Predictive Modeling for Auto Insurance Risk Assessment Using Machine Learning Algorithms

Auto insurance is indispensable as it protects the insured against a spiritual and financial loss that arises in the case of unforeseen events or accidents. However, determining the premium to be charged is very difficult and has far-reaching consequences both for the company and the insured. Premia, which are too low, can lead to insolvency, whereas premia charged which are too high will not be competitive and, thus, are inconvenient for customers. Therefore, insurance companies are continuously facing the challenge of setting the correct premium, which is a very complex task. In recent years, insurances have uncovered their data treasures, with the revolution of big data and machine learning and have gained knowledge in how to utilize data and tools for risk assignment. Thus, the task of risk classification has changed, and quantitative models can now assist the underwriting decision. Machine learning algorithms have undergone rapid advancement and have produced useful outcomes, and many companies have already started to utilize the predictive powers by using supervised and unsupervised machine learning algorithms. Such approaches have enabled effective customer setup and thus have reduced overall costs for both insurance companies and their customers. This thesis aims to apply the supervised machine learning algorithms of gradient boosting, random forest, and artificial neural networks to insurance customer data. Risk classification can be assigned through the discrete outcome “number of claims in a given period”, which is also called loss count, and has been done by applying these models to the underlying dataset of different policyholders. The primary goal of the thesis concerns the evaluation of predictive accuracy, which can evaluate how well the model generalizes to unseen data. In addition, a risk classification will be produced for customers of two generations by using the models. Furthermore, the thesis aims to identify influential features by using machine learning algorithms that can predict loss count accurately since knowing which features play an important role in risk classification can assist the underwriting process.