Beyond Discrete‐Time Hazard Regression: Bayesian Regularization for Dynamic Churn Modeling

ABSTRACT Discrete‐time hazard regression is widely applied to firm‐level time‐to‐event analyses and increasingly to individual churn behaviors. While the discrete logit hazard model for panel data remains the workhorse in OM, both fixed‐effects and random‐effects specifications have key limitations. The Grassia(II)‐Geometric (G2G) model—allowing for dynamic churn and asymmetric heterogeneity—offers a compelling alternative for hazard regression and churn prediction. To address high‐dimensional data, we propose Bayesian sparsity modeling via spike‐and‐slab LASSO (SSL) that enables simultaneous regularization and inference—both crucial for explanatory hypothesis testing. Simulation studies show that even when data are generated from logit functions with additive heterogeneity, asymmetric heterogeneity undermines logit regression, whereas G2G‐SSL achieves stronger out‐of‐sample fit with lower false discovery rates. We further validate its predictive power in an empirical study on voluntary employee turnover at a Fortune 500 manufacturer. The proposed approach outperforms a range of advanced methods, improving both churn time prediction and cost efficiency for prescriptive workforce planning. Finally, we outline methodological extensions and provide a modeling procedure for discrete hazard regression analysis. By importing insights from marketing and Bayesian statistics, our work enables OM scholars to move beyond traditional hazard models and better capture real‐world heterogeneity across a broad range of contractual churn scenarios.