In the next SmartMobility cluster meeting this Friday (Oct 28), 12:30-13:30h, in Pav. K16, Sander Peters will give a talk about his recently finished graduation project; see below for the abstract.
Title: Throughput time and time window estimation for business processes using historical data
Abstract: Throughput time prediction is important for companies to plan resources or provide customers with an estimated completion time of their case. In order to increase the reliability of these estimates the new developed technique generates empirical probability distributions that describe when each event in the process occurs, based upon historical data. These empirical distributions are combined into a new empirical distribution to estimate the throughput time for a certain event. Using these empirical distributions time window estimates are derived with a certain precision.
Based upon the related work in the field of throughput time prediction a gap in literature on statistical techniques for throughput time prediction is found. The gap consists of the lack of the ability to use statistical techniques to obtain time window estimates. The new technique is first evaluated on an artificially generated dataset and outperforms the existing techniques for this dataset. A case study at Van Opzeeland is performed to evaluate the new technique on processes, in which there is a single possible execution path without choices, on a real-life dataset. The throughput time estimates are about 50 percent improved compared to the existing techniques and in 70 percent of the cases the time range wherein a certain even is finished is smaller with the same precision compared to the existing techniques. In order to also handle processes in which choices can be made an extension has been developed to the new technique. This extension uses the probabilities of each variant to estimate throughput times and provide estimates for time windows. The extension has been verified on a case study using the real-life dataset from the BPI 2015 challenge. For the case study the new technique performs less well than the average value for the throughput time estimates. Also the time range is larger than the average value with a markup value added to it. This might be an issue of the technique or could be caused by the assumption that each process time of an activity is independent of the previous activities. The new technique provides statistically reliable time windows and therefore can improve, for example, resource scheduling, since reliable estimates provide indicators when a resource is needed for a specific case.