MASTER Assignment
FORECASTING CUSTOMER LIFETIME VALUE THROUGH RISK PREDICTION : AN EXPLAINABLE MACHINE LEARNING APPROACH FOR THE TELECOMMUNICATION INDUSTRY
Type : Master M-BIT
Period: Mar, 2023 - Aug, 2023
Student : Edo Belva Firmansyah, (Edo, Student M-BIT)
Date Final project: August 29, 2023
Supervisors:
Abstract:
As businesses focus more on understanding Customer Lifetime Value (CLV) to guide customer strategies, a gap has been identified: limited consideration of customer risk in CLV calculations, particularly in the telecommunications sector. This gap is more apparent as this sector has yet to explore risk-adjusted CLV predictions using machine learning (ML). This thesis aims to address this gap by predicting risk-adjusted CLV in the telecommunications industry using ML techniques. The study introduces a Risk-Adjusted Return (RAR) metric to incorporate customer risk into CLV calculations. Employing Design Science Research Methodology (DSRM) and Cross-Industry Standard Process for Data Mining (CRISP-DM), the study constructs ML models. Customer risk for RAR includes churn probability and beta value, adjusting the discount rate for risk. Four RAR calculation approaches are proposed, and ML models (Logistic Regression, XGBoost, CatBoost, and Random Forest) predict churn probability and RAR. Validation employs eXplainable AI (XAI) techniques like feature importance, SHAP global explanation, SHAP local explanation, and LIME. Results reveal statistical differences in RAR approaches, validating their distinctiveness. Churn model accuracy is 85%, and RAR models exhibit strong performance (R^2=0.92, MAPE≈20%). XGBoost excels in churn prediction, while CatBoost excels in RAR prediction. Prominent features for RAR include loyalty points, average revenue, revenue variability, churn probability, and beta value. These align with traditional RAR calculations, enhancing model validity. The study concludes that proposed RAR is statistically significant, and models exhibit robustness, with feature importance aligning with traditional methods. Future research could explore different risks, dissect revenue components for individual RAR, employ advanced ML algorithms, hyperparameter tuning, and further integrate ML in XAI.