Estimating individual promotional campaign impacts through Bayesian inference

Purpose – The purpose of this paper is to estimate individual promotional campaign impacts through Bayesian inference. Conventional statistics have worked well for analyzing the impact of direct marketing promotions on purchase behavior. However, many modern marketing programs must drive multiple purchase objectives, requiring more precise arbitration between multiple offers and collection of more data with which to differentiate individuals. This often results in datasets that are highly dimensional, yet also sparse, straining the power of statistical methods to properly estimate the effect of promotional treatments. Design/methodology/approach – Improvements in computing power have enabled new techniques for predicting individual behavior. This work investigates a probabilistic machine-learned Bayesian approach to predict individual impacts driven by promotional campaign offers for a leading global travel and hospitality chain. Comparisons were made to a linear regression, representative of the current state of practice. Findings – The findings of this work focus on comparing a machine-learned Bayesian approach with linear regression (which is representative of the current state of practice among industry practitioners) in the analysis of a promotional campaign across three key areas: highly dimensional data, sparse data and likelihood matching. Research limitations/implications – Because the findings are based on a single campaign, future work includes generalizing results across multiple promotional campaigns. Also of interest for future work are comparisons of the technique developed here with other techniques from academia. Practical implications – Because the Bayesian approach allows estimation of the influence of the promotion for each hypothetical customer’s set of promotional attributes, even when no exact look-alikes exist in the control group, a number of possible applications exist. These include optimal campaign design (given the ability to estimate the promotional attributes that are likely to drive the greatest incremental spend in a hypothetical deployment) and operationalizing efficient audience selection given the model’s individualized estimates, reducing the risk of marketing overcommunication, which can prompt costly unsubscriptions. Originality/value – The original contribution is the application of machine-learning to Bayesian Belief Network construction in the context of analyzing a multi-channel promotional campaign’s impact on individual customers. This is of value to practitioners seeking alternatives for campaign analysis for applications in which more commonly used models are not well-suited, such as the three key areas that this paper highlights: highly dimensional data, sparse data and likelihood matching.