Browse Theses

Forecasting demand is important to guide decision making in operations management. However, in many applications, there is not much data. For example, though many retailers have abundant historical data for existing products or from related environments, there is no historical data for new products or a new environment for which a demand forecast is needed. Limited data makes building accurate forecasting models very challenging. One idea is to extrapolate from historical data for existing products or a related environment. I explore this idea in two settings.The first is the problem of forecasting demand when a new product is introduced. Regression type methods use covariate information and historical sales for existing products to predict demand for existing and new products. However, they ignore cannibalization which is important in forecasting. Limited data makes estimating cannibalization in the forecasting problem more difficult. From the real problem of forecasting demand for new and existing products in the Body Shop, I show the importance of incorporating cannibalization effects in the forecasting problem and provide an approach to estimating cannibalization when a new product with no historical sales data is introduced. I propose a hybrid demand model that combines linear model and choice models to incorporate cannibalization with limited data. The linear model captures the effects of time varying factors like discounts, promotions on holidays which affect not only the sales on a given product but also total demand. The choice model is used to model the market share of products. It naturally leads to an estimate of cannibalization when a new product is introduced. When building the choice model, I cluster products with similar characteristics, and build choice models over clusters to address the issue of having the large choice set and no observed "choices" of the new products. I apply this hybrid model to a real application of the Body Shop.The second problem considered is that of estimating a choice model with limited data. It is assumed, however, that a related dataset is available. A method is proposed to combine both datasets by selecting the parameters of the choice model so that it approximately fits both. The objective of the proposed mothod can be regarded as a generalized version of ridge regression but is different in that it shrinks estimates towards the Maximum Log-likelihood Estimator (MLE) of the old data set βold, and prioritizes being close to components of βold that are "reliable". The advantage of the approach is illustrated to be relative to classical regularization methods by simulations.