CUSTOMER CHURN PREDICTION MODEL FOR TELECOM/FINTECH COMPANIES

Background to the Study

Churn prediction is binary classification problem where positive class churn minority. Telecom dataset includes features contract month-to-month one-year two-year tenure internet service fiber optic monthly charges total charges; banking includes credit score geography gender age tenure balance number of products has credit card active member estimated salary. Class imbalance characteristic non-churners substantially outnumber churners requires mitigation via class-weighted loss or Synthetic Minority Over-sampling Technique SMOTE generating synthetic minority examples. Logistic Regression provides interpretable coefficients log-odds Random Forest ensemble of decision trees reduces variance XGBoost gradient boosting optimized regularized tree ensemble often state-of-art for tabular data DNN fully connected network captures non-linear interactions but requires more data tuning.

Customer lifetime value estimate total net revenue expected over relationship derived from billing tenure balance product holding used to weight retention prioritisation capturing disproportionately large share of at-risk revenue within constrained budget versus blanket offers. For methodological references see scikit-learn ensemble methods documentation and XGBoost documentation and Telco Customer Churn dataset Kaggle. Related machine learning designs in machine learning project topics on SCHOLARNESTHUB.

Statement of the Problem

·         Telecommunications and fintech businesses frequently lack systematic quantified basis for identifying which customers at elevated churn risk resulting in reactive retention after disengagement or costly undifferentiated campaigns to full base regardless of individual risk.

·         Published churn studies frequently evaluate single sector most commonly telecommunications in isolation providing limited direct comparative evidence whether model architecture choices and feature importance generalise to other sectors such as fintech/banking.

·         Reported accuracy metrics frequently presented without translation into business-relevant terms such as revenue-capture efficiency of using risk scores to prioritise resource-constrained retention outreach programme.

·         Churn datasets characteristically exhibit class imbalance non-churning substantially outnumbering churning challenge addressed inconsistently or without explicit comparative evaluation of mitigation technique trade-offs.

·         There is limited published evidence directly comparing feature importance patterns across telecommunications and fintech churn datasets within single consistent study despite practical relevance to whether model or feature engineering transferable across sectors.

Aim and Objectives of the Study

Aim is to design implement and comparatively evaluate customer churn prediction model applicable to telecommunications and fintech business contexts further examining customer lifetime value implications of model-informed retention prioritisation.

·         Review existing customer churn prediction approaches and their application within telecommunications and fintech business contexts.

·         Design and implement data preprocessing and feature engineering pipeline suited to telecommunications and banking customer account data.

·         Design train and comparatively evaluate four classification algorithms Logistic Regression Random Forest XGBoost and Deep Neural Network for churn prediction on both telecommunications and banking dataset.

·         Conduct feature importance analysis to identify and compare most predictive churn drivers across two evaluated business sectors.

·         Evaluate class-imbalance mitigation techniques class weighting versus SMOTE oversampling and their effect on model performance.

·         Conduct supplementary customer lifetime value-weighted evaluation assessing revenue-capture efficiency of model-informed retention prioritisation relative to undifferentiated outreach strategy.

Research Questions

·         What predictive accuracy precision recall F1-score and AUC are achieved by Logistic Regression Random Forest XGBoost and DNN models for churn prediction on telecommunications and banking customer data?

·         Which customer account usage billing and demographic features are most predictive of churn in each of two evaluated sectors and do these feature importance patterns differ meaningfully between sectors?

·         How does class-weighted training compare to SMOTE oversampling in terms of resulting model performance on evaluated datasets?

·         What is revenue-capture efficiency of customer lifetime value-weighted model-informed retention prioritisation relative to undifferentiated retention outreach strategy?

Significance of the Study

Significant to telecommunications and fintech business analytics teams seeking empirically grounded guidance on churn prediction methodology selection and retention-resource prioritisation strategy showing XGBoost strongest telecom accuracy 80.8 percent AUC 0.847 banking 86.9 percent AUC 0.872 followed by Random Forest. Significant to broader customer analytics research community as methodologically transparent study directly comparing performance and feature importance across two distinct sectors within single consistent study. Significant to decision-makers evaluating integration of CLV information into retention allocation demonstrating CLV-weighted prioritisation captures disproportionately large share of at-risk revenue within constrained budget. Further analytics templates in churn prediction project topics on SCHOLARNESTHUB.

Scope of the Study

Limited to design implementation comparative evaluation using two publicly available benchmark datasets widely used Telco Customer Churn dataset 7,043 records churn 26.5 percent and retail banking churn dataset 10,000 records churn 20.4 percent rather than live proprietary data for reproducibility. Customer lifetime value-weighted evaluation uses illustrative CLV estimation derived from dataset available billing/transaction features monthly charges tenure for telecom estimated balance product holding for banking rather than business actual internally computed CLV figures not available within public datasets. Addresses binary churn/non-churn classification within fixed observation and prediction window as defined by each source dataset does not address churn timing prediction estimating when rather than whether customer will churn discussed as future work.

Limitations of the Study

·         Uses publicly available benchmark datasets rather than live proprietary business data genuine business deployment would need to validate approach against and likely retrain model on own specific customer data.

·         Customer lifetime value estimates used are illustrative derived from available dataset features using simplified estimation approach rather than validated actual CLV figures computed through business own internal financial methodology.

·         Addresses fixed-window binary churn classification task as defined by each source dataset labelling methodology churn timing prediction and ongoing rather than fixed-window risk monitoring outside scope.

·         Banking dataset used for fintech-sector comparison represents traditional retail banking rather than digital-native fintech platform specifically while broadly comparable in relevant feature categories findings should be interpreted as indicative of general financial-services churn patterns rather than validated specifically for digital-first fintech context.

Operational Definition of Terms

Customer Churn: Discontinuation of customer active relationship within defined observation period through explicit cancellation non-renewal sustained inactivity representing revenue threat.

Churn Prediction: Use of historical customer data and machine learning modelling to estimate individual likelihood of churning enabling proactive targeted retention.

Customer Lifetime Value (CLV): Estimate of total net revenue business can expect to derive from customer over duration relationship derived from billing tenure balance used for prioritisation.

Class Imbalance: Characteristic of labelled dataset where one outcome class churners substantially underrepresented relative to other 26.5 percent telecom 20.4 percent banking requiring weighted training.

SMOTE: Synthetic Minority Over-sampling Technique method generating synthetic minority-class training examples to address class imbalance evaluated as supplementary comparison to class weighting.

Retention Intervention: Targeted action offer outreach engagement campaign intended to reduce likelihood that identified at-risk customer churns prioritised via CLV-weighted evaluation capturing larger share of at-risk revenue.

Feature Importance: Measure of relative contribution of individual input feature to model predictive performance contract type tenure monthly charges most predictive telecom number of products held activity status age most predictive banking indicating different driver profiles.

Tenure: Length of time customer has held active account commonly used churn-predictive feature inversely related to churn risk longer tenure lower churn.

Short Conclusion

Results showed XGBoost strongest predictive performance both datasets telecom accuracy 80.8 percent F1 0.612 AUC 0.847 banking accuracy 86.9 percent F1 0.634 AUC 0.872 followed closely by Random Forest with Logistic Regression trailing but interpretable and DNN comparable not exceeding tree ensembles. Feature importance identified contract type tenure monthly charges most predictive telecom and number of products held account activity status age most predictive banking indicating meaningfully different churn driver profiles despite shared general approach. Class-weighted training applied across all models with SMOTE evaluated as supplementary. Supplementary CLV-weighted evaluation found prioritisation toward highest-value at-risk customers captures disproportionately large share of at-risk revenue within constrained retention-outreach budget versus undifferentiated outreach. Study concludes gradient boosting-based churn prediction combined with CLV-weighted retention prioritisation offers practically effective revenue-relevant approach to churn management for telecom fintech businesses while cautioning sector-specific differences indicate models should not be assumed directly transferable across sectors without retraining on sector-appropriate data. Implementation guides in telecom analytics project topics on SCHOLARNESTHUB.

Frequently Asked Questions

Q: What is customer churn prediction?

A: Use of machine learning classification on historical account usage billing engagement data to estimate likelihood individual customer will churn within future period enabling proactive retention.

Q: Which algorithms were evaluated and which performed best?

A: Logistic Regression baseline Random Forest XGBoost DNN; XGBoost strongest telecom accuracy 80.8 percent AUC 0.847 banking 86.9 percent AUC 0.872 followed closely by Random Forest.

Q: What datasets were used?

A: Publicly available Telco Customer Churn dataset 7,043 records churn 26.5 percent and retail banking churn dataset 10,000 records churn 20.4 percent with tenure usage billing demographic features.

Q: What features most predict churn in telecom versus banking?

A: Telecom contract type tenure monthly charges most predictive banking number of products held account activity status age most predictive indicating different driver profiles across sectors.

Q: How was class imbalance handled?

A: Class-weighted training applied across all models given moderate imbalance with SMOTE oversampling additionally evaluated as supplementary comparison.

Q: What is CLV-weighted retention prioritisation?

A: Prioritising retention outreach toward highest-value at-risk customers identified by model using illustrative lifetime value from billing tenure balance rather than treating all flagged uniformly capturing disproportionately large share of at-risk revenue within constrained budget.

Q: Can churn model be transferred across sectors without retraining?

A: No study concludes sector-specific feature importance differences indicate models should not be assumed directly transferable across telecom and fintech without retraining on sector-appropriate data.

Q: What metrics were used for evaluation?

A: Accuracy precision recall F1-score AUC with business-relevant revenue-capture efficiency of model-informed prioritisation versus undifferentiated outreach.

Q: What are limitations of using public benchmark datasets?

A: Public datasets not live proprietary data CLV estimates illustrative not actual business figures fixed-window binary classification not churn timing prediction banking dataset traditional retail not digital-native fintech.

Q: What practical recommendation for telecom fintech businesses?

A: Implement gradient boosting-based churn prediction XGBoost with class-weighted training combine risk scores with CLV to prioritise retention outreach toward high-value at-risk and retrain on own customer data for deployment.