Model Evaluation Stores (MES) are not supported for time series forecasting models.
The input dataset for the evaluation recipe should contain at least:
The time column
The time series identifiers columns (if any)
The target column
The external features columns (if any)
The evaluation recipe computes the evaluation dataset by moving the forecast/evaluation window (of size forecast horizon) from the end of the input dataset to the beginning as many times as possible (given the size of the timeseries), or a fixed number of times if the Max. nb. forecast horizons is set.
The output metrics dataset contains the computed metrics. If the input dataset contains multiple time series, choose between aggregated metrics (one single row, default) or per time series metrics (one row per series).
Statistical models (ARIMA and Seasonal LOESS) can be refit on the input data before evaluation.
Evaluation with Seasonal LOESS only works with refitting enabled.