Hold-out periods

You can use hold-out periods to define the number of periods before the forecast begins. The history in this range is disregarded during the forecast calculation, and a forecast is calculated for the hold-out periods. The hold-out forecast is then compared to the history in the hold-out range to calculate the error analysis. The winning method is then recalculated with the entire history including the hold-out periods to determine the resulting forecast. The resulting forecast is written to the database.

This diagram shows how a winning forecast is selected using Competitive Forecasting with hold-out periods:

In an 18-period history with six hold-out periods, Competitive Forecast uses periods 1 to 12 of the period history and compares them to the Mean Absolute Percentage Error (MAPE) in periods 13 to 18. Use Competitive Forecasting to select the winning method using minimum MAPE, exluding methods that cannot calculate.

When the winning method is determined, the forecast of the winning method is calculated using the entire 19-period history. A forecast for the entire horizon (extrapolation range) is then produced.