Date of Award

5-2025

Document Type

Project

Degree Name

Master of Science in Information Systems and Technology

Department

Information and Decision Sciences

First Reader/Committee Chair

Dorota Huizinga

Abstract

Addressing customer churn remains one of the greatest difficulties in the auto insurance sector, frequently caused by unmanaged complaints, service inaction, and insufficient communication regarding the customer’s needs. The goal of this research is to create a framework based on machine learning algorithms that predict whether a customer complaint will be confirmed by the regulatory authority and use this as a proxy to estimate customer churn in the absence of direct cancellation data. Through predicting complaint confirmation, which acts as an indicator of genuine service failure, the model helps in detecting customers with greater risk and aids in formulating individualized retention plans. The information used for this study includes auto insurance complaints submitted from the Texas Department of Insurance by constituents residing in Austin, Texas. Although the dataset is geographically restricted and missing some demographic information, it provides rich data pertaining to service-related problems and behavioral indicators of churn. The research addresses four key questions: Q1. Which service-related factors lead to confirming a complaint and how do they indicate customer dissatisfaction and possible churn risk? Q2. How accurately can a stacked machine learning model combining both XGBoost and MLP predict complaint confirmation outcomes, and how do their performance compare to individual models? Q3. How can the insights from the churn prediction models be translated into effective, data-driven retention strategies for the auto insurance companies? Q4. How can feature importance from the churn prediction model be used to support strategic decision-making and improve customer retention efforts? The findings and conclusions of the analysis are respectively as follows: For Q1. purpose, the most significant drivers of confirmed complaints which is a proxy of churn appear to be resolution time, complaint handling and reason for complaint. All these features indicate how important customer satisfaction and operational productivity service management are to customer loyalty. For Q2. a more sophisticated approach where the model comprised a stack of tree-based and neural network methods performed better than the individual models, exhibiting lower generalizability and predictive performance. This confirmed that hybrid models are less fragile when it comes to measuring complex customer behavior. With respect to Q3. This study used TF-IDF-based topic modeling to extract predominant themes from the complaint texts. These themes were associated with the scores for churn probability to develop targeted company retention strategies which incorporated redesigns of the workflows, escalation protocols, and enhancements in communications. Q4. The analysis carried out SHAP analysis to explain the conclusions of the models which enabled business executives to follow the trail of churn to services they failed to deliver. The use of these interpretations provided customers support teams with clear data to take operational actions based on changes customers wanted to see engineered in their service delivery processes. By adding demographic, transactional, and policy variables, future research might enhance the framework's efficiency in model performance and customization. Further work may include using A/B testing to assess the impact of implementing machine learning-informed retention campaigns within a live CRM environment.

Download

  Contact Author

Share

COinS