Customer churn, often referred to as customer attrition, is the process by which customers stop doing business with a company. Churn is a critical metric for businesses as it directly impacts revenue. When customers leave, they stop contributing to the company’s revenue, and lead to increased costs associated with acquiring new customers to replace them.Â
High churn rates can indicate dissatisfaction with a product or service, and understanding the reasons behind churn is essential for businesses to improve their offerings and customer retention strategies. In the financial services sector, customer churn can have profound implications. The industry is characterized by high customer acquisition costs and significant lifetime value per customer.Â
Therefore, understanding why customers leave is important. Explaining customer churn is crucial for financial services because it allows businesses to pinpoint the exact reasons for churn, whether due to poor customer service, high fees, or better offers from competitors. This insight enables targeted interventions to retain valuable customers and reduce churn rates.
In this blog, we’ll look at how to apply Generative AI on top of predictive ML models to enhance explainability. Using Large Language Models (LLMs) on Snowflake AI Data Cloud, we’ll extract detailed natural-language descriptions to help business associates understand complex quantitative predictions.
How Churn is Analyzed and Monitored
Churn analytics involves using data analysis techniques to understand and predict customer churn. Various methods include:
Descriptive Analytics: Summarizing historical data helps understand the current churn state.
Diagnostic Analytics: This type of analytics goes a step further to explain why churn is happening.
Predictive Analytics: Predictive models use historical data to predict future churn. Techniques include machine learning algorithms such as logistic regression, decision trees, and neural networks.
Prescriptive Analytics: This analytics type suggests actions to prevent churn by leveraging insights from predictive models.
Machine learning models are often employed to predict churn. One effective method for understanding these models is SHAP (SHapley Additive exPlanations). SHAP provides a way to explain the output of any machine learning model (shown below in Figure 1), assigning each feature impact for a particular prediction. Although Figure 1 is a useful data science plot, this plot might be confusing to a wider audience: which is why we are proposing a Generative AI solution using Snowflake Cortex.Â
Snowflake Cortex is a powerful tool for building automated pipelines for churn analysis. It enables businesses to integrate data from various sources, apply machine learning models, and generate actionable insights. With Cortex, companies can automate the process of predicting churn, understanding its reasons, and implementing interventions to reduce it.
Explaining Customer Churn with Snowflake Cortex
Consider a churn prediction model used by Snowflake Cortex. We use Kaggle’s Bank Customer Churn Dataset with a Random Forest model in this example. After the model runs, it generates SHAP values to explain its predictions. For instance, if the SHAP output indicates that the customer’s age and the number of products used lead to churn, Cortex can provide detailed explanations of these findings.
Cortex can leverage the importance of SHAP features to delve into the reasons behind churn and provide comprehensive insights by combining SHAP explanations with other churn analytics. For example, it could show that high fees, frequent account issues, and lack of customer support are the primary drivers of churn for a particular segment of customers. Below in Figure 2 we can see the application flow of a customer churn solution using the Snowflake framework.
This solution helps businesses understand the multifaceted nature of churn. To see this in action, we see a simple example of Snowflake Cortex using an SQL query that joins churn metrics and the top ten important model features. The output shows detailed analysis and insight using the mixtral-8x7b LLM.
SELECT SNOWFLAKE.CORTEX.COMPLETE('mixtral-8x7b', CONCAT('Critique this churn analysis with model feature importance and diagnose customer behavior: ', 'churn_rate = 0.2037, retention_rate = 0.7963, n_start = 10000, n_churn 2037, account_tenure_average = 5.01 days, account_tenure_median = 5.01 days, feature_age = 0.183, feature_num_products = 0.155, feature_is_active = 0.083, feature_credit_score_given_age = 0.052, feature_geography_germany = 0.048, feature_gender_female = 0.023, feature_account_balance = 0.020, feature_account_balance_salary_ratio = 0.018, feature_tenure_by_age = 0.016, feature_gender_male = 0.015'));
Based on the information provided, here is a critique of the churn analysis and a diagnosis of customer behavior:
1. Churn Rate and Retention Rate: The churn rate of 0.2037 indicates that one-fifth of customers leave the service within a certain period. The retention rate of 0.7963 indicates that the majority of customers (79.63%) are staying with the service. This is a relatively high churn rate, which suggests there may be customer satisfaction or loyalty issues.
2. Account Tenure: The average and median account tenure is 5.01 days, which is quite low. This suggests that customers may not use the service for long before they churn, and it could also indicate that the service cannot retain customers long.
3. Feature Importance: The feature with the highest importance is “”feature\_age”” with a value of 0.183. This suggests that the customer’s age is a significant factor in determining whether they will churn. Other important features include “”feature\_num\_products”” (0.155), “”feature\_is\_active”” (0.083), and “”feature\_credit\_score\_given\_age”” (0.052). These features suggest that the number of products a customer has, their level of activity, and their credit score (given their age) are also important factors in determining churn.
4. Diagnosis of Customer Behavior: Based on the information provided, older customers have more products, are more active, have higher credit scores (given their age), and are less likely to churn. On the other hand, younger customers have fewer products, are less active, and have lower credit scores (given their age) are more likely to churn. This suggests that the service may need to improve customer satisfaction and loyalty among younger, less active, and lower credit score customers to reduce churn. Additionally, the low account tenure suggests that the service may need to find ways to encourage customers to use the service for longer periods.
Cortex can also suggest targeted marketing interventions based on SHAP values and churn analytics insights. These interventions can be tailored to address the specific reasons behind churn, thereby improving customer retention. The next example shows possible market interventions to decrease customer churn.
SELECT SNOWFLAKE.CORTEX.COMPLETE('mixtral-8x7b', CONCAT('Given this churn analysis with model feature importance suggest 5 ranked market interventions to decrease churn: ', 'churn_rate = 0.2037, retention_rate = 0.7963, n_start = 10000, n_churn 2037, account_tenure_average = 5.01 days, account_tenure_median = 5.01 days, feature_age = 0.183, feature_num_products = 0.155, feature_is_active = 0.083, feature_credit_score_given_age = 0.052, feature_geography_germany = 0.048, feature_gender_female = 0.023, feature_account_balance = 0.020, feature_account_balance_salary_ratio = 0.018, feature_tenure_by_age = 0.016, feature_gender_male = 0.015'));
Based on the given churn analysis and model feature importance, here are five ranked market interventions to decrease churn:
1. Encourage Account Activity: The feature “”feature\_is\_active”” has a high importance score, indicating that encouraging customers to be more active on the platform could help decrease churn. This could be achieved through targeted marketing campaigns, personalized recommendations, or incentives for using the platform more frequently.
2. Tailor Offerings to Customer Age: The feature “”feature\_age”” has a high importance score, suggesting that tailoring offerings to different age groups could help decrease churn. This could involve creating targeted marketing campaigns, offering age-specific promotions, or developing new products or services that cater to specific age groups.
3. Promote Additional Products: The feature “”feature\_num\_products”” has a high importance score, indicating that promoting additional products to customers could help decrease churn. This could be achieved through cross-selling or upselling strategies or by offering discounts or incentives for purchasing multiple products.
4. Improve Credit Score Offerings for Older Customers: The feature “feature_credit_score_given_age” has a moderate importance score, suggesting that improving credit score offerings for older customers could help decrease churn. This could involve offering credit score improvement programs, credit counseling services, or credit-building products.
5. Target Marketing Efforts in Germany: The feature “feature_geography_germany” has a moderate importance score, indicating that targeting marketing efforts in Germany could help decrease churn. This could involve localizing marketing campaigns, partnering with local influencers or brands, or offering products or services tailored to the German market.
It’s important to note that these interventions should be tested and evaluated to determine their effectiveness in reducing churn. Additionally, it may be beneficial to consider other factors beyond the top-ranked features, as there may be interactions or complex relationships between different variables that could impact churn.
Conclusion
Understanding and managing customer churn is crucial for the financial services industry. By leveraging churn analytics, SHAP feature importance, and tools like Snowflake Cortex, businesses can gain valuable insights into the reasons behind churn and take proactive steps to address them.Â
Automated pipelines powered by Cortex provide accurate churn predictions and offer detailed explanations and actionable interventions. By addressing the root causes of churn, financial services can enhance customer satisfaction, reduce attrition rates, and ultimately improve their bottom line.
If you’re interested in taking your data insights to the next level, schedule a consultation with one of our architects at phData to explore how we can empower your data and AI journey.
