Multi-Output Insurance Purchase Prediction Pipeline

An end-to-end, leakage-free Machine Learning pipeline that predicts customer choices across 7 distinct insurance policy components (A through G) simultaneously. Built on the sequential customer shopping transaction logs of the Allstate Purchase Prediction Challenge (Kaggle), this project combines behavioral feature engineering with multi-output gradient boosting — and, crucially, evaluates every model against the honest baseline that dominates this dataset.

---

🚀 Project Overview

In the insurance industry, understanding and predicting customer behavior during the quote-generation funnel is critical for dynamic pricing and targeted cross-selling. Each customer in the data views a sequence of quotes (shopping points) before buying; the final purchase is recorded separately.

Instead of treating the 7 policy options as isolated problems, this repository implements a Multi-Output Classification framework, and instead of reporting accuracy in a vacuum, it measures every model against the "last quote" baseline — simply predicting that the customer buys exactly what they saw in their last viewed quote. That baseline already reaches ~92.5% per-option accuracy, so the real scientific question of this project is: can a model learn when a customer will deviate from their last quote?

📊 Key Performance Metrics (held-out test set, 19,402 customers)

• Naive Baseline (Last Quote): 92.48% average per-option accuracy — the honest reference point.
• Champion Model: Tuned XGBoost, 92.54% average per-option accuracy, beating the baseline on 6 of 7 components.
• Hardest Component: Target G — baseline 86.75% → model 87.04%. This is where customers change their minds most often, and where the model adds the most value.
• Strict Metric (All-7 Exact Match): ~68.9% for both the baseline and the best models — an honest reminder of how dominant the last-quote signal is (even winning Kaggle solutions barely moved this metric).
• Stability: 3-fold cross-validation on G: 0.866 ± 0.002; train-test gap ≤ 1pp on every target (no overfitting).

---

🛠️ Data Pipeline Architecture

The project is modularly structured into three sequential Jupyter Notebooks:

1. 01_EDA.ipynb (Exploratory Data Analysis)

• Statistical profiling of transaction lengths, demographic variables, and price states.
• Correlation heatmap of the numeric features.
• Class distributions for all 7 multi-class target variables.
• The baseline discovery: how often the purchased option equals the last viewed quote (~92.5% per option, ~68.8% for the whole package) — the finding that shapes the entire evaluation strategy.

2. 02_Preprocessing.ipynb (The Feature Engineering Engine)

• Extracts the last_shopping_point (the final quote a user reviewed before checking out) as the current-state feature vector.
• Engineers historical aggregate features per customer journey: cost statistics (mean/min/max/std), price volatility, and cost change from first to last quote.
• Option dynamics features: for each option A-G, how many distinct values the customer viewed (_nunique) and how many times the option switched between consecutive quotes (_n_changes) — the signal that tells the model when to distrust the last quote.
• Deterministic missing-value fills only (C_previous/duration_previous → 0, car_value → 'Unknown'). Fitted transformations (median imputation, categorical encoding) are deliberately deferred to the modeling notebook so they can be learned from the train split only — no data leakage.

3. 03_Modeling.ipynb (Multi-Output Framework & Honest Evaluation)

• Train/test split first, then train-only fitting of risk_factor median imputation and state/car_value encodings.
• Trains and benchmarks Random Forest, XGBoost, and CatBoost inside MultiOutputClassifier wrappers, side by side with the naive baseline.
• Hyperparameter tuning with RandomizedSearchCV on the hardest target (G), 3-fold cross-validation, and a per-target overfitting check.
• Deep dive on target G: classification report (macro F1), normalized confusion matrix, baseline-vs-model comparison split by customer history stability, feature importance, and SHAP interpretability.
• Inference demo: predicting the full 7-option package for an unseen customer and decoding labels back to the original scale.

---

📈 Model Performance Benchmark (Real Results)

Test set accuracy across all 7 insurance options, including the baseline every model must beat:

Target Component	Baseline: Last Quote	Random Forest	XGBoost 🏆	CatBoost
A	0.9272	0.9274	0.9276	0.9275
B	0.9353	0.9353	0.9353	0.9352
C	0.9316	0.9312	0.9320	0.9317
D	0.9502	0.9502	0.9504	0.9503
E	0.9362	0.9362	0.9363	0.9362
F	0.9260	0.9245	0.9260	0.9260
G	0.8675	0.8651	0.8701	0.8682
📊 OVERALL AVERAGE	0.9248	0.9243	0.9254	0.9250
🎯 ALL-7 EXACT MATCH	0.6889	0.6842	0.6884	0.6888

The tuned XGBoost (n_estimators=150, max_depth=6, learning_rate=0.1) keeps the 0.9254 overall average and raises G to 0.8704.

Business Insight: The margins over the baseline are small in absolute terms — and that is the honest, well-known reality of this dataset. The model's real value is concentrated where it matters: on target G and on customers with unstable shopping histories (baseline 83.8% → model 84.2%), exactly the customers a recommendation system needs help with. For everyone else, "show them their last quote" is already near-optimal, and the model correctly learns to reproduce it.

---

💻 Tech Stack & Frameworks

• Language: Python 3.10+
• Core ML Libraries: scikit-learn, xgboost, catboost
• Data Manipulation: pandas, numpy
• Visualization: matplotlib, seaborn
• Interpretability: shap

---

⚙️ How To Run This Project

1. Clone the Repository:

   git clone https://github.com/Khayal07/Insurance-Product-Purchase-Prediction-ML.git
   cd Insurance-Product-Purchase-Prediction-ML

2. Install Dependencies:

   pip install -r requirements.txt

3. Execute the Pipeline (order matters):

   • Run 01_EDA.ipynb to analyze the distributions and the baseline.
   • Run 02_Preprocessing.ipynb to build preprocessed_features.csv / preprocessed_targets.csv.
   • Run 03_Modeling.ipynb to train models, tune, evaluate, and view all visualizations.
   • Note: 03 must always be run after 02 — the fitted preprocessing steps (imputation, encoding) live in 03 by design.

4. Future Improvements

   • Sequence models (e.g., customer-journey transformers) to squeeze more signal out of the quote-order dynamics.
   • Joint prediction of the 7 options (e.g., classifier chains) to directly optimize the all-7 exact match metric.
   • API Deployment: wrap the winning Multi-Output framework with FastAPI to serve real-time multi-label vector predictions (the inference demo in 03 is the blueprint).