Walk Forward Validation
Ordinary cross validation shuffles data, which is invalid for time series because it lets a model train on the future. Walk forward validation keeps time order intact.
The procedure
- Train on an initial block of early data.
- Forecast the next slice and score it against the truth.
- Extend the training window to include that slice, then forecast the next one.
- Repeat marching forward through the series.
This mimics how a model is used in production, where you always predict the unseen next period from the known past.
Expanding and sliding
- An expanding window grows the training set at each step, using all history.
- A sliding window keeps a fixed length, dropping the oldest data so the model focuses on recent behavior.
Why it matters
Walk forward validation gives an honest estimate of forecast error and exposes how performance changes over time. It costs more compute because the model is refit at each step, but it prevents the optimistic bias of shuffled splits.
Key idea
Walk forward validation trains on the past and tests on the next slice repeatedly, giving an honest time aware estimate of forecast error.