Zobecněný lineární model s využitím neuronových sítí pro predikci počtu škodních událostí

The claim frequency, along with the extent of damage, is a crucial part of risk estimation in the insurance sector. Using appropriate data, our goal is to compare traditional statistical methods, particularly generalized linear models, with methods based on machine learning, especially neural networks, in the context of predicting the number of insurance events in mandatory liability insurance. Both mentioned approaches were applied to real data similar to the Czech mandatory insurance market, allowing us to estimate the predictive capabilities of both methods overall and in individual risk segments and to compare their advantages. While traditional approaches offer proven models with good interpretability, modern techniques provide advanced options for modeling complex patterns in the data. On categorized data for generalized linear models, the advantages of statistical learning, which more often made predictive errors in less represented segments, were not sufficiently demonstrated. An advantage of the neural networks is its ability to work with categorical variables that have many possible values, which can be quickly clustered and visualized using embedded layers. Primarily, the future of risk prediction in insurance is seen especially in the use of hybrid models that combine the important interpretability and associated expert scalability of generalized linear models with the identification of more complex dependencies that can be appropriately uncovered using neural networks.