Improving automotive garage operations by categorical forecasts using a large number of variables

Cost effective job scheduling for garage management relies upon assigning repair times into appropriate categories rather than using exact repair time lengths. In this paper, we employ an ordinal logit model with least absolute shrinkage and selection operator (LASSO) to forecast such repair time categories for automotive engines. Our study is based on a unique dataset of maintenance records from the network of 64 UK garages of BT Fleet Solutions, and we consider a large number of predictor variables, with condition, manufacturing, geographical, and calendar-related information. The application of LASSO enables the identification of relevant predictor variables for forecasting purposes. Based on the Brier score and the ranked probability score (and their skill scores), we document substantial predictive ability of our method which outperforms five benchmarks, including the method used by the company. More importantly, we demonstrate explicitly how to associate the predicted probabilities with a loss function in order to make operational decisions in garages. We find that the best choice of job scheduling does not always correspond to the predicted categories, especially when the loss function is asymmetric. We show that scheduling jobs on the basis of our method can help the company reduce loss value. Finally, we identify opportunities for further improvements in the operations of the company and for garage maintenance operations in general.