ABSTRACT

It is challenging to forecast total sales of short life-cycle products due to a lack of historical sales data. Multi-source information combination methods make it possible to depict different kinds of characteristics and features, given a limited volume of samples. Evidence theory is a common approach used for multi-source combinations. This paper proposes a new method, named ‘Multi-Evidence Dynamic Weighted Combination Forecasting (MEDWCF)’, based on improvements in the application of Evidence theory. Two kinds of machine learning methods are used to solve the basic probability assignment generation problem pertaining to Evidence theory, so a dynamic update combination algorithm is proposed. These innovations improve the classical one-step static combination rules. Samples of 313 films launched within 2016 and 2017 proved that compared with other forecasting methods, MEDWCF has more effectiveness and better generalisation ability. Effective product sales forecast by MEDWCF may help managers make correct decisions in manufacturing and marketing before the product launched.

摘要

由于缺乏历史销售数据，预测短生命周期产品的总销售额具有挑战性。在样本量有限的情况下，多源信息组合方法可以描述不同类型的特征和特征。证据理论是用于多源组合的常用方法。本文在改进证据理论应用的基础上，提出了一种新方法，称为“多证据动态加权组合预测（MEDWCF）”。利用两种机器学习方法解决证据理论中的基本概率分配生成问题，提出一种动态更新组合算法。这些创新改进了经典的一步静态组合规则。2016年和2017年上映的313部电影样本证明，与其他预测方法相比，MEDWCF具有更高的有效性和更好的泛化能力。MEDWCF 有效的产品销售预测可以帮助管理人员在产品推出之前做出正确的制造和营销决策。