Improving Automotive Aftermarket Forecasting with MLOps

All vehicles on the road require maintenance as parts wear out. To respond to the need for new units without over or underestimating, it is essential to have a forecast model that indicates the quantity of parts needed with minimal possible error. All vehicles have thousands of parts, and each part is compatible with only a small set of vehicle series, increasing the total number of different parts in the market that need to be forecasted. Managing thousands of Machine Learning (ML) models with remarkable complex data in highly complex environments requires a Machine Learning Operations (MLOps) approach to have a reliable ML system from the beginning to the end, with a well-defined and transparent protocol to declare all ML life cycle workflows. In this study, with an MLOps system, we analyze the forecast capacity in two horizons, 6 and 12 months, with two Deep Learning (DL) approaches in a real-world automotive parts dataset. The DL approaches assume two different forecasting strategies: a multi-time series (one single model for all predictions) and a single-time series (one model for each time series). The multi-time series approach outperformed the single-time series, regardless of the horizon.