Demand Planning (SCM-APO-FCS)

First-Order Exponential Smoothing

Exponential smoothing methods are the most widely accepted time series techniques in use today. They were originally called "exponentially weighted moving averages." The basic premise of single exponential smoothing is that the sales values for more recent periods have more impact on the forecast and should therefore be given more weight, while the weights for older periods will decrease at an exponential rate. In addition, because the calculations require more recent sales history, data storage is minimized (or at least reduced) as a result of the minimal historical data required.

First-order exponential smoothing, also known as single exponential smoothing, uses a smoothing constant (alpha) to which a value between 0 and 1 is assigned. The larger its value (closer to 1), the more weight it assigns to recent sales history. A large alpha (.8) is comparable to using a small number of time periods (n) in a moving average model. A small n allows greater emphasis to be placed on recent periods. Conversely, a small alpha (.1) is similar to using a large number of time periods in the moving average, because the impact of recent data is lessened.

The strengths of exponential smoothing models are that they:

• Are reasonably simple to understand and use

• Provide more weight to recent data periods

• Do not require much data storage

• Have fairly good accuracy for short-term forecasts (one to three periods out into the future)

The weaknesses of exponential smoothing models are that:

• A great deal of research may be required to find the correct alpha value

• They are usually weak models to use for medium or long-range forecasting (three periods and beyond)

• Forecasts can be thrown into great error because of large random fluctuations in recent data. Because they rely heavily on past history and on a smoothing factor to predict the future, exponential smoothing models cannot easily predict turning points in recent data. At least one to three periods are usually needed to correct for extreme fluctuations in recent data.

Second-Order Exponential Smoothing

The method of first-order exponential smoothing is theoretically appropriate when the data series contains a horizontal pattern (that is, it does not have a trend). If first-order exponential smoothing is used with a data series that contains a consistent trend, the forecasts will trail behind (lag) that trend. Second-order exponential smoothing, also known as Holt's linear exponential smoothing, avoids this problem by explicitly recognizing and taking into consideration the presence of a trend. It prepares a smoothed estimate of the trend in a data series.