Accuracy in Sales Forecasting

Forecast Accuracy

Calculating the Absolute Error
The Mean Absolute Error has strong capabilities for assessing forecast accuracy in the context of inventory optimization and it is very simple to calculate and use.
The Absolute Error is the absolute difference between forecasted and actual value in number of items. Intuitively, we can think of the Absolute Error as the number of items the forecast is off from what actually happens. Absolute means that the formula disregards whether the forecast is too high, or too low, all that counts is by how many items the forecast is off the actual value, negative algebraic signs are therefore not regarded.

Calculating MAPE
MAPE stands for Mean Absolute Percentage Error. We arrive at MAPE by dividing the Absolute Error by the forecasted value. Intuitively, MAPE as a percentage provides us with a measurement of the forecast error relative to the actual value.
reIn this example we take a look at a time series of three actuals and the respective forecasts, for which we calculate the MAPE. The absolute error of the first data pair is 110 -- 90 = 12 items. By dividing this value by the forecasted value we arrive at a percentage value, in this case 0.11. We can read this as the forecast has an absolute percentage error of 11percent. In the example we have done this for all data pairs, arriving at absolute percentage errors of 11percent, 25percent and 11percent. By taking the mean we arrive at the MAPE, which in this example is 16percent.

Lets take a look why MAPE is frequently not suited to compare forecasts.

MAPE is in most cases not suited to compare sales and demand forecasts. The main issue is its sensitivity to sparse time series. Sparse time series are items selling in very low quantities; most retailers have a large amount of this type of low selling products in the long tail of their product portfolio.

What is a good forecast?

A quantitative assessment of the accuracy of a forecast is meaningless without the context. For example, 5percent forecasting error is a bad result when forecasting the national electricity consumption 24h ahead while 80percent forecasting error is extremely good when forecasting first day sales of a new product.
Factors affecting the accuracy of a demand forecast are manifold, such as volatility of demand, data aggregation level, the amount of available historic data, the forecast horizon, sparse data, availability of event data and many more.
The easiest way to set the context is by comparing a forecast to the status quo or a competing forecast. A good forecast is the forecast that is more accurate than the next best alternative! Please remember, when comparing accuracies the Absolute Error should be used as it avoids distortions due to sparse time series.

Category:

Education

Tags:

• MAPE
• sales forecasting
• accuracy
• absolute error