Customer Churn Prediction in Small and Medium Enterprises Using Gradient Boosting with RFM-Based Behavioral Features

The increasingly widespread use of digital marketing has played an important role in encouraging repeat purchases and maintaining customer engagement. Loyal customers are seen as one of the important factors for micro, small, and medium enterprises (MSMEs) in maintaining a presence in the market and building consumer loyalty in the long term [1]. Customer loyalty is considered to play a role in reducing promotional costs because customers tend to make repeat purchases. Technological developments today are increasingly used to analyze customer loyalty levels and predict repurchase behavior [2]. The use of data-driven analytics and the implementation of digital business models show that the modernization process in the MSME sector in Indonesia is ongoing. Data-driven analytics and digital business models are increasingly being used to understand market behavior and improve the effectiveness of marketing strategies. However, despite the increasingly massive development of digitalization, MSMEs still face significant challenges in predicting customer loyalty. Limited resources capable of analyzing data and the low utilization of predictive methods lead to inaccurate assessments [3]. This phenomenon results in less optimal strategic decision-making for MSMEs.

Customer loyalty is closely related to churn, which is the tendency of customers to end a relationship with a company or switch to another service provider within a certain period [4]. One of the problems that MSMEs still face is the limitation in accurately identifying churn rates. Various previous studies have examined the phenomenon of churn in telecommunications companies and credit card services to evaluate customer loyalty. Research on churn prediction in MSME digital marketing is still relatively limited. This research gap causes the data-driven decision-making process to not be utilized optimally and reduces the ability of MSMEs to develop more targeted and effective customer retention strategies.

One of the most commonly used methods to predict customer churn is the decision tree. This machine learning method models the decision-making process in the form of a tree structure based on predetermined rules [5]. Each branch represents the conditions derived from the predictor variable, while the terminal node shows the customer's classification results. Customers can be grouped into categories that have the potential to stop transacting and those that are expected to remain persistent.

Although widely used in customer churn predictions, the decision tree method still has a number of limitations. These models can leverage historical data to predict customer behavior, but their accuracy is generally lower than ensemble models such as Extreme Gradient Boosting (XGBoost) [6]. Overly simple tree structures tend to result in models that are prone to overfitting; their performance decreases when faced with new data that has not been studied before [7]. Churn prediction models are often evaluated using accuracy-based metrics, even though these measures do not always align with the main business goals [8]. Another limitation is the relatively low ability of decision trees to capture consumer behavior patterns as well as changes in customer preferences that occur over time. The characteristics of consumers in Indonesia are heterogeneous, making churn prediction more difficult if based only on one decision tree model. The implication is that predictive performance tends to decline when models have to deal with high variations in customer behavior.

The decision tree is a single model that is simple to interpret because the decision-making process is represented through logical rules arranged in a hierarchical manner. These characteristics allow decision trees to be used not only as a prediction tool, but also as a means to explore data patterns and understand customer behavior. The simplicity of such structures limits the model's ability to capture nonlinear relationships. This limitation has the potential to increase prediction errors and reduce the model's accuracy, resulting in less optimal churn analysis outcomes and potentially biased conclusions. Currently, a stronger analytical approach is needed to improve churn prediction performance compared to using a decision tree model alone.

A better approach as a solution to the limitations of a simple decision tree is gradient boosting. Each decision tree is trained sequentially to learn from and correct the prediction errors generated by the previous tree. This improvement process is carried out by directing the construction of new trees toward residual values through a gradient-based optimization approach. Gradient boosting produces a much stronger and more accurate model than a single decision tree by combining many simple decision trees.

Gradient boosting overcomes these limitations by repeatedly integrating many simple decision trees into a single unified model. Each decision tree built at the next stage can help reduce prediction errors [9]. These mechanisms result in a more stable and resilient model, especially when faced with heterogeneous customer behavior. Customer churn analysis in digital marketing with a gradient boosting approach tends to be ideal because it can increase the accuracy of predictions. This modeling method is suitable for MSMEs because it offers more reliable predictive capabilities than a simpler single decision tree model.

The gradient boosting algorithm was developed to improve convergence efficiency in predictive modeling. Based on previous studies, empirical evaluations have shown that gradient boosting is capable of achieving high predictive performance [10]. Gradient boosting tends to be less sensitive to parameter shrinkage. Gradient boosting produces more compact ensembles. This is achieved with a smaller number of trees without sacrificing predictive accuracy. Research on the application of gradient boosting in data-driven analytics among MSMEs is still relatively limited, especially in Indonesia. This is caused by resource limitations for conducting research, particularly in the MSME sector. Further research is needed to assess customer churn among MSMEs using gradient boosting to improve the accuracy of customer loyalty projections. Many MSMEs possess large volumes of digital marketing data but have not yet fully utilized advanced predictive algorithms. There are many advantages if insights can be generated as a basis for decision-making. This condition indicates a research gap related to the implementation and effectiveness of ensemble models for churn prediction. Based on this phenomenon, this study aims to analyze the performance of gradient boosting in predicting customer churn among MSMEs.

The main contribution of this study is to examine the effectiveness of gradient boosting in predicting customer churn among MSMEs. Gradient boosting is an approach that remains relatively underexplored. Previous studies have mostly focused on churn prediction in large-scale industries, such as telecommunications and banking services, because these sectors generally have data availability and analytical data infrastructure. MSMEs face limitations in both data availability and analytical capability. The application and effectiveness of advanced ensemble models such as the Gradient Boosted Tree (GBT) have not yet been widely studied in MSMEs. This study employs several behavioral variables based on the Recency, Frequency, and Monetary Value (RFM) framework [11], namely recency, which refers to the number of days since the last purchase; frequency, which indicates the number of transactions; and monetary value, which represents total customer spending. The combination of these variables is used to generate reliable predictions through an ensemble learning approach.

This study analyzes descriptive data consisting of total purchase, frequency, and days since last purchase, as presented in Table 2.

Table 2. Descriptive analysis

Table 2 shows that three main variables, namely transaction value (total purchase), spending intensity (frequency), and recency (days since last purchase), are used to identify customer behavioral profiles. The average customer expenditure, expressed in Indonesian Rupiah (IDR), is IDR 12,283,323, with a very wide range, from IDR 2,141,000 to IDR 25,180,000. This condition indicates a high degree of heterogeneity in purchasing power among different customer segments.

Based on data obtained from MSMEs, the average customer purchase frequency was recorded at 10.31 transactions. The median value indicates that half of the total customers had more than 11 transactions. In the days since last purchase variable, an average of 77.38 days was obtained, with a maximum value of 160 days. The high recency value in the third quartile, which is 109 days, indicates a group of customers who have not purchased for a long time. This condition can be a signal of increasing customer potential to stop transacting (customer churn). The analysis process begins with the data selection stage using the Column Filter feature to ensure that only specific attributes are used in the modelling. The target churn variable, which was originally an integer data type, is converted into a string so that it can be recognized as a class label in the classification process.

Customer churn is an important challenge for business sustainability and therefore requires careful analysis. MSMEs require an accurate predictive model to support effective customer retention strategies. Heatmap analysis plays an important role in descriptive churn prediction. A heatmap depicts customer attributes and purchasing behavior as visual patterns. This is used to describe variables and provide information about customer characteristics and purchasing activity patterns. A heatmap plays an important role in data analysis by visually representing patterns and intensity levels [15]. Heatmaps in predictive modeling help reveal patterns and relationships that may be less apparent when examining numerical summaries. Heatmap analysis aims to improve model interpretability.

Figure 2 shows the heatmap visualization that reveals a loyalty threshold at the 10th transaction frequency. The heatmap applies an exploratory visualization tool to identify customer behavior patterns. The visualization results show a clear polarity: customers with low transaction frequencies, ranging from 1–9 transactions, have a greater likelihood of experiencing churn. The heatmap visualization also shows that the highest concentration of risk is in the 5th transaction. These results indicate that customers who have exceeded the threshold of 10 transactions exhibit a stable retention rate. This finding illustrates that the initial phase is a very crucial period.

tu_pian_2.png

Figure 2. Heatmap

Based on the GBT algorithm in Table 3, the confusion matrix for 248 testing observations shows strong performance. From 50 customers who actually experienced churn, 42 customers were classified as True Positive (TP), and the other 8 customers were classified as False Negative (FN). Furthermore, from 198 the model predicted 176 customers were classified as True Negative (TN), and 22 customers were classified as False Positive (FP). These results had an accuracy rate of 87.9%.

Table 3. Confusion matrix

A recall value of 84% showed the model's ability to capture churn signals. The precision value of 65.6% reflected that the model was efficient in prioritizing targeted marketing interventions. An acceptable error rate in False Positive and False Negative classifications shows that the model does not experience overfitting. The model was able to generalize non-linear consumer behavior patterns. This model can also serve as a decision-support tool to help reduce customers’ churn risk and improve customers’ effectiveness retention programs.

Table 4 shows a comparison of gradient boosting with several classification algorithms to evaluate model effectiveness, namely logistic regression, decision tree, and random forest. The test results show that gradient boosting produced the best performance on the evaluation metrics. This model achieved an accuracy rate of 87.9%, a precision of 65.6%, a recall of 84.0%, and an area under the curve (AUC) value of 0.88. Gradient boosting showed a significant improvement in performance indicators compared to logistic regression and decision tree. The random forest test results produced competitive values with an accuracy of 86.2% and an AUC value of 0.85, but the results were still below the performance of gradient boosting. These findings indicated that the boosting mechanism can improve the model's ability to identify customer behavior patterns.

Table 4. Performance comparison of churn prediction models

This study applied stratified 5-fold cross-validation, as shown in Table 5. The results show that the performance difference between the train-test split approach and cross-validation is relatively small. The validation shows 0.3% in accuracy, 0.5% in precision, 0.2% in recall, and 0.01 in AUC, respectively. This finding indicated that the model's performance is relatively consistent across various data splits. The test results also indicated that the model has a good level of stability and reliability in predicting customer churn.

Table 5. Cross-validation results

Table 6 shows that the model achieved an accuracy of 87.9%, indicating a high level of agreement between the predicted classifications and the actual customer statuses. These results indicate that the model has a good ability to distinguish between customers who have the potential to experience churn and customers who remain active. The model's ability to recognize customers who do not experience churn is also relatively good. This is reflected in the specificity value of 88.8%, which indicates that most loyal customers can be identified accurately. The model has proven to be effective in detecting churn risks and is able to maintain classification accuracy.

Table 6. Performance metrics summary

The precision value of 65.6% indicates that not all customers predicted to experience churn actually stop transacting. There are still a number of customers who remain active but are classified as churn customers. This condition is acceptable because the model is still able to provide predictive results. This classification error has the potential to cause retention programs to loyal customers. As a result, some of the resources allocated to retain customers are not used optimally. The high recall value indicates that the model is more effective in identifying customers who have potential to experience churn. The F1-score of 0.736 indicates a good balance between the model's ability to detect churn cases and its ability to reduce misclassification errors. These results indicate that the model performs well in terms of accuracy. Based on the test results, the model is reliable to support data-driven decision-making.

The study evaluates model performance using Cohen's Kappa statistics to measure classification reliability. The testing showed a Kappa value of 0.72. Based on the established scale [16], this value showed a substantial agreement level. This finding indicates that high model performance is not the result of class imbalance. Based on the performance testing, this reflected the algorithm's ability to extract valid predictive patterns from customer behavior variables.

This finding showed that gradient boosting has proven effective in predicting customer churn. The GBT model was shown to have achieved the highest accuracy rate of 87.9% among the evaluated models. An AUC value of 0.88 proved to significantly surpass the performance of a single decision tree. A precision rate of 65.6% shows that there are still several false positives, namely customers who are predicted to experience churn but actually remain loyal. This phenomenon is a common challenge in modeling nonlinear human behavior.

The model's ability to find customers who are truly at risk of leaving the company is a very important factor in predicting customer churn. Recall is often used as a key metric in model evaluation. The higher the recall value, the greater the company's chances of recognizing customers who show indications of stopping transactions. This ability is needed by the company before consumers really switch. This condition helps companies reduce risk of customer loss.

The precision value does not always have to be at a very high level to be able to provide benefits for MSMEs. In many cases, precision in the range of 60–70% is still considered adequate. These considerations arise because the costs incurred to provide promotions or incentives to customers who are not actually at risk of churn are generally relatively small. Failure to detect customers who are actually about to leave the company can result in greater losses resulting from lost opportunities to maintain revenue in the future.

The study findings show the advantages of an ensemble learning approach in modeling customer behavior. Compared to conventional classification methods, this approach is better able to capture complex and not always linear relationship patterns. The combination of several learning models allows the process of identifying customer characteristics to be carried out in more depth, so that the relationship between variables can be better understood. These capabilities contribute to improved prediction accuracy and result in more reliable models to support business decision-making.

The confusion matrix analysis results show that each form of misclassification has a different impact on business activities. False positive errors arise when customers who are actually still active are categorized as customers who have the potential to experience churn. This situation can cause companies to run promotional or retention programs for customers that do not require special intervention, resulting in less-than-optimal use of budgets and resources.

A false negative error occurs when a customer who is at risk of leaving company is not successfully detected by the model. Compared to false positives, this type of error has greater consequences. This is because companies lose the opportunity to take retention actions before customers stop transacting. If this condition occurs, the potential income that should still be maintained is at risk of being lost. Based on the test results, the relatively small number of false negatives suggests that the model is able to recognize the vast majority of customers who have a tendency to churn. These findings indicated that the model is effective enough to be applied in support of customer retention programs. Companies can easily determine which customer groups need to be prioritized in maintaining customer relationships.

The findings of this study are consistent with various machine learning studies that explain decision tree model limitations. The model is known to be quite sensitive to small changes in data and prone to overfitting. This happens when the decision tree grows too complex or when the distribution of classes in the dataset is unbalanced. This condition can reduce the model's ability to produce accurate predictions on new data. The GBT approach offers a solution to these weaknesses through a gradual learning process that is carried out iteratively. At each stage, a new tree is built to correct any remaining prediction errors from the previous tree. The model is able to examine the relationships between complex variables, including non-linear patterns. This is done without sacrificing the ability to generalize to data that has never been studied before. The results are in line with previous research that boosting-based methods, such as GBT and XGBoost, tend to produce lower error rates than using a single decision tree model. This advantage can be observed in consumer behavior data in service sector.

MSME transaction data can be analyzed periodically to produce a churn probability score for each customer. This information can be used by management and Customer Relationship Management (CRM) systems to identify customers who are at high risk of leaving the company. Based on the results, companies can develop more targeted retention programs, for example through personalized promotional offers. Special incentive strategies or other marketing approaches can be carried out according to customer characteristics. The developed model allows application to batch-based data processing environments and can be integrated with various business intelligence systems [17]. The integration provides support for faster, data-driven decision-making processes. Integrations help companies anticipate potential churn before it impacts a decline in customer numbers and revenue.

Gradient boosting is more suitable for structured tabular data because it does not require excessive computational resources. Gradient boosting tends to be particularly suitable when the number of input features is relatively limited. The gradient boosting model implementation is feasible in MSME environments that have limited resources. This model can be implemented in batch processing, where transaction data is processed to generate churn predictions.

The development of ensemble boosting models continues to experience significant progress in improving customer churn prediction accuracy. One example is the Enhanced Gradient Boosting Model (EGBM), which uses a Support Vector Machine with a Radial Basis Function kernel (SVM-RBF) as the base learner. The model applies an exponential loss function to improve the learning process in the Gradient Boosting Machine (GBM) [18]. Gradient boosting models are able to produce more accurate predictions than linear models, decision trees, and random forests. This finding strengthens the position of gradient boosting as an effective machine learning approach for capturing relationships in non-linear data [19].

Empirical studies show that XGBoost achieves the highest testing accuracy, while conventional GBM produces more stable generalization ability. Boosting methods produce a simpler and more interpretable leaf-view structure. This is done while still maintaining important information at a higher level. Random forest models tend to produce a denser and more redundant concept lattice [20].