Definition
Walk-forward analysis divides historical data into sequential in-sample (training) and out-of-sample (testing) windows that roll forward through time. The strategy is optimized on each in-sample window and tested on the following out-of-sample period. This simulates real-world conditions where strategies must be periodically recalibrated and provides a more realistic assessment than simple in-sample/out-of-sample splits.
lightbulb Example
Divide 10 years of data into 5 rolling windows: optimize on years 1-4, test year 5; optimize 2-5, test 6; etc. Average out-of-sample performance across all windows gives a more robust estimate than a single backtest.
verified_user Key Points
- More realistic than fixed in-sample/out-of-sample split
- Simulates periodic strategy recalibration
- Average out-of-sample performance is the key metric
- Helps detect overfitting and regime dependence