The goal of this project is to build a machine learning system that predicts insurance agents who are likely to have a NILL month (i.e., zero new policies sold in the upcoming month). This allows the company to proactively engage and support underperforming agents before their productivity drops.
- ✅ Custom target engineering for future-month prediction
- 📊 Comprehensive EDA and correlation + mutual information analysis
- 🕒 Advanced time-based and windowed features to capture short-term behavior shifts
- 📉 Class imbalance handling using XGBoost with custom parameters (
scale_pos_weight,gpu_hist) - 💡 Actionable recommendations for at-risk agents based on predictive factors
├── Explanation Docs/
│ ├── Agent Performance Analysis.pdf
│ ├── Predict NILL Agents Report.pdf
│ └── README.md
│
├── Dashboard/
│ ├── PowerBI\_Dashboard.pbix
│ ├── Evidence\_Screenshots/
│ └── README.md
│
├── Notebooks/
│ ├── EDA.ipynb
│ ├── Model\_Building.ipynb
│ ├── Draft\_Notebooks/
│ └── README.md
│
├── Models/
│ ├── xgb\_model\_fold\_0.pkl
│ ├── ...
│ └── xgb\_model\_fold\_9.pkl
│ └── README.md
-
Explanation Docs/
Contains the core documentation of our approach — including agent behavior analysis and detailed model report. -
Dashboard/
Includes the Power BI dashboard and visual evidence to support our findings and predictions. -
Notebooks/
All Jupyter notebooks for data exploration, model development, and experimentation. -
Models/
Serialized XGBoost models trained on each fold of the dataset using stratified cross-validation.
Here are some visual examples from our Power BI dashboard:
- Focused on recall for minority class (NILL agents) to better serve business goals.
- Engineered features and trends improved model interpretability and robustness.
- Delivered both predictive insights and practical actions to re-engage underperforming agents.
Team Cognic AI
Feel free to reach out if you want to collaborate, ask questions, or learn more about our approach!