Forecast methods

The first purpose of the forecast module of frePPLe is to compute statistical forecast based on the demand history. The forecast table of FrePPLe contains the intersections for which a forecast should be computed. By default, this table is automatically populated by frePPLe based on the intersections found in the demand history (see parameter forecast.populateForecastTable for more details).

The forecast table contains a field named method where the planner can choose among different methods which one should be picked by frePPLe. FrePPLe implements 5 times series methods to compute the statistical forecast. By default frePPLe automatically selects the method that provides the lowest forecast error.

Check this feature on a live example

  1. Moving average: This method simply calculates the average of the demand history for the last N time buckets where N is a configurable parameter (see parameter forecast.MovingAverage_order).
    In real-life datasets you will normally only see this method being selected when there is only very limited history (eg < 5 time periods).
  2. Constant: The constant method is an implementation of the single exponential smoothing method. It is used when the demand history doesn’t really evolve in time.
    Constant method
  3. Trend: The trend method is an implementation of the double exponential smoothing method. It is used when a trend, either positive or negative, is observed in the sales history.
    Trend method
  4. Seasonal: The seasonal method is an implementation of the Holt-Winter’s method and is used when a seasonal pattern is detected in the demand history.
    A sesasonal forecast has a recurring pattern in time such as ice creams being much more sold during summer compared to winter.
    FrePPLe automatically detects for the periodicity of the recurring patterns.
    Seasonal method
  5. Intermittent: The intermittent method is an implementation of the Croston method. This method is suitable for sales history with intermittence, i.e. a sales history with many zeros.
    Intermittent method

The implementation of these methods requires some parameters. FrePPLe will automatically tune these parametes to minimize the forecast error. Advanced users can tune the allowed range of these parameters to tweak the results.

The forecast module also provides 3 other options to compute the statistical forecast:

  1. Automatic: This option is the one by default. FrePPLe will test all 5 statistical forecast methods described above and will pick the one that minimizes the forecast error.
  2. Manual: This option prevents frePPLe from computing any forecast. The forecast will be provided by the planner through overrides.
  3. Aggregate: It means that the forecast should be equal to the sum of the child intersections’ forecast. This option is deprecated as all parent intersections are automatically calculated without having to specify a record in the forecast table for the parent.