Simple Forecasts Tend To Work Best
The best we can hope for from a forecast is to be correct on average. A forecast should not add more noise than is present in the sales data, and it shouldn’t be biased.
|
Rolling average forecast
A simple moving average forecast is a good way of stripping out the noise and detecting underlying signals and trends within the sales data.
The adjacent graph shows forecast bias for the same client data set as used in the bias section, now with the bias of a four week rolling average forecast included.
The bias has clearly been reduced very significantly by the rolling average forecast, proving that the bias extremes in the client forecast are a result of human interventions.
|
| |
Exponentially weighted forecast
One weakness of the rolling average forecast is a tendency to be slow at picking up signals in the sales data. This can result in the planners having insufficient time to respond to the signals.
An exponentially weighted forecast is a good way of tackling this. Such a forecast gives recent weeks more prominence in the forecast calculation than less recent weeks.
Seasonality and Events
Seasonality and events need special treatment. While we cannot forecast noise which should be buffered against using safety stock, we should be able to forecast the effects of seasonality, and anticipate the effects of special events, such as promotions and new product introductions.
These deal with a longer time horizon than the operational forecast, and are therefore included within Sales and Operations Planning. Once the longer term forecast has been generated, it should be overlaid onto the operational forecast.
|
|