Currently, TimeSeriesPredictor deals with missing values in the data by first imputing them via forward-/backward-filling, and then training all models as if there are no missing values. This strategy may lead to poor accuracy on datasets with a large portion of missing values since training data will include regions with constant values arising from ffilling.

A better alternative is to keep the missing values represented by nan in the data, and let models handle it. This requires following modifications to the code:

• Update metric implementations to handle missing values in target ([timeseries] Ensure that all metrics handle missing values in the target #3966)
• Update preprocessing logic in TimeSeriesPredictor._check_and_prepare_data_frame
• Make sure that all models can handle missing values. This means, all models can train normally and produce forecasts with no NaN values, even if training data contains NaNs.
  • GluonTS models (DeepAR, TFT, PatchTST, DLinear)
  • StatsForecast models (AutoETS, AutoARIMA, Theta, intermittent demand models)
  • Baseline models (Naive, SeasonalNaive, Average, SeasonalAverage, Zero)
  • MLForecast models (DirectTabular, RecursiveTabular)