• Mikhail I. Morozov, Yugra state university (Khanty-Mansiysk, Russia)

This study examines the primary limitations of tree-based decision models in the insurance sector, namely their sensitivity to noise and outliers, susceptibility to overfitting, and limited generalization capabilities. To overcome these challenges, we propose leveraging advanced machine learning techniques, neural networks, and stacking ensemble methods. We hypothesize that the quality of loss prediction can be slightly improved by combining the predictions of multiple base models (XGBoost, CatBoost, Logistic Regression, Random Forest, and TabTransformer), alongside the original dataset features, into a more robust ensemble. Specifically, the meta-model is constructed as a fully connected neural network with an additional input layer for the original features. The empirical study utilizes Russian auto insurance data from 2023 to 2025, detailing the experimental framework for training heterogeneous machine learning and neural network models. The data preprocessing pipeline is described, with a particular focus on a specific categorical encoding scheme based on the European car classification by luxury segment. Special attention is given to the severe imbalance of the target class, as losses account for less than 3% of the total dataset. Based on the trained base models, ensemble models are constructed using five stacking techniques: Simple Average, Constrained Logistic Regression, Adaptive Regression by Mixing (ARM), ARM-Tweedie, and Multilayer Perceptron (MLP). A comprehensive comparative analysis of the base models and the ensembles is conducted using both classification metrics (ROC-AUC, PR-AUC) and probabilistic metrics (Log Loss, Brier Score). The results indicate that the best-performing base model is TabTransformer (Brier Score = 0.037, Log Loss = 0.185, ROC-AUC = 0.635, PR-AUC = 0.054). The MLP-based ensemble outperformed it across all key metrics (Brier Score = 0.026, Log Loss = 0.131, ROC-AUC = 0.649, PR-AUC = 0.055). These findings confirm the initial hypothesis: the fully connected neural network ensemble yields a slightly improvement in the primary evaluation metrics. This demonstrates the methodological value of the proposed approach: even under conditions of a weak data signal, the neural network ensemble, augmented with original features, ensures a more effective aggregation of predictions compared to traditional predictive analytics and alternative stacking ensemble methods.

decision trees, neural networks, machine learning, vehicle insurance, loss ratio, stacking ensemble

2026-09-03