Forecasting of customer demands for production planning by local k-nearest neighbor models

Demand forecasting is of major importance for manufacturing companies since it provides a basis for production planning. However, demand forecasting can be a difficult task because customer demands often fluctuate due to several influences. Methods of nonlinear dynamics have shown promising results in numerous applications but they have mostly been neglected in the context of demand forecasting. This paper evaluates the forecasting performance of local k-nearest neighbor models, which base on the theory of dynamical systems, in a comprehensive empirical study utilizing a large dataset of industrial time series of the M3-Competition. After a broad literature review, the theoretical background is described. Subsequently, different parameter configurations and model selection strategies are compared. A locally constant mean and a locally constant median are compared to locally linear regression models with four different regularization methods and different parameter configurations. In this comparison, the locally constant mean and the locally linear ridge regression with high regularization parameters provide the best trade-offs between forecast accuracy and computation times. Finally, these models achieve a high performance regarding low forecast errors, short computation times as well as high service levels in an inventory simulation compared to established benchmark methods. In particular, they obtain the best results among all applied methods regarding short time series. Moreover, they achieve the lowest errors considering the original accuracy criterion of the M3-Competition. Hence, local k-nearest neighbor models can be regarded as a valid alternative for demand forecasting in an industrial context, accomplishing high forecast accuracy with short computation times.