Customer Churn Prediction — Classification Project
Telecommunication companies, such as Vodafone Corporation, face a significant challenge with customer churn, where customers cease using their services. To address this issue effectively, it’s crucial to anticipate which customers are at risk of churning and implement proactive retention strategies. Leveraging machine learning models can provide a solution by predicting potential churners based on various factors, including usage patterns, payment history, and demographic data.
For the purpose of this article, we will focus more on the conceptual and strategic aspects rather than delving deep into technical details. Readers who are keen on exploring the technical aspects can find the detailed code and methodologies on my GitHub repository.
Project Overview
In this project, we analyze customer churn within Vodafone Corporation using the provided data. Our objective is to identify the key factors influencing customer churn and develop a predictive model to pinpoint potential churners. The methodology followed adheres to the CRISP-DM framework, ensuring a structured and systematic approach to our analysis.
Project Hypothesis
Our hypothesis aims to test the churn rate differences between customers with tenures of 0–36 months and those with tenures of 36 months and beyond. We hypothesize that after the initial 3 years, the rate at which customers churn decreases significantly. Additionally, we believe that by identifying strategies to reduce churn within the first 3 years, we can effectively manage and control customer attrition.
Our hypothesis testing led to the rejection of the null hypothesis, revealing a significant difference in churn rates between customers with tenures of the first 36 months and those with tenures beyond 36 months. Upon further analysis, we found that the churn rate was higher within the initial 3 years compared to tenures beyond 3 years. This difference is visualized in the pie charts below.
We further observed that the likelihood of churning decreases as tenure increases. This trend is clearly depicted in the line graph below.
Therefore, considering that churn rates peak between 0 and 20 months of tenure, it’s essential to adopt proactive measures aimed at engaging and retaining customers during this initial phase. By doing so, customers are more likely to reach a tenure of around 30 months, where the churn rate stabilizes. This stabilization signifies a critical period in customer relationships, presenting valuable opportunities for implementing targeted retention strategies.
Analytical Questions
In this chapter, we will explore various factors that influence churn rates. Our analysis will be guided by the following analytical questions:
1. How does method of payment affect churning?
Payment methods are nearly equally distributed among customers, with the Electronic Check mode being slightly more popular especially among churners. While Electronic Check has a slight edge as the largest customer segment compared to other payment methods, the differences are minimal. Despite this, the choice of payment method significantly influences churn rates. Specifically, customers using Electronic Check show a higher likelihood to churn compared to those opting for alternative payment methods. The graph below illustrates the contribution of each payment method to customer attrition.
2. Does the type of contract affect rate at which customers leave?
Most churners are enrolled in the Month-to-Month contract, making up 55.6% of the churners, and this contract also has the highest number of customers overall. In contrast, the Two-Year contract has the lowest churn rate, with only 2.4% (30 out of 1,243) of customers churning. The One-Year contract falls in between, being the second most churning contract but at a significant distance from the Month-to-Month contract, accounting for 12.1% of churners. Clearly, a discernible trend emerges: the longer the contract duration, the less likely the customer is to churn.
3. What is the impact of monthly charges to customer churning?
In our analysis, we found that the overall average monthly charge for customers is $65.09. Interestingly, churners have a higher average monthly charge of $75.2 compared to non-churners, whose average stands at $61.4. This suggests that monthly charges play a significant role in influencing churn rates. Specifically, customers with monthly charges exceeding the overall average are more susceptible to churn than those with charges below the average. Understanding this correlation between monthly charges and churn rates can guide businesses in developing targeted strategies to retain customers and optimize revenue.
4. Does presence of a partner affect likelyhood of churn?
Our analysis reveals a notable difference in churn rates based on customers’ partnership status. Specifically, 32.5% of customers without partners churned, whereas only 20% of customers with partners chose to leave. Consequently, it’s evident that customers without partners are more inclined to churn compared to their counterparts who have partners. This insight underscores the importance of considering relationship status when devising retention strategies, as it can significantly impact customer loyalty and retention rates.
5. What is the relationship of type of internet service and churning?
Our analysis indicates a significant correlation between the type of internet service and churn rates. Specifically, customers with fiber optic internet service exhibit a higher likelihood of churning compared to those using DSL or having no internet service. This observation suggests that the type of internet service could be a determining factor in customer satisfaction and, consequently, retention rates. Understanding these nuances can help businesses tailor their offerings and improve customer satisfaction to mitigate churn effectively.
6. Which gender is churning at a higher rate?
Our analysis indicates that gender does not appear to be a significant factor influencing churn rates. Thus, we can conclude that gender is not a strong predictor for churn. This insight suggests that other factors likely play a more critical role in customer retention, and the company should focus on identifying and addressing those factors to reduce churn effectively.
Churn Prediction Model
Our goal is simple: predict which customers might leave, so we can keep them. Missing a churn prediction means missing a chance to keep a customer and losing potential revenue. When we miss predicting a churn, it hits our bottom line. Losing a customer not only means lost sales but also the higher cost of attracting a new one. It’s about more than just money; it’s about maintaining customer trust and satisfaction.
Choosing the Right Model
We chose the Logistic Regression model because it’s great at catching potential churners. While its overall accuracy is 80%, it excels in reducing false negatives — predicting that a customer will stay when they won’t. It’s okay if the model sometimes flags loyal customers for retention actions. While this might lead to some extra efforts, it’s better than missing a real churn risk.
Below is the confusion matrix representing the performance of our final best-performing model.
For a clearer grasp of the confusion matrix, I recommend reading Racheal Appiah’s blog post “The Confusion Matrix — Explained Like You Were Five”. It offers a straightforward explanation that’s perfect for understanding this essential tool in machine learning.
Conclusion
While this blog provides a high-level overview, I’ve omitted many technical details for brevity. If you’re interested in diving deeper into the technical aspects, I invite you to visit my GitHub repository. There, you’ll find comprehensive information on the model development process, including pipelines, hyperparameter tuning, and threshold selection, among other technical nuances not covered in this article.
