What is bias ?
If you observe forecast error going in one direction or other, then you have a forecast which is possibly biased. In the example below try to observe the bias and see if you are able to find the forecast versions with bias.
As you would have pointed out already following is the observation on above example:
- Forecast1 has a positive forecast bias
- Forecast2 has a negative forecast bias
- Forecast3 has mixed bias and the cumulative bias in such case could be insignificant
Decision on whether a forecast is biased or not can be made by reviewing the forecast error as well, study the example below.
Why it exists?
As final forecast is combination of statistical forecast and manual intelligent updates, the bias can slide into Final forecast through any of the mentioned input gateways. While statistical forecast bias are normally specific to items(local bias), bias building up due of manual intervention usually has impact on all the items(global bias).
Manual updates done by forecasters/demand planners at times lead to building of a biased forecast. These updates are driven by various factors like:
- Increasing forecast to match up with volume targets
- Optimistic or pessimistic approach towards forecasting
- Expecting Promotional incremental sales volume
This kind of bias usually has impact on all the items.
Bias found in statistical forecast is most of the times under-forecasting or over-forecasting situation for a specific item-location combination. Few of the cases which lead to such observation can be list as:
- Under-forecasting for complete future horizon due of recent months sales volume showing a down trend.
- A persistent trend in sales volume
Such forecast bias is normally particular to an item or an item-customer combination.
Getting rid of Bias?
Bias shall be removed from forecast as it can assist in improving your forecast accuracy, which eventually reflects across supply chain health. An overall reduction on forecast across all items can take out the global bias(e.g. 15% decrement on forecast numbers all across). For bias which is specific to item, one needs to identify and fix them for every incident by adjusting the forecast.
