Predicting claims in auto insurance using deep neural networks

In the modern world, the insurance market is subject to significant changes, including under the influence of the use of digital technologies and the introduction of machine learning methods in insurance scoring. The object of the study is a data set with records of insurance policies. The study uses a deep nonlinear neural network to predict the occurrence of claim loss on auto insurance policies. Before using a multilayer neural network, data is preprocessed, and possible data leakage is eliminated. At the output of the neural network model, the resulting loss probability value is converted to a binary value. The model is evaluated using the ROC-AUC metric, with a graph of the ROC curve. The results show that the obtained model has predictive accuracy, but not high enough accuracy for industrial applications of the chosen model. The findings indicate the need for further research on ways to solve this problem using other machine learning methods.