Stochastic Boosting in Retail Banking: Predicting the "When" and "Why" of Customer Churn

TL;DR

Predicting customer attrition is a billion-dollar problem for banks. This study moves beyond simple "will they leave?" binary classification to "when will they leave?" over a 6-month horizon. By leveraging real-world data from 130,000 customers and utilizing Stochastic Boosting, the researchers achieved over 96% AUC in forecasting near-term churn, identifying external banking relationships as a critical "red flag" for customer loyalty.

Problem & Motivation: The Non-Contractual "Silent" Exit

In retail banking, there is rarely a "breakup" letter. Unlike a gym membership or a Netflix subscription, a bank customer simply stops swiping their card or transfers their balance elsewhere. This is known as a non-contractual setting, and it makes identifying churners incredibly difficult.

The authors identify two fatal flaws in prior work:

1. Imbalance Ignorance: Churners represent less than 1% of the population, leading many models to simply "guess" everyone stays to achieve high (but useless) accuracy.
2. Temporal Ambiguity: Knowing a customer might leave in the next year is too vague for a marketing team. To be actionable, a model must predict if the churn is imminent (1 month) or distant (6 months).

Methodology: Rolling Windows and Stochastic Boosting

The core innovation lies in the data structuring. Instead of a single snapshot, the researchers used Rolling Time Windows.

• The Framework: For a 24-month dataset, the model looks at months 1-6 (features) to predict month 7 (label). It then slides to months 2-7 to predict month 8, and so on.
• The Algorithms: Six models were tested: Random Forests (RF), SVM, Stochastic Boosting (SB), Logistic Regression (LR), CART, and MARS.
• The "Acid Test": To prove the model wasn't just identifying "unhappy people" but actually "when they leave," the authors tested the model on a dataset consisting only of churners, forcing the model to distinguish between "churched now" and "churned later."

Experiments & Results: Identifying the Leaders

The results were conclusive: Stochastic Boosting (SB) and Random Forests (RF) dominated the leaderboard.

Key Findings:

• Stochastic Boosting (SB) reached an AUC of 96.53% for 1-month horizons.
• As the time horizon increased (e.g., predicting 6 months out), accuracy naturally decayed but remained robust (SB AUC ~94.45%).
• Feature Importance:
  • 1-2 Months Horizon: Success depends on "Total value of bank products" and "Existence of cards in another bank."
  • 3-4 Months Horizon: Predictive power shifts to "Number of transactions" and "External mortgage loans."

Deep Insight: Why Does This Work?

The effectiveness of Stochastic Boosting here stems from its ability to handle non-linearities and rare events. By iteratively re-weighting misclassified samples, SB focuses on the difficult-to-predict churners that simpler models (like Logistic Regression) miss.

Furthermore, the survival analysis (Cox Regression) revealed a fascinating social dimension: Age and Education are strong stay-indicators, while Web-based Acquisition (opening an account online) actually correlates with higher loyalty, likely due to the convenience of the digital ecosystem reducing "friction" that causes churn.

Conclusion & Future Look

This research provides a production-ready blueprint for banks. By shifting from static survival curves to dynamic machine learning horizons, institutions can prioritize high-value customers who are most at risk of leaving in the immediate future.

Limitations: The study relies on a 2-year window; longer-term patterns (macroeconomic shifts) remain unexplored. Future work integrating "Next Best Product" recommendations could allow banks to not only predict churn but automatically present a tailor-made incentive to prevent it.