Walk-Forward Validation

Standard backtesting trains and evaluates a strategy on the same historical data. This conflates pattern recognition with prediction. A strategy may identify a pattern that existed between 2015 and 2020 but has no basis for persistence.

Walk-forward validation addresses this by splitting the data into sequential train and test windows. The strategy is optimized on the training period, then evaluated on the subsequent test period that was not available during optimization. This process repeats across multiple expanding windows.

We support both expanding windows, where the training set grows over time, and rolling windows, where the training set moves forward with a fixed length. A strategy must demonstrate consistent out-of-sample performance across multiple windows. A single good window is not sufficient.

This methodology is not novel. It is standard practice at institutional quantitative funds. We apply it because strategies that fail walk-forward validation have no place in a portfolio, regardless of their in-sample statistics.