🤖 machine learning

A Rolling-Window Framework for Churn Prediction and Behavioral Driver Identification

This paper proposes a temporally explicit, rolling-window framework for churn prediction in non-contractual service environments that integrates feature-based and sequence-based learning to achieve robust, interpretable, and deployment-ready risk assessment, demonstrating high accuracy and stability on real-world data without requiring frequent model retraining.

Original authors: Muhammad Jawad Mufti, Omar Hammad, Haitham Saleh, Muqaddas Gull

Published 2026-06-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Jawad Mufti, Omar Hammad, Haitham Saleh, Muqaddas Gull

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you run a busy, on-demand car wash service. You don't have contracts; people just show up when they need a wash. The big problem? You never know when a customer is about to stop coming until they've already stopped. They don't send a "I'm quitting" letter; they just go silent.

This paper is about building a smart radar system to spot those silent customers before they disappear, using a method that looks at their behavior like a rolling movie reel rather than a single frozen photo.

Here is how the authors solved this, explained simply:

1. The Problem with Old Methods: The "Yearbook Photo"

Most old ways of predicting churn are like looking at a customer's yearbook photo. They take all the data from a customer's entire life with the company, mash it into one big summary, and ask: "Will this person leave?"

The authors say this is flawed because:

It's static. It doesn't see that a customer was happy last month but has been silent for the last two weeks.
It's too slow. By the time the "yearbook" is updated, the customer might have already left.

2. The New Solution: The "Rolling Window"

Instead of a yearbook, the authors use a rolling window. Imagine a camera that slides forward one day at a time.

The Observation Window (The Past 30 Days): The camera looks at what the customer did in the last month. Did they book? Did they cancel? Did they pay?
The Evaluation Window (The Next 30 Days): The camera then looks at the future month to see if they actually stopped coming.

If the customer is silent in the "future" month, the system learns: "Ah, the behavior in the 'past' month predicted this silence."

The magic is that this camera rolls forward every single day. It doesn't wait for a month to end. As soon as a new day passes, the window slides, and the system re-evaluates the customer's risk. It's like checking the weather every hour instead of just once a season.

3. Two Different "Detectives"

The paper tested two different types of AI detectives to see who could spot the silent customers better:

Detective A (The Feature-Based Model): This detective looks at a summary report of the last 30 days. It counts things like "total bookings," "average cost," and "how many days since the last visit." It's like a very organized accountant who spots patterns in the numbers.
- Result: This detective was very stable and accurate (87.6% accuracy). It was great at seeing the big picture.
Detective B (The Sequence-Based Model): This detective looks at the story of the last 30 days. It doesn't just count the bookings; it looks at the order they happened. Did the customer book every day and then suddenly stop? Or did they book once and then wait? It's like a movie director who notices the pacing of the story.
- Result: This detective was a master at catching the people who were about to leave (96% recall). It was very sensitive to the "story" of their silence, though it sometimes got a little too nervous and thought some active people were leaving (false alarms).

4. The "Future Test" (Time Travel)

To make sure these detectives weren't just memorizing the past, the authors did a special test. They trained the models on data from 2024 and then sent them to predict behavior in June and July 2025 (a future time they hadn't seen before).

The Result: Both detectives still worked well without needing to be retrained. The "Accountant" (Feature model) stayed very steady, while the "Director" (Sequence model) kept catching the leavers, even though the world had changed slightly. This proves the system is robust enough for real-world use.

5. Why Do People Leave? (The "Why" Factor)

The paper didn't just predict who would leave; it explained why. Using a tool called SHAP (which acts like a magnifying glass for the AI's brain), they found the top reasons customers were at risk:

The Trend: If the number of bookings is going down, that's a huge red flag.
The Silence: The longer it's been since the last booking, the higher the risk.
The Weekend: If they usually book on weekends and suddenly stop, that's a sign.

The Big Takeaway

The main lesson of this paper isn't about using the most complex, expensive AI. It's about how you frame the problem.

The authors found that if you design your system to look at time correctly (using these rolling windows) and separate the "past behavior" from the "future outcome" clearly, you get a much better result than just using a fancy, complicated model on bad data.

In short: Don't just look at the customer's whole history; watch their recent movie reel, day by day, to catch the moment they start to fade away.

Technical Summary: A Rolling-Window Framework for Churn Prediction and Behavioral Driver Identification

Problem Definition
Customer churn prediction in non-contractual, pay-per-use service environments presents unique challenges distinct from contractual models. In these settings, disengagement is not an explicit event (like contract termination) but must be inferred from sustained behavioral inactivity. Existing literature often relies on static, single-point representations of customer behavior or simplified temporal assumptions. These approaches frequently aggregate data into coarse-grained summaries (e.g., lifetime RFM or quarterly statistics), obscuring short-term dynamics that precede churn. Furthermore, many studies lack explicit temporal problem formulations, with undefined observation windows and prediction horizons, limiting the interpretability and operational readiness of models for continuous risk assessment.

Methodology
The study proposes a temporally explicit churn prediction framework designed specifically for non-contractual service platforms. The core of the methodology involves a rolling-window instance construction strategy that transforms raw event-level data into a time-indexed panel of behavioral instances.

Temporal Formulation: For each customer, the framework defines a fixed 30-day behavioral observation window ( $W_b$ ) followed by a 30-day churn evaluation window ( $W_c$ ). Churn is operationalized as the absence of qualifying activity during the evaluation window. This ensures a clear temporal separation between behavioral evidence (inputs) and churn outcomes (labels).
Instance Generation: Windows advance with a daily stride. To avoid redundancy, consecutive windows yielding identical behavioral summaries are collapsed; new instances are generated only when a behavioral change (e.g., a new booking) occurs within the observation window.
Feature Engineering: The framework extracts 40 engineered features from each observation window, categorized into:
- Booking Activity: Volume and success/cancellation rates.
- Monetary Behavior: Spending patterns and wallet usage (derived only from successful bookings).
- Spatial & Service Characteristics: Geographic dispersion and service type preferences.
- Temporal Context: Recency indicators and calendar attributes.
- Categorical Preferences: Dominant payment methods and sources.
- Trend-Based Indicators: Linear trends in engagement and spending to capture short-term dynamics.
Modeling Approaches: Two complementary paradigms are evaluated within this unified temporal design:
1. Feature-Based Learning: An XGBoost (Gradient Boosting) model trained on aggregated feature vectors.
2. Sequence-Based Learning: A Long Short-Term Memory (LSTM) network trained on ordered sequences of window-level behavioral summaries to capture temporal dependencies.
Evaluation Protocol: Models are trained on historical data (2024) and evaluated on a temporally held-out future dataset (June–July 2025) without retraining to assess robustness against temporal shifts. Interpretability is achieved via gain-based feature importance and SHAP (Shapley Additive Explanations) analysis.

Key Results
The framework was evaluated on a real-world dataset from a non-contractual on-demand car wash service comprising over 100,000 customers and 400,000 booking events.

Predictive Performance:
- The XGBoost model achieved an accuracy of 87.6% and a ROC-AUC of 0.94 on the rolling-window test set, demonstrating balanced classification and strong stability.
- The LSTM model achieved a higher recall of 96.1% on the test set, indicating superior sensitivity in detecting temporal disengagement patterns, though with a slightly lower overall accuracy compared to XGBoost.
Temporal Generalizability: When applied to the unseen future dataset (2025) without retraining:
- XGBoost maintained an accuracy of 83.14% and ROC-AUC of 0.911, showing strong robustness to distributional shifts.
- LSTM maintained a high recall of 91.80% for churn users, though its ROC-AUC declined to 0.830.
Churn Drivers: Analysis identified trend_total_bookings, weekend_success_rate, total_bookings, and days_since_last_booking as the dominant global predictors. SHAP analysis confirmed that declining booking trends, prolonged inactivity, and sparse engagement significantly increase churn probability.

Significance and Claims
The paper argues that the primary contribution lies not in model complexity, but in the careful design of temporal framing. The study demonstrates that:

Explicit Temporal Separation: Defining clear observation and evaluation windows is critical for realistic churn modeling in non-contractual environments, preventing information leakage and ensuring prospective risk assessment.
Rolling-Window Necessity: Moving away from static snapshots to rolling, instance-level assessments allows for the detection of short-term behavioral dynamics that static aggregation obscures.
Interpretability and Deployment: The framework balances predictive performance with explainability. By linking feature importance to specific time windows, the model supports actionable decision-making, allowing providers to identify when and why a customer is at risk, rather than just if they are at risk.
Robustness over Complexity: The results suggest that well-structured temporal framing enables even lightweight models (like XGBoost) to achieve robust, deployment-ready performance, challenging the notion that deep learning architectures are strictly necessary for high-quality churn prediction in these contexts.

The study concludes that this framework provides a practical foundation for churn-oriented decision support in dynamic service environments, emphasizing that temporal rigor is as vital as algorithmic sophistication.