walkforward

Walkforward is a methodology used in time series modeling and trading to evaluate predictive models and strategies by advancing a time window through historical data. In finance, walk-forward optimization (often called walk-forward testing) involves repeatedly splitting data into an in-sample window used for calibration and an out-of-sample window used for evaluation, then shifting the window forward and re-optimizing. This approach aims to approximate how a strategy would perform when deployed in real time, by testing on data that was not available during the calibration stage.

Process and variants: At the start, one selects lengths for the in-sample and out-of-sample windows and a step size. For each step, parameters or rules are optimized using only the in-sample data, and the resulting strategy is then applied to the subsequent out-of-sample data to generate performance metrics. The step is repeated as time progresses. A common variant is walk-forward optimization, where parameters are re-optimized at every step. A stricter form uses a fixed parameter set for all out-of-sample windows to assess stability, though this is less typical in finance.

Purpose and considerations: Walkforward seeks to mitigate overfitting by testing on unseen data and by simulating live deployment. It provides insight into robustness across changing market conditions and different regimes. Limitations include sensitivity to window choices, potential data-snooping from repeated optimizations, computational intensity, and remaining vulnerability to non-stationarity. Proper implementation requires strict separation of in-sample and out-of-sample data to avoid look-ahead bias.

Related concepts include rolling window, rolling-origin forecast, backtesting, and time-series cross-validation.