Improving the forecasting performance of Purac

Corbion-Purac is an expert partner in technology- and application development for Poly Lactic Acid (PLA) and the worldwide market leader in Lactic Acid, Lactic Acid derivatives and Lactides. From five production locations products are distributed over the whole world. To manage the logistics of the Supply-Chains, forecasts of the demand are used extensively. The challenge for this master-thesis project is to support Purac in improving the forecasting performance. Forecasting performance is defined as the efficiency of the forecasting process. This means that besides the forecast accuracy, the required effort is also important. If less effort is required to get the same accuracy, this is seen as a better forecasting performance. First, a thorough analysis of the forecasting practices at Purac was done. This lead to a list of nine measures that can help to improve the forecasting performance of Purac. Seven of the nine measures concern statistical time-series forecasting. So, there is a strong focus in this list on statistical time-series forecasting. The reason is that the process that leads to the judgmental forecasts was found to be quite well organized. It has the right incentives, checks and balances and the people that have the most knowledge of the client and market are ultimately responsible for forecasting. Explanatory models have also been left out, because these are more difficult, currently not used and require different data. The list of nine measures was shrunk down to a smaller set of five promising measures. This is done with the following two selection criteria: ease of operation (use and maintenance) and ease of implementation (in ICT-system). For each selection criterion a questionnaire was made. The combined end-results of this lead to a reduction of the list of nine measures to a shorter list of five measures: totally freeing the t-2 forecast for products with a KPI based on the t-1 forecast, optimized parameters for forecasting models (Moving Average and Single Exponential Smoothing), selecting optimal forecasting method, optimization based on the fixed period of a product and using alternative loss-functions for optimization. For these five measures a quantitative analysis is done. The five measures were tested on 85 products for which already a statistical forecasting method is used at Purac. The increase of the accuracy of the statistical forecast by totally freeing the t-2 forecast is estimated to be around 1,5%. The effect on the final forecast is estimated to be 2,3% for the Alif (?) and 1,6% for the S?n (?) region. Using optimized parameters for the forecasting methods can considerably improve the forecast accuracy of the statistical forecasts. For the products for which the Moving Average (MA) method is used, this is estimated to be 20% and for the Single Exponential Smoothing products around 5%. Optimizing the forecasts on the fixed period does not have any effect on the accuracy. For the other tests the results were inconclusive. To gain more confidence in the wider applicability of the results detailed testing was done regarding: the optimized models compared to judgmental forecasting, the relative performance with different loss-functions, the relative performance of three alternative forecasting methods (MA, SES, best fit method) and the effect of freeing the t-2 forecast on the accuracy of the statistical forecast. For this a dataset was selected in a specific way. Five Strategic Product Groups (SPGs) were selected that have a relative high volume and low accuracy. The idea was that for these products there is more potential to improve the accuracy, because of the current low accuracy and that it has a big effect overall, because it has a relative high volume. From the detailed testing, some important conclusions are drawn. Instead of just focussing on the exact numerical results, the conclusions are summed up here: • The effect of freeing the t-2 forecast is bigger than in the previous tests. This is now estimated to be around 2%. This effect is smaller for the currently used statistical methods and bigger for the alternatives. • It is better to report the median forecast accuracy alongside the average forecast accuracy when measuring the accuracy over multiple months. The reason is that the average forecast accuracy is sensitive for outliers. If a forecasting method has overall good results, but in one or two periods very bad results, the average accuracy will be low. But the median accuracy will hardly be affected by this. Both statistics together give a completer picture. • The selection method that was used to find SPGs for which the statistical forecasting methods were supposed to have a high accuracy compared to the judgmental forecast, is actually not a good one. The main reason is that with this method especially the less-smooth products are selected, which are the most difficult to forecast with statistical methods. • Statistical forecasting cannot be applied blindly to every product. Segmentation is needed on the item hierarchy level instead of the SPG level. • A sigma of 1,5 gives an additional increase of the accuracy of about 1,5%-2,0%. A period of 12 months is used for fitting. This gives good results for both MA and SES. A maximum for the N of the MA-method (or a minimum for the ? of the SES) is not used, as this gives conflicting results and does not have a big impact. • MAD is the only loss-function among the tested four, which gave good results for the MA and SES method for both datasets. • Among the three fitting methods, the SES method gave here the best results and the MA method the worst. In the previous test, the SES-method was the worst. In both cases the best-fit method had the second best results. That is why it is decided to use the best-fit method to work out the implementation. Finally, implementation alternatives were worked out for two regions as an example. The implementation alternatives concern how to segment the products and use the combination of a statistical and judgmental forecast for each segment. None of the proposed segmentation alternatives leads to the ability of identifying products for which the statistical forecast leads to a better forecast than the judgmental forecast. Among all the implementation alternatives volume segmentation is the most preferred. Using volume segmentation leads to a consistency of the accuracy among the segments (A segment has a higher accuracy for the statistical forecast than the B segment etc.) and the number of products and volume stay stable over the years. Besides, the important products are differentiated from the less important ones based on the volume. This can help the planners to focus on the right products. Based on this it is proposed not to use a combined volume-accuracy segmentation (which would lead to nine segments), but only volume segmentation (which leads to three segments). Using two dimensions would make it unnecessarily complex without much additional benefit. Three sub-alternatives have been worked out that are used in combination with volume segmentation. The first two concerned the type of forecast to use for the C-segment. For these sub-alternatives, like currently for the A and B segment judgmental forecasting is used. Sub-alternative 1 and 2 are the easiest to understand. Because of this, these are expected to be easier to implement. From these alternatives it was also concluded that for the C-segment an accuracy increase/decrease of 1% has an overall effect in the order of 0,1%. The C-segment makes up more than 50% of the items. Sub-alternative 3 gives a specific forecasting strategy for each segment. It results in the end in the highest accuracy and a focus on the most important items. Sub-alternative 1 and 2 can be used as intermediate steps towards this final solution. The following is proposed for the end-implementation: 1. A-segment: use the statistical forecast as a reference and comparison to the judgmental forecast. It is recommended to have an intensive contact with the customers of these products and gain as much as possible information that is of importance for the forecast. 2. B-segment: use the statistical forecast as the standard, but thoroughly check it and make judgmental adjustments. 3. C-segment: use the statistical forecast as the standard and only make changes when felt necessary. The forecast accuracy of this group is not that important, because it does not have much effect on the overall forecast accuracy. So this group should be given the least attention. The details regarding the implementation can be found in appendix I. Through the research recommendations are done regarding how to improve the demand planning process and increase the forecast accuracy. These are summed up together in the chapter “Conclusion and recommendations” (p.81).