Non-Heuristic Selection via Hybrid Regularized and Machine Learning Models for Insurance

TL;DR

This paper explores a hybrid modeling approach combining regularized feature selection with machine learning classifiers to predict insurance purchase behavior, achieving high accuracy while maintaining interpretability.

Contribution

It introduces a hybrid modeling framework that integrates regularized models with advanced machine learning classifiers for improved insurance customer prediction.

Findings

Hybrid CatBoost with Lasso achieved AUC of 0.861

Hybrid models outperformed individual classifiers in predictive metrics

Regularized feature selection maintained interpretability while enhancing accuracy

Abstract

In this study, machine learning models were tested to predict whether or not a customer of an insurance company would purchase a travel insurance product. For this purpose, secondary data provided by an open-source website that compiles databases from statistical modeling competitions were used. The dataset used presents approximately 2,700 records from an unidentified company in the tourism insurance sector. Initially, the feature engineering stage was carried out, which were selected through regularized models: Ridge, Lasso and Elastic-Net. In this phase, gains were observed not only in relation to dimensionality, but also in the maintenance of interpretative capacity, through the coefficients obtained. After this process, five classification models were evaluated (Random Forests, XGBoost, H2O GBM, LightGBM and CatBoost) separately and in a hybrid way with the previous regularized models, all these stages using the k-fold stratified cross-validation technique. The evaluations were conducted by traditional metrics, including AUC, precision, recall and F1 score. A very competitive hybrid model was obtained using CatBoost combined with Lasso feature selection, achieving an AUC of 0.861 and an F1 score of 0.808. These findings motivate us to present the effectiveness of using hybrid models as a way to obtain high predictive power and maintain the interpretability of the estimation process