Data-driven project buffer sizing in critical chains

In critical chain project management, the probability of timely project completion can usually be increased by setting a suitable project buffer. Existing studies tend to explicitly construct a formula to calculate the project buffer size, but there is a lack of a data-driven framework that sizes the project buffer through mining the buffer related information hidden in the project data. Therefore, from a perspective that is different from the current studies, we propose a data-driven project buffer sizing approach to size the project buffer in a prediction manner in the project planning phase. Our approach first uses a full-factor design of experiments and Monte Carlo simulation to construct the required dataset. Then, support vector regression is adopted to train the project buffer prediction model. The parameters of the support vector regression are fine-tuned via grid search and cross validation. Extensive computational experiments on the PSPLIB J30 dataset are conducted to validate our approach. Six performance measures (i.e., the mean squared error, the mean magnitude relative error, the percentage relative error deviation, the proportion of project buffer consumed, the average project buffer size and the probability of exceeding the planned duration) are employed to assess our approach. The experimental results show that our approach is competitive compared to classical project buffer sizing methods.