New and Short-life-cycle Product Procurement Decisions: Machine Learning and Risk Control

Description

Fashion and seasonal products, which are characterized by short selling seasons, highly uncertain demand, and often long lead times, have brought a great challenge for retailers trying to match supply with demand. As a result, retailers often suffer from significant losses in profit due to either lost sales from stockouts or clearances of excess inventory at the end of the selling season. The situation for new, seasonal products is even worse, as new products do not have historical demand/sales data. However, a firm may have been selling similar products in the past and keeps a record of them. In addition to demand/sales figures, the data record may contain rich information about the attributes (features) of the products, such as retail price, color, item type (e.g., shirt, pant), fabric, design style (e.g., casual, sporty), store type, and season, in the case of fashion cloths. Such attributes are called covariate information to demand.In this project we attempt to link a new product, by using covariate information, to "similar" products that were sold historically. Weights are used to measure similarities between the new product and historical products, and the values of those weights are estimated by employing machine learning methods such as k-nearest neighbors, classification and regression tree, and random forests, to the data. Then, the pair of the realized demand of a similar historical product and its associated weight, together with those from other similar products, are utilized to approximate the expected profit and other quantities which take on the (conditional) demand distribution.Two basic model settings are considered. The first is the single-period/newsvendor type, and the second is the two-period type, which we believe to be most representative of realistic situations for many short-life-cycle products. As to the objective criterion, the firm determines the order quantities with a chance constraint that requires meeting a profit target at a chosen confidence level, while aximizing expected profit. The chance constraint is also called a value-at-risk (VaR) constraint, reflecting the fact that managers seek to meet profit targets because missing a target affects both the stock price and bonuses. The approximated formulation using weights can be transformed into a mixed integer programming, for which an efficient algorithm will be developed. We will prove the proposed approximation to be asymptotically optimal even with the sample-dependent approximation for the VaR constraint. We will also use real-world data to verify our models and methods.