested on the same dataset, making it likely that some appear profitable purely by chance.
Plain-English explainer from The Dispatch Investors Encyclopedia. General information, not financial advice.