How zdanovic/insurane_claim works
Project type and purpose
• This is an offline, two-stage machine-learning pipeline for insurance claim forecasting (initial reserve at FNOL and final loss modeling).
• It runs as a series of Jupyter notebooks, ingesting CSV data, performing EDA, training models, evaluating performance, and producing interpretability outputs.
File structure
• data/ (CSV data stores for policies, raw claims, merged data)
• notebooks/
– 01_preprocessing_eda.ipynb (data cleaning, feature engineering, EDA)
– 02_fnol_model.ipynb (train & compare GLM Tweedie and XGBoost for FNOL reserves)
– 03_final_loss_model.ipynb (train final loss models, ensemble stacking)
• README.md, .gitignore (repo metadata)
Extract insights from README
• Two sequential modeling stages with strict no-leakage rules
• Ensemble approach combining statistical (GLM Tweedie) and ML (XGBoost)
• Interpretability via SHAP and error-case analysis
• Targets: MAE, RMSE, R² across two forecast horizons
Main components
a. Data Storage
– Raw CSV files (policies, claims)
– Merged/processed dataset
b. Preprocessing & EDA Module
– Notebook 01: cleaning, transforms, rolling windows, encoding
c. FNOL Modeling Module
– Notebook 02: GLM Tweedie vs. XGBoost training & validation
d. Final Loss Modeling Module
– Notebook 03: ingest FNOL outputs + features, train ensemble (Stacking Regressor)
e. Evaluation & Interpretability
– Metrics calculation (MAE, RMSE, R²)
– SHAP feature-importance plots, error case visuals
f. Downstream Integration
– Exported model artifacts & reports for dashboards/audit pipelines
g. Execution Environment
– Jupyter Notebook server, Python 3.8+, key libs (pandas, scikit-learn, xgboost, shap, matplotlib)
Relationships and data flows
1. Raw Data → Preprocessing & EDA
2. Processed Data → FNOL Modeling
3. FNOL Predictions + Features → Final Loss Modeling
4. Model Outputs → Evaluation & Interpretability
5. Artifacts & Visuals → Downstream Integration
Architectural patterns & principles
• Two-stage sequential pipeline (FNOL → Final Loss)
• Modular notebooks as components
• Strict data-leakage controls
• Ensemble design pattern (stacking regressors)
• Interpretability baked into pipeline (SHAP)