Predictive performance of count regression models versus machine learning techniques: A comparative analysis using an...

Area of Science:

• Insurance Analytics
• Statistical Modeling
• Machine Learning

Background:

• Accurate automobile insurance claim frequency forecasting is vital for risk assessment and pricing.
• Traditional Poisson models struggle with zero-inflated data, common in insurance.
• Limited research exists on machine learning for zero-inflated insurance datasets.

Purpose of the Study:

• To investigate the effectiveness of machine learning (ML) models in improving claim frequency forecast accuracy.
• To compare ML models against traditional statistical models on zero-inflated automobile insurance data.
• To identify the best-performing models for handling zero-inflation in insurance data.

Main Methods:

• Comparative evaluation of Poisson, Negative Binomial (NB), Zero-Inflated Poisson (ZIP), Hurdle Poisson, Zero-Inflated Negative Binomial (ZINB), Hurdle Negative Binomial, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) models.
• Utilized Mean Absolute Error (MAE) for performance assessment.
• Applied models to a specific automobile insurance dataset.

Main Results:

• Support Vector Machine (SVM) exhibited the highest predictive accuracy, effectively handling zero-inflation.
• Zero-Inflated Poisson (ZIP) and Zero-Inflated Negative Binomial (ZINB) models also showed strong performance.
• Traditional Poisson and NB models demonstrated lower predictive capabilities compared to ML and zero-inflated models.

Conclusions:

• Machine learning, particularly SVM, offers significant improvements in forecasting claim frequency for zero-inflated automobile insurance data.
• The findings suggest a shift towards advanced modeling techniques for better risk management in the insurance sector.
• Further research on diverse datasets is recommended to validate these findings.