View the model
Download the model

predicts future values by using a weighted average to "smooth" past values. The basic model for exponential smoothing is to predict sales in period t+1 by:

P( t+1) = P( t) + alpha*( D( t) - P( t) )
where
P( t+1) = predicted value for period t+1, computed in period t,
D( t) = actual value at period t
alpha = smoothing constant that measures how much we adjust our forecaset based on most recent error.

If alpha is set to 1, the forecast for the next period is based entirely on the actual value from the last period.
If alpha is set to 0, the actual value from the last period is completely ignored.
Because neither of these extreme cases will provide much insight into future data, we'll constrain alpha to be between .01 and .99.

There are extensions to simple exponential smoothing that allow one to also take into account
trends and seasonality.

Objective of Optimization:
Determine the best smoothing constant alpha that minimizes the one-period-ahead squared forecast error.