AI-Driven Customer Churn Prediction and Retention Intelligence from Behavioral Signals

Abstract

Customer retention represents a strategic priority for organizations seeking to reduce revenue loss and strengthen long term customer relationships in increasingly competitive markets. This study utilizes two openly available customer analytics datasets drawn from the telecommunications and retail banking sectors to examine engagement patterns, transaction history, service usage, and account activity. Data preprocessing, behavioral feature engineering, exploratory analysis, and class imbalance treatment prepare a combined sample of 17,032 customer records for predictive modeling. Six behavioral indicators, recency, frequency, monetary value, engagement decline, complaint intensity, and inactivity duration, are engineered from the raw fields available in each dataset and capture early signs of customer disengagement across sectors. The study benchmarks five predictive models, Logistic Regression, Random Forest, Gradient Boosting, XGBoost, and LightGBM, for churn classification performance under a unified evaluation protocol. Model interpretability analysis relies on SHAP values to identify the primary behavioral drivers of customer departure, and k-means clustering profiles at-risk customer segments to support targeted retention strategies. Results show that Gradient Boosting achieves the strongest overall balance of performance, reaching an F1 score of 0.55 and an area under the curve of 0.80, while XGBoost records the highest raw accuracy at 74.2 percent, with every benchmarked model falling in a modest 66 to 74 percent accuracy range. Frequency and engagement decline emerge as the dominant behavioral predictors of churn, ahead of monetary value, inactivity duration, complaint intensity, and recency. Clustering analysis identifies two behaviorally distinct customer segments, a large lower-risk Stable Loyalists group and a smaller higher-risk At-Risk Decliners group, each warranting a differentiated retention approach. The novelty of the study lies in combining real, openly accessible cross sector behavioral data with interpretable ensemble modeling and segment level retention profiling within a single transparent analytical framework.