Skip to main content

Filip Axelsson: Premieskattning med hjälp utav mer tolkningsbara trädbaserade metoder

Master thesis final presentation

Time: Wed 2023-06-07 11.25

Location: Meeting room 9, floor 2, house 1, Albano

Respondent: Filip Axelsson

Supervisor: Mathias Lindholm

Export to calendar


For many years, the same method has been used for pricing in property insurance. This method is known as generalized linear models (GLM), which are robust, reliable, and interpretable, making them still used today. However, we are witnessing a clear development in machine learning where methods are becoming increasingly better, thus providing more accurate predictions. In this report, two machine learning methods will be used: Gradient Boosting Machine (GBM) and Random Forest. Both of these models are tree-based models but also black-box models. This means that while you can observe their input and output, the process that occurs in between is unclear or complex. This leads to their drawback, as they are dicult to interpret, and there are different ways to increase understanding of the models, but these often provide only a local perspective or apply to only a part of a larger dataset. An alternative way to make these machine learning methods more interpretable is to use surrogate models. In this report, simple decision trees will be fitted to the predictions of the black-box models. By using the existing dataset `MCcase" with predefined tariffs, we can investigate if any of the black-box models perform better than GLM and also assess the quality of these surrogate models. The results showed that only the Random Forest model and its surrogate model outperformed GLM for this specific dataset. However, all models performed at a similar level. One possible reason for GBM performing the worst could be the lack of correlation in the dataset, as one of GBM's main strengths is handling such correlations. Despite this, we were able to demonstrate that the surrogate models served their purpose and could potentially offer a new method for premium estimations.