Double Exponential Smoothing
…like regular exponential smoothing, except includes a component to pick up trends.
- Ft = Unadjusted forecast (before trend)
- Tt = Estimated trend
- AFt = Trend-adjusted forecast
Ft = a* At-1 + (1- a) * (Ft-1 + Tt-1)
Tt = b* (At-1-Ft-1) + (1- b) * Tt-1
AFt = Ft + Tt
To start, we assume no trend and set our “initial” forecast to Period 1 demand.
… We then calculate our forecast for Period 2.
… But Period 2 demand turns out to be 20.
What is the trend estimate for Period 3?
By Period 4, the model is starting to pick up on the trend …
…And after a few periods the model “locks on” to the correct trend value.
(Of course, this example is simplified to make the numbers clearer.)