Detecting Customer Trends for Optimal Promotion Targeting

Abstract

Retailers have become increasingly interested in personalizing their products and services such as promotions. For this, we need new personalized demand models. Unfortunately, social data is not available to many retailers for cost and/or privacy issues. Thus, we focus on the problem of detecting customer relationships from transactional data, and using them to target promotions to the right customers.Academic / Practical Relevance: From an academic point of view, this paper solves the novel problem of jointly detecting customer trends and using them for optimal promotion targeting. Notably, we estimate the causal customer-to-customer trend effect solely from transactional data, and target promotions for multiple items and time periods. In practice, we provide a new tool for Oracle Retail clients that personalizes promotions. Methodology:We develop a novel probabilistic demand model distinguishing between a base purchase probability, capturing factors such as price and seasonality, and a customer trend probability, capturing customer-to-customer trend effects. The estimation procedure is based on regularized bounded variables least squares and instrumental variable methods. The resulting customer trend estimates feed into the dynamic promotion targeting optimization problem, formulated as a non-linear mixed-integer optimization model.Though it is NP-hard, we propose an adaptive greedy algorithm.Results: We prove our customer-to-customer trend estimates are statistically consistent, and the adaptive greedy algorithm is provably good. Having access to Oracle Retail fashion client data, we show that our demand model reduces the WMAPE by 11% on average. Also, we provide evidence of the causality of our estimates. Finally, we demonstrate that the optimal policy increases profits by 3-11%.Managerial Implications: The demand model with customer trend and the optimization model for targeted promotions form a decision support tool for promotion planning. Next to general planning, it also helps to find important customers and target them to generate additional sales.

On Dynamic Pricing with Covariates

“…• Network/peer effects: A wide range of products are influenced by network or peer effects (Seiler et al 2017, Goolsbee and Klenow 2002, Bailey et al 2019, Nasr and Elshar 2018. Baardman et al (2020) show that demand prediction can be improved by incorporating such effects. Thus, the customers' features are more likely to exhibit short-term dependencies.…”

Data Analytics in Operations Management: A Review

“…However, for most retailers, acquiring these data is costly and poses challenges in terms of privacy concerns. Baardman et al (2018b) focus on the problem of detecting customer relationships from transactional data and using them to offer targeted price promotions to the right customers at the right time. The paper develops a novel demand model that distinguishes between a base purchase probability, capturing factors such as price and seasonality, and a customer-trend probability, capturing customer-tocustomer trend effects.…”

“…The relevance of causal inference to empirical OM has been highlighted previously (Ho et al 2017), but less research has highlighted the importance of considering causality in prescriptive modeling. We have already discussed one example, the work of Baardman et al (2018b), which develops a specialized estimation method for detecting customer-to-customer trends that uses the instrumental variables approach. Another example is the work of Bertsimas and Kallus (2016), which considers the problem of pricing from observational data.…”

Convex Loss Functions for Contextual Pricing with Observational Posted-Price Data

“…A popular approach in practice is a predict then optimize, or direct method approach, whereby an intermediate contextual demand function is estimated to predict the probability a customer purchases at a given price, and then optimized to maximize revenue (Chen et al 2015, Ferreira et al 2016, Dubé and Misra 2017, Alley et al 2019, Baardman et al 2020, Biggs et al 2021b.…”