CARRE (Counterfactual Action Retrieval and Reason Evaluation) is a new framework for churn prescription. It operates in three stages: first, it retrieves a predefined catalog of retention actions. Second, it estimates model-predicted churn-risk changes using explicit feature transformations and assigns a cost-aware counterfactual score to each action. Finally, it generates a structured churn reason and a profile-grounded explanation for the selected action. The framework was evaluated on the IBM Telco Customer Churn dataset. Results show CARRE achieves a 79.8% greater mean model-predicted risk reduction than the plain SHAP baseline and an 80.4% greater reduction than the cost-controlled SHAP+Cost baseline across 313 high-risk test cases. Its cost-normalized efficiency is 10.5% higher than plain SHAP. Diagnosis-driven prompt refinement increased weak-label agreement from 79.4% to 90.4%. A deterministic audit found no contradictions among 66 verifiable profile claims. Retrieval ablations indicated that k=5 provides the best balance between candidate coverage and downstream reasoning agreement. This framework separates and jointly evaluates retrieval, model-based scoring, and language generation in a churn-prescription pipeline.
Source: https://arxiv.org/abs/2609.09766