Calling the ATT forecast engine with Best algorithm

This section provides an overview of the required conditions to call the forecast engines, and to generate the forecasts for the selected item/location, when the CallForecastEngine command is used.

When the engine is of the type ATT, the algorithm specified is Best (or Null), the system selects the Best algorithm from multiple methods. The engine selects a model for the data using these techniques based on the history length (See, the algorithm documentation for more details):

  • General algorithm:
    • Least squares with forecast trend damping of 0.95
    • 4 and 6 point Moving Averages
    • 4 point Exponential Moving Average
    • BATS (Bayesian Analysis of Time-Series) with additive seasonality and forecast trend damping of 0.95
    • Constant BATS
    • Linear BATS with forecast trend damping of 0.95
    • Naïve with Drift and forecast trend damping of 0.95
    • Naive with Seasonality
  • Algorithm using the MEDIANS initialization method:
    • Constant Holt-Winters
    • Linear Holt-Winters with trend damping of 0.95
  • Algorithm using the AVERAGE initialization method:
    • Constant Holt-Winters
    • Where the history length => 2 * the periodicity of the data:
      • Linear Holt-Winters, with trend damping of 0.95
      • Holt-Winters with additive seasonality and trend damping of 0.95
Note: 
  • The initialization Type parameter is not considered for these techniques.
  • The History Trend Damping factor parameter and BATS Growth Damping parameter are not considered for the trend damping specified in the algorithms.
  • By default, the engine selects the model with the lowest value for the specified decision criteria.
  • If the number of history data points is less than the required moving average points, the initialization techniques cannot be used. For example, if 5 history points are passed for an item and location combination, the 6 point moving average technique cannot be used.