An Adaptive Data-driven Model for Erratic Demand Forecasting

Abstract

Abstract Erratic demand patterns are common in spare parts supply networks. Owing to the characteristics of nonlinear dynamics, aperiodic variations and deep uncertainties, erratic demand forecasting remains to be a challenge. This paper is devoted to forecast this type of demand in a more difficult situation where regular explanatory variables are not available and historical data are limited. To address this problem, we propose an Adaptive Autoregressive Support Vector Machine model in which: 1) autocorrelated attributes are generated from historical demand time-series data automatically; 2) the attribute dimension and the suitable nonlinear kernel mapping function are identified in a data-driven manner; and 3) key model parameters are adaptively controlled by a parallel heuristic algorithm to cope with data uncertainties and guarantee the model generalization ability. We test this model by erratic demands of heavy truck spare parts. Computational results demonstrate that the proposed model outperforms seven sound time-series forecasting models and general Support Vector Machine models.