When forecasting multivariate time series, components don't have necessarily the same scale and unit, therefore basic metrics such as MAE or MSE, widely used in the litterature, are not suitable when they are averaged/aggregated across components. In darts, some metrics can be used for that, such as sMAPE or MASE. However, some of them need specific parameters when computing them, we should therefore change the BenchmarkMetric class and its behavior in the benchmark module.
When forecasting multivariate time series, components don't have necessarily the same scale and unit, therefore basic metrics such as MAE or MSE, widely used in the litterature, are not suitable when they are averaged/aggregated across components. In darts, some metrics can be used for that, such as sMAPE or MASE. However, some of them need specific parameters when computing them, we should therefore change the BenchmarkMetric class and its behavior in the benchmark module.