Evaluation of forecasting techniques and forecast errors with focus on intermittent demand
Abstract: To decide in advance the amount of resources that is required next week or next month can be both a complicated and hazardous task depending on the situation, despite the known time frame when the resources are needed. Intermittent demand, or slow-moving demand, that is when there are time periods without demand and then suddenly a time period with demand, becomes even more difficult to forecast. If the demand is underestimated it will lead to lost sales and therefore lost revenues. If the demand is overestimated, in the best case the stock is increased or in worst case, the items lie unsold until they become obsolete. The items with intermittent demand can have a value of up to 60% of the total stock value for all items. This thesis addresses the topic of forecasting intermittent demand and how to measure the accuracy of the chosen forecast method or methods. Four forecasting methods are tested on almost 18 months of empirical demand data from a manufacturing company. The tested forecasting method are single exponential smoothing, Croston and two modification of the Croston method, one by Syntetos and Boylan the other by Segerstedt (modified Croston). Four start values and eight smoothing constants are tested. The methods are evaluated with different accuracy measures; variance (MSE and MAD), bias (CFE, the maximum and minimum value of CFE) and sMAPE. In addition with a new complementary measure of bias; Periods in Stock (PIS), PIS considers the time aspect, when the forecast error occurred not just the error size. Also two variants of MAD and MSE are tested. To improve the evaluation of the bias measures, the percentages of demand occasions that can not be fulfilled are used. The relationship between the different errors for a certain method is examined with principal component analysis (PCA). The errors are also examined with logistic regression to find out if a certain forecasting method is favoured by certain accuracy measures. The logistic regression is based on descriptive statistics for time series plus the mean absolute change that considers the sequence of the time series as well as the variation. Ranking and error quotients between different methods are other applied methods. The results of the research both confirm and contradict earlier findings. Among the confirming research results are the bias among the different methods. Croston and Modified Croston are overestimating the demand, Syntetos and Boylan's Croston variant has a tendency to underestimate the demand. Single exponential smoothing is relatively biasfree when low smoothing constants are concerned. The contradictive results are that CFE is not a suitable measure of bias at least when the number of forecasting periods is limited. The value of CFE can indicate a nonbiased forecast when both PIS and the percentage of unmet demands indicate a biased forecast. PIS is also less sensitive to transient demand events that can distort CFE. PIS is recommended as a bias measure for limited time series, especially considering intermittent demand, along with the percentage of unmet demand. Another result is that MAD is not reliable since the method in certain circumstances favours methods that underestimate the demand.
This dissertation MIGHT be available in PDF-format. Check this page to see if it is available for download.