This project applies predictive analytics to improve customer acquisition strategies for The Insurance Company (TIC), using the CoIL Challenge 2000 dataset (5,822 customers × 86 variables).
By testing Logistic Regression and Random Forest, the study demonstrates how predictive modeling can:
- Improve customer targeting for mobile home insurance.
- Reduce marketing acquisition costs by prioritizing high-probability leads.
- Provide data-driven insights for long-term customer relationship management.
Key takeaway: Logistic Regression provided interpretability for deployment, while Random Forest delivered higher predictive accuracy by capturing complex interactions.
📄 Full report PDF → Improving-Customer-Targeting-Through-Predictive-Modeling.pdf
Business Challenge
- Traditional marketing = high cost + low conversion.
- TIC needed to identify which customers are most likely to purchase mobile home insurance policies.
Solution Approach
- Apply predictive modeling to customer demographic & product ownership data.
- Compare Logistic Regression (interpretable baseline) vs. Random Forest (higher complexity).
- Evaluate models on accuracy, precision, recall, F1-score, ROC-AUC.
- Source: CoIL Challenge 2000 (real-world business data).
- Size: 5,822 rows × 86 columns.
- Features:
- Demographic (age, household size, income, education).
- Product ownership (car, life, home policies).
- Purchasing power & socio-economic indicators.
- Target:
CARAVAN→ binary (owns mobile home insurance: 1 = yes, 0 = no).
- Converted categorical/factor variables.
- Split into 80/20 training vs. testing sets.
- Addressed severe class imbalance (only ~6% policyholders).
- Logistic Regression (GLM)
- Probability-based, interpretable, fast to deploy.
- Random Forest
- Ensemble method, improved accuracy, captured nonlinear relationships.
- Metrics: Accuracy, Precision, Recall, F1-score, ROC-AUC.
- Visuals: Decision trees, feature importance, confusion matrix, ROC curve.
-
Logistic Regression:
- Strong interpretability, useful for pilot deployment.
- High precision (78.1%), but low recall (7.1%) → missed many true buyers.
-
Random Forest:
- Better handling of complex feature interactions.
- Feature importance: Purchasing behavior (PBRAND), demographics (MOSTYPE), car/life insurance ownership (APERSAUT, PPERSAUT), and purchasing power (MKOOPKLA) drove predictions.
-
Business Value:
- Improved targeting efficiency → fewer wasted marketing efforts.
- Actionable insights for segmentation & cross-selling.
- Foundation for CRM integration and predictive lead scoring.
- Predictive Modeling: Logistic Regression & Random Forest.
- Data Wrangling: Handling imbalanced data, feature engineering, correlation analysis.
- Model Evaluation: Confusion matrices, precision-recall tradeoffs, ROC-AUC.
- Business Translation: Linking analytics results to marketing efficiency & acquisition cost reduction.
- Reproducibility: Structured process with R Markdown, version control in GitHub.
Environment: RStudio, R 4.2.1
Packages: stats, randomForest, caret, ggplot2
# Example: Logistic Regression
glm_model <- glm(CARAVAN ~ ., data = train, family = "binomial")
summary(glm_model)
# Example: Random Forest
library(randomForest)
rf_model <- randomForest(CARAVAN ~ ., data = train, ntree = 500, importance = TRUE)
varImpPlot(rf_model)- Severe class imbalance limited recall → explore SMOTE or cost-sensitive learning.
- Add hyperparameter tuning (grid search, Bayesian optimization).
- Deploy models with PMML for integration into CRM systems.
- Pilot test with TIC’s customer data before full-scale deployment.
Released under the MIT License. See LICENSE.
predictive-analytics, logistic-regression, random-forest, insurance, customer-targeting, machine-learning, data-science, coil-challenge, r, marketing-analytics