DPM: A State Space Model for Large-Scale Direct Marketing

TL;DR

The paper introduces the Dynamic Propensity Model, a latent variable time series approach that leverages marketing and purchase data to optimize channel selection, offers, and timing in large-scale direct marketing campaigns.

Contribution

It presents a novel statistical methodology combining particle methods and stochastic gradient descent for fast parameter estimation in a large-scale latent variable model.

Findings

Model effectively predicts customer responses across channels and products.

Marketing touch effects vary significantly by channel and product.

Outperforms lagged variable logistic regression in predictive accuracy.

Abstract

We propose a novel statistical model to answer three challenges in direct marketing: which channel to use, which offer to make, and when to offer. There are several potential applications for the proposed model, for example, developing personalized marketing strategies and monitoring members' needs. Furthermore, the results from the model can complement and can be integrated with other existing models. The proposed model, named Dynamic Propensity Model, is a latent variable time series model that utilizes both marketing and purchase histories of a customer. The latent variable in the model represents the customer's propensity to buy a product. The propensity derives from purchases and other observable responses. Marketing touches increase a member's propensity, and propensity score attenuates and propagates over time as governed by data-driven parameters. To estimate the parameters of the model, a new statistical methodology has been developed. This methodology makes use of particle methods with a stochastic gradient descent approach, resulting in fast estimation of the model coefficients even from big datasets. The model is validated using six months' marketing records from one of the largest insurance companies in the U.S. Experimental results indicate that the effects of marketing touches vary depending on both channels and products. We compare the predictive performance of the proposed model with lagged variable logistic regression. Limitations and extensions of the proposed algorithm are also discussed.