Predictive Churn Model with High-Risk Customer Identification for Insurance

This code implements an end-to-end solution to analyze and predict customer churn (cancellation) in the insurance industry using machine learning. The workflow includes data exploration, preprocessing, modeling, and actionable insights generation.

---

Project Stages

1. Exploratory Data Analysis (EDA)

• Generates statistics on the distribution of churn in the dataset.
• Calculates correlations between numerical variables and churn rates.
• Identifies initial patterns to guide modeling decisions.

2. Data Preprocessing

• Encoding: Converts categorical variables (gender, location, policy_type, coverage_level) to numerical values using LabelEncoder.
• Feature Selection: Defines predictor variables such as:
  • age (idade)
  • gender (genero)
  • location (localizacao)
  • customer_tenure_years (tempo_cliente_anos)
  • policy_type (tipo_politica)
  • monthly_premium (premio_mensal)
  • coverage_level (nivel_cobertura)
  • claims_last_year (sinistros_ultimo_ano)
  • payment_delay_days (atraso_pagamento_dias)
  • customer_satisfaction (satisfacao_cliente)
  • customer_service_interactions (numero_interacoes_atendimento).
• Data Handling:
  • Imputes missing values with the median (SimpleImputer).
  • Normalizes data using StandardScaler.

3. Predictive Modeling

• Algorithm: Random Forest with adjustments for class imbalance (class_weight='balanced').
• Evaluation:
  • Metrics: Precision, recall, F1-score, confusion matrix, and AUC-ROC.
  • Strategy: Stratified data split (80% training / 20% testing).

4. High-Risk Customer Identification

• Classification: Lists the top 10 customers with the highest churn probability.
• Explainability:
  • Highlights critical drivers (e.g., monthly_premium above average, payment_delay_days).
  • Displays feature importance to explain model decisions.

---

Technologies and Methods

• Libraries:
  • Pandas and NumPy for data manipulation.
  • Scikit-learn for machine learning and preprocessing.
  • Seaborn/Matplotlib for visualization (implied by imports).
• Techniques:
  • Handling imbalanced data.
  • Model interpretability (rule-based explanations).

---

Model Output

1. Performance Report:

   • Detailed classification metrics (precision, recall, F1).
   • AUC-ROC score to evaluate model discrimination capability.

2. Priority Action List:

   • High-risk customers flagged for intervention, with specific reasons (e.g., low satisfaction, recent claims history).

---

Ideal for: Insurance companies aiming to reduce customer churn through data-driven strategies and machine learning.