Improving predictive performance within driver's insurance cross-border claims through a collaboration between ML and human experts

In a globalized world there is a large number of car accidents involving people visiting a country instead of being resident of the country. Driver insurance multilateral agreements such as the Green Card System in Europe functions to cover these situations, so that third party victims are compensated for loss caused both in material damage and in bodily injury. These cross-border arrangements encompass an entire new array of challenges when compared to local arrangements, as they must deal with different laws and regulations, protocols, prices, currencies, languages, and more. Because of its complexity, it is important for insurance companies in this context to make accurate prediction of costs ahead of payment dates. Consequently, Artificial Intelligence and supervised Machine Learning become very useful tools. However, there is a vast amount of data generated before, during, and after the car accident. Some of the data relates to the 'who', 'what', 'when', 'where', and 'why' of the car accident. This study dives into big data in order to investigate, through Global Interpretation Methods applied to black-box models in the form of Random Forest and CatBoost algorithms, which type of predictors are more important and how when predicting a driver´s insurance claim cost. It finds that some of the most relevant predictors are not directly related to the car or object damaged, but to the context such as time, location, parties involved, and cause of the accident. Additionally, this study configures a setting in which AI is used in collaboration with human expert claim handlers to achieve better predictive performance than either AI or human experts by themselves. The increase in predictive performance when including human expert input into the trained models is remarkable. This sheds new light into the currently discussed topic of AI versus Human Expert and suggests a successful implementation of a synergy between the two into insurance claim handling. Finally, it investigates how the progressive inflow of information about the car accident during the lifecycle of a claim changes the relative importance of AI versus human expert input for the model to make an output of the cost.