Overfitting
AdvancedOverfitting is when a statistical or machine learning model fits its training data too closely, capturing random noise rather than the underlying pattern, so that it performs well on past data but poorly on new data.
A model that is too complex for the amount of data available can effectively memorise the past, including random fluctuations that will not repeat. It looks impressive on the data used to build it but fails to generalise.
In finance, overfitting is a common problem in developing trading strategies. With enough parameters and enough trial and error, it is possible to find a strategy that would have made large profits in the past purely by chance.
For example, a trading rule that buys shares only on Tuesdays when the temperature in a particular city is above 20 degrees might show excellent historical returns, but there is no reason to expect it to work in future.
Ways to reduce overfitting include keeping models simple, using more data, testing models on data not used to build them, a practice called out-of-sample testing, and techniques such as cross-validation and regularisation, which penalise unnecessary complexity. A strategy that works only with very specific settings should be treated with suspicion.
For worked examples and full lessons, visit notesbysara.org.