Rollingorigin

Rolling origin, also called rolling-origin evaluation or walk-forward validation, is a method for evaluating time series forecasting models. In rolling-origin evaluation, a model is trained on historical data up to a forecast origin, and a forecast is produced for a fixed horizon ahead. The origin then advances in time, the training set is updated (by adding new observations and, depending on the approach, discarding oldest data), and the model is re-trained to forecast the next block. This process continues, providing an empirical measure of how well a model performs as data accumulate.

Purpose and advantages: The method preserves temporal order, avoids look-ahead bias, and mimics real-world updating of forecasts. It is particularly useful for comparing models and tuning parameters under realistic data availability. By aggregating errors across multiple origins, it offers a robust estimate of out-of-sample performance.

Procedure: Start with an initial training window ending at time t0. Fit the model and generate forecasts for t0+1 through t0+h. When outcomes are observed, compute error metrics such as MAE, RMSE, or MAPE. Move the origin forward by a chosen step, update the training data, re-train, and forecast again. Repeat until the end of the series. For multi-step horizons, forecasts for all horizon steps are produced at each origin.

Variants and considerations: Some implementations use a fixed forecast origin with a rolling training window, others use a growing training window. The choice affects bias and variance of the error estimates. It can be computationally intensive, and care is needed with non-stationary data or regime changes.

See also: time-series cross-validation, walk-forward validation, forecast origin. Rolling-origin evaluation is widely supported in forecasting tools and libraries as a standard approach for assessing forecast performance.