Time Series Forecasting using external features

AVerhagen Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered Posts: 6


I'm having trouble getting time series forecasts with DeepAR and Auto ARIMA models when using an external feature. I have provided future values of the external feature, however the model only gives backtests instead of future predictions.

In my case I have a F1 fan site where I can see the amount of people that browse on it. As an external feature I want to give as input the days that there will be a raceday and historical data that shows on which days there has been a raceday in the past.

The historical data spans 2 years, and I have provided 30 days of future values with a forecast horizon of 8 days. Why don't I see a forecast and only a backtest?

Thanks in advance!

Best Answer

  • StanG
    StanG Dataiker, Registered Posts: 52 Dataiker
    Answer ✓


    During training, models train on all the dataset provided and only make backtest forecasts (and not future forecasts), so you shouldn't provide future values of external features in the training dataset.

    You need to use the scoring recipe after deploying the trained model on the dataset that contains the future values of external features in order to get the future forecasts.

Setup Info
      Help me…