quantam

The personal website of Edward Yu.

Preventing Overfitting in Quantitative Finance

Recently, I gave an informal talk at Columbia University about how to prevent overfitting of algorithmic trading models. Personally, I think there is a general lack of statistical rigor in finance, with many traders not following recommended best practices in data science.

You can find the slides of my talk here.

TL;DR

  1. Obtain good data that is of high granularity, up to date, and doesn’t contain survivorship bias.
  2. Make the problem easier by transforming inputs and outputs.
  3. Apply regularization to penalize the complexity of the model.
  4. Split data into test/training/validation sets for backtesting.
  5. Use walk-forward optimization and don’t only optimize for portfolio returns; I recommend optimizing for Sharpe ratio.
  6. Don’t ever assume financial data is generated by a Gaussian distribution.