High-Low Promotion Policies for Peak-End Demand Models - Georgia Perakis, MIT Sloan School of Management
Management Science Research Centre & Bensadoun School of Retail Management Present
Georgia Perakis
MIT Sloan School of Management
Ìý
Friday, November 30, 2018
1:30PM - 3:30PM
BRONF 340
High-Low Promotion Policies for Peak-End Demand Models
Ìý
Abstract:
Promotions are a highly effective marketing tool that can have a significant impact on a retailer’s profit. A strong understanding of how changes in the price affect consumers' purchasing behavior can lead to more effective promotions policies and as a result, to a substantial increase in profit for retailers. Incorporating important consumer behavioral effects in the demand model is crucial in order to better predict demand. In this talk, we will present a new demand model that relies not only on current and past period prices but more importantly, on the minimum price set within a set of past periods (bounded memory peak-end). Furthermore, using these as features and employing machine learning tools, we show that this new demand model predicts actual sales more accurately than current methods. We test our prediction approach on sales data from a large retailer and demonstrate that there is a 9% relative improvement in the precision of the demand prediction.
This new demand model also allows us to determine the optimal promotion strategy more efficiently. That is, subsequently, we suggest a compact Dynamic Programming (DP) approach that uses the proposed demand model. We examine when this DP solves the problem optimally. That is, we establish when, for some commonly used demand models (including the one proposed in this talk), the proposed DP solves the promotion planning optimization problem exactly. In fact, we confirm a common practice by retailers, that is when and for what demand models, the optimal promotion strategy is to either promote (always at the same level of promotion) or not promote an item at all. For demand models where these conditions do not hold, we provide an analytical guarantee for our proposed DP and illustrate that still the proposed DP yields near optimal solutions fast. Furthermore, on the same sales data we tested our demand prediction approach on, we demonstrate that the proposed DP yields on average a 9.1% increase in profit relative to the retailer's current practices.
Ìý
All are cordially invited to attend.
