Abstract
As product variety is an important feature for modern enterprises, multi-product demand forecasting is essential to support order decision-making and inventory management. However, these well-established forecasting approaches for multi-dimensional time series, such as Vector Autoregression (VAR) or dynamic factor model (DFM), all cannot deal very well with time series with high or ultra-high dimensionality, especially when the time series are short. Considering that besides the demand trends in historical data, that of associated products (including highly correlated ones or ones having significantly causality) can also provide rich information for prediction, we propose new forecasting approaches for multiple products in this study. The demand of associated products is treated as predictors to add in AR model to improve its prediction accuracy. If there are many time series associated with the object, we introduce two schemes to simplify variables to avoid over-fitting. Then procurement data from a grid company in China is applied to test forecasting performance of the proposed approaches. The empirical results reveal that compared with four conventional models, namely single exponential smoothing (SES), autoregression (AR), VAR and DFM respectively, the new approaches perform better in terms of forecasting errors and inventory simulation performance. They can provide more effective guidance for actual operational activities.
Highlights
Demand forecasting, a prerequisite for inventory decision-making, plays a vital role in supply chain management
By summarizing previous literature related to multi-dimensional time series analysis, we find that these methods all fail to deal with the situation where product dimension is large but time dimension is small
CI and CII refer to the approach based on correlations adopting scheme (I) and scheme (II) to control model size respectively
Summary
Demand forecasting, a prerequisite for inventory decision-making, plays a vital role in supply chain management. How to improve prediction accuracy has always been the focus of academic circles and enterprises. Traditional time series algorithms cannot well adapt to the complex high- or even ultra-high dimensionality, resulting in inferior predictive effectiveness in multi-product scenarios. It is worth noting that the demand of multiple products is not completely isolated, but rather complex relationships exist between them. According to the relevant literature, there are two common association relationships between different products: correlation and Granger causality. Materials used in engineering projects have a clear sequence, so Granger causality exists in their demand [3]. Capturing and making full use of such potential information can be helpful to obtain more accurate prediction results. What’s more, when the time series are short, historic trend cannot provide enough information for future demand.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have