Timeseries forecasting

While direct timeseries prediction is a work in progress Ludwig can ingest timeseries input feature data and make numerical predictions. Below is an example of a model trained to forecast timeseries at five different horizons.

timeseries_data y1 y2 y3 y4 y5
15.07 14.89 14.45 ... 16.92 16.67 16.48 17.00 17.02
14.89 14.45 14.30 ... 16.67 16.48 17.00 17.02 16.48
14.45 14.3 14.94 ... 16.48 17.00 17.02 16.48 15.82
ludwig experiment \
--dataset timeseries_data.csv \
  --config_file config.yaml

With config.yaml:

input_features:
    -
        name: timeseries_data
        type: timeseries

output_features:
    -
        name: y1
        type: numerical
    -
        name: y2
        type: numerical
    -
        name: y3
        type: numerical
    -
        name: y4
        type: numerical
    -
        name: y5
        type: numerical