HEALTH-INSURANCE-CROSS-SELL-PREDICTION

📖 Introduction

Cross-selling to existing clients has been one of the primary methods of generating new revenue for many businesses. Identifying the potential buyers among existing clients can help a company plan its communication strategy accordingly, thereby optimizing its business model and increasing revenue. In this project, the goal is to use the Health Insurance Cross-Sell dataset to understand Vehicle Insurance Cross Sale Responses, apply machine learning techniques to identify Vehicle Insurance buyers among pre-existing policyholders and provide explanations from the best classifying model to understand factors affecting customer responses.

📖 EDA Observations and findings

The important observations identified during EDA were:

1. Most customers that purchased vehicle insurance had existing damages in their vehicles, were not previously insured and owned vehicles aged over a year.
2. It was also observed that most policyholders of the health insurance company come from three region codes: 28, 8 and 41.
3. The policy sales channels most preferred to reach out to the clients were channel numbers 152, 26 and 124

The obstacles identified in the EDA phase of the project were:

1. The highly imbalanced dataset, with only 12.25 per cent of positive responses observed.
2. Categorical attributes like Region Code and Policy Sales Channel contained over fifty categories.
3. Vehicle Age attribute categorized a vehicle’s age into three categories in the form of strings.
4. The numerical features, Annual Premium, contained a notable number of outliers.

📖 Feature Engineering

These obstacles were overcome in the Feature Engineering phase of the project, when:

1. Categories that occur less than 5 per cent of the time were binned together and recognized as ‘Rare’.
2. The age of the client’s vehicle was numerically encoded.
3. Outliers in Annual Premium were capped at their lower and upper quartiles.
4. Categorical features were One Hot Encoded in order to be interpreted as categories.

📖 Sampling Methods

After Feature Engineering, the dataset was scaled and split into training and testing sets. Three sampling techniques, namely:

• Tomek Links
• Synthetic Minority OversamplingTechnique (SMOTE), and
• SMOTE-EditedNearestNeighbors were performed on the training set in order to counter the Imbalanced Dataset.

📖 ML Models Evaluated

Eight models, namely

1. Logistic Regression
2. Gaussian Naive Bayes
3. Decision Trees
4. Gradient Boosting
5. CatBoosting
6. LightGBM
7. Bagging, and
8. Random Forest were tested

📖 Evaluation Metric

• F2 score was the preferred metric as it gave emphasis in minimizing false negatives and was relevant as it could help identify maximum a number of potential customers.

📖 Results

• Upon model deployment, tuning and evaluation, we found that the best performing model was the Gaussian Naive Bayes classifier trained on SMOTE sampled training set with an F2 score of 0.634614, precision of 27.58 per cent and recall of 94.02 per cent.

• The finalized model’s predictions were dependent on the customer’s previous insurance status and existing damages on the vehicle.

📖 Conclusions

• When the dataset was explored earlier in the project, approximately, one in every ten clients had a positive cross-sale response when they were approached for cross-sale (12.25 % success rate).
• From the confusion matrix, we were able to tell that the model had correctly identified 94.04 per cent of the positive responses with a success rate of 27.58 per cent in predicting positive responders.
• This means that, approximately, three out of every ten predicted buyers produced a positive response.

Therefore, the model not only helped identify a large part of the potential vehicle insurance buyers but had also increased the success rate of cross-sales, helping the company save a significant amount of time and resources by generating better leads.

📋 References

1. Christopher M. Bishop, “Pattern Recognition and Machine Learning”, Pg. 137-139
2. Zhi-Hua Zhou, “Ensemble Methods Foundations and Algorithms”, Pg. 57-58
3. John T. Hancock and Taghi M. Khoshgoftaar, “CatBoost for big data: An Interdisciplinary Review”
4. Essam Al Daoud, “Comparison between XGBoost, LightGBM and CatBoost Using a Home Credit Dataset”.
5. Jason Brownlee,“Imbalanced Classification with Python: Better Metrics, Balance Skewed Classes, Cost-Sensitive Learning”.

📋 Execution Instruction

The given IPython Notebook can be either downloaded to be run locally on Jupyter Notebook or on Google Colab via browser.

📜 Credits

• Project Done by Mahin Arvind Chanthira Sekaran
• Project Verified by Almabetter

☎ Contact