Paper Type

full

Description

Firms seek to better understand heterogeneity in the customer response to marketing campaigns, which can boost customer targeting effectiveness. Motivated by the success of modern machine learning techniques, this paper presents a framework that leverages deep-learning algorithms and field experiment response heterogeneity to enhance customer targeting effectiveness. We recommend firms run a pilot randomized experiment and use the data to train various deep-learning models. By incorporating recurrent neural nets and deep perceptron nets, our optimal deep-learning model can capture both temporal and network effects in the purchase history, after addressing the common issues in most predictive models such as imbalanced training, data sparsity, temporality, and scalability. We then apply the learned optimal model to identify customer targets from the large amount of remaining customers with the highest predicted purchase probabilities. Our application with a large department store on a total of 2.8 million customers supports that optimal deep-learning models can identify higher-value customer targets and lead to better sales performance of marketing campaigns, compared to industry common practices of targeting by past purchase frequency or spending amount. We demonstrate that companies may achieve sub-optimal customer targeting not because they offer inferior campaign incentives, but because they leverage worse targeting rules and select low-value customer targets. The results inform managers that beyond gauging the causal impact of marketing interventions, data from field experiments can also be leveraged to identify high-value customer targets. Overall, deep-learning algorithms can be integrated with field experiment response heterogeneity to improve the effectiveness of targeted campaigns.

Share

COinS
 

Leveraging Deep-learning and Field Experiment Response Heterogeneity to Enhance Customer Targeting Effectiveness

Firms seek to better understand heterogeneity in the customer response to marketing campaigns, which can boost customer targeting effectiveness. Motivated by the success of modern machine learning techniques, this paper presents a framework that leverages deep-learning algorithms and field experiment response heterogeneity to enhance customer targeting effectiveness. We recommend firms run a pilot randomized experiment and use the data to train various deep-learning models. By incorporating recurrent neural nets and deep perceptron nets, our optimal deep-learning model can capture both temporal and network effects in the purchase history, after addressing the common issues in most predictive models such as imbalanced training, data sparsity, temporality, and scalability. We then apply the learned optimal model to identify customer targets from the large amount of remaining customers with the highest predicted purchase probabilities. Our application with a large department store on a total of 2.8 million customers supports that optimal deep-learning models can identify higher-value customer targets and lead to better sales performance of marketing campaigns, compared to industry common practices of targeting by past purchase frequency or spending amount. We demonstrate that companies may achieve sub-optimal customer targeting not because they offer inferior campaign incentives, but because they leverage worse targeting rules and select low-value customer targets. The results inform managers that beyond gauging the causal impact of marketing interventions, data from field experiments can also be leveraged to identify high-value customer targets. Overall, deep-learning algorithms can be integrated with field experiment response heterogeneity to improve the effectiveness of targeted campaigns.