Multi-channel multi-model customer conversion prediction in the insurance domain

This study addresses the challenge of predicting customer conversion across multiple sales channels within the insurance domain, an area that remains relatively underexplored compared to the extensive literature on e-commerce and online customer conversion. To bridge this gap, we evaluate the predictive capabilities of six popular machine learning (ML) models in estimating the likelihood of existing customers subscribing to additional services within 30 days of receiving an offer. The dataset utilized incorporates demographic profiles, vehicle specifica- tions, portfolio compositions, and multi-channel engagement. The data exhibits varying degrees of class imbalance for customer conversion depending on the specific sales channel. Our findings indicate distinct customer behaviors across sales channels, requiring channel-specific predictive models to optimize customer acquisition and conversion prediction. Among the evaluated models, CatBoost demonstrates superior performance, achieving a balanced accuracy (BA) of 82.20% with an F1 score of 87.7% on digital sales data and 73.66% BA with an F1 score of 73.6% for physical sales. While Neural Networks perform competitively (82.10% BA in digital and 72.70% BA in physical channels), CatBoost consistently outperforms by a marginal yet notable margin. The observed 10% performance disparity between channels reflects fundamental differences in customer behavior.