Skip to content
GetProfitable
Search
Dictionary

Overfitting (curve fitting)

Tuning a strategy so closely to past data that it captures the noise as well as the pattern, and therefore fails on new data.

Any dataset contains accidents. If you search hard enough you will find rules that explain them perfectly, and those rules carry no information about tomorrow. The tell is a backtest that looks excellent and an out-of-sample result that looks like a coin flip.

A simple demonstration: generate 1,000 series of pure random-walk prices and test 100 moving-average rules on each. On the best-performing combination you will see an impressive equity curve with a good sharpe-ratio, drawn entirely from noise, because you took the maximum of 100,000 random numbers.

Defences are all variations on the same idea: fewer parameters, more data, honest counting of how many things you tried, and results that must survive data the fitting never touched. See walk-forward-analysis and purged-cross-validation.

Related: underfitting, out-of-sample, data-snooping, multiple-testing, degrees-of-freedom

Educational only, not advice. Spotted an error? Post in Site Feedback.