Predict and Prevent Customer Churn

This skill provides expertise in churn prediction modeling, specializing in reliable, interpretable models that generate practical business insights, covering feature engineering, model selection, evaluation metrics, and translating predictions into retention strategies. Core principles start with precisely defining churn per business context - for SaaS, no login for 30+ days or subscription cancellation; for telecom, contract termination or 90+ days of inactivity; for retail, no purchases for 12+ months. Time-aware feature engineering builds features respecting temporal relationships, using an observation window (e.g. 90 days of behavior) to predict a future window (e.g. the next 30 days), while carefully avoiding data leakage by ensuring features use only historical data. Class imbalance handling addresses the reality that churn typically affects only 5-20% of customers, using stratified sampling, SMOTE, or class weights rather than naive oversampling, and focusing evaluation on precision-recall metrics rather than raw accuracy.

The feature engineering framework computes a comprehensive feature set per customer: recency features (days since last login/purchase), frequency features (login/purchase/support-ticket frequency normalized by window length), monetary features (total spend, average order value, and a spend trend computed via linear regression slope), engagement features (feature-usage breadth, average session duration, bounce rate), and lifecycle features (customer age in days, and a tenure bucket categorizing customers as new/growing/mature/veteran). Model selection and training compares multiple algorithms - Logistic Regression, Random Forest, XGBoost, and Gradient Boosting - using a time-aware train/test split to prevent data leakage, SMOTE to balance the training set, and class-weight or scale_pos_weight parameters tuned to the actual class imbalance ratio. Evaluation metrics and business impact focus on precision-recall curves and ROC-AUC rather than accuracy, translating model output into business-relevant retention insights.

Use this skill when building a churn prediction model that needs a business-precise churn definition, leakage-free time-aware features, and proper handling of severe class imbalance. It is well suited to subscription, telecom, or retail businesses with historical customer behavior data to train on. It is not meant for businesses with too little historical data to establish a meaningful observation window, or for churn definitions so ambiguous that a precise business definition can't be agreed on first.

Input: historical customer behavior data (logins, purchases, support tickets, revenue) and a business-precise churn definition.

Output: an engineered feature set, trained and compared churn models, and business-focused evaluation metrics. Example time-aware feature engineering pattern:

def create_churn_features(df, observation_end_date, window_days=90):
    observation_start = observation_end_date - timedelta(days=window_days)
    features = {
        'days_since_last_login': (observation_end_date - df.groupby('customer_id')['last_login_date'].max()).dt.days,
        'login_frequency': df.groupby('customer_id')['login_count'].sum() / window_days,
    }
    return pd.DataFrame(features)

Works with Python's scikit-learn, XGBoost, and imbalanced-learn (SMOTE) for model training, and pandas/numpy for time-aware feature engineering.

Data scientists building customer churn models for subscription, telecom, or retail businesses, and teams that need leakage-free, business-precise churn predictions rather than a naive classification pass.