This study explores the insurance pricing domain in the motor insurance industry, focusing on the creation of “technical models” which are essentially obtained after combining the frequency model (the expected number of claims per unit of exposure) and the severity model (the expected amount per claim). Technical models are designed to predict the loss costs (the product of frequency and severity, i.e., the expected claim amount per unit of exposure) and this is a main factor that is taken into account for pricing insurance policies. Other factors for pricing include the company expenses, investments, reinsurance, underwriting, and other regulatory restrictions. Different machine learning methodologies, including the Generalised Linear Model (GLM), Gradient Boosting Machine (GBM), Artificial Neural Networks (ANN), and a unique hybrid model that combines GLM and ANN, were explored for creating the technical models. This study was conducted on the French Motor Third Party Liability datasets, “freMTPL2freq” and “freMTPL2sev” included in the R package CASdatasets. After building the aforementioned models, they were evaluated and it was observed that the hybrid model which combines GLM and ANN outperformed all other models. ANN also demonstrated better predictions closely aligning with the performance of the hybrid model. The better performance of neural network models points to the need for actuarial science and the insurance industry to look beyond traditional modelling methodologies like GLM.

中文翻译：

---

广义线性模型与机器学习方法在预测汽车保险损失成本方面的比较

本研究探讨了车险行业的保险定价领域，重点关注“技术模型”的创建，这些模型本质上是结合频率模型（每单位风险敞口的预期索赔数量）和严重性模型（预期金额）每项索赔）。技术模型旨在预测损失成本（频率和严重程度的乘积，即每单位风险敞口的预期索赔金额），这是保险单定价时考虑的主要因素。定价的其他因素包括公司费用、投资、再保险、承保和其他监管限制。探索了不同的机器学习方法来创建技术模型，包括广义线性模型 (GLM)、梯度提升机 (GBM)、人工神经网络 (ANN) 以及结合 GLM 和 ANN 的独特混合模型。这项研究是在法国汽车第三方责任数据集“freMTPL2freq”和“freMTPL2sev”上进行的，这些数据集包含在 R 包 CASdatasets 中。建立上述模型后，对它们进行了评估，发现结合 GLM 和 ANN 的混合模型优于所有其他模型。人工神经网络还展示了与混合模型的性能密切相关的更好的预测。神经网络模型的更好性能表明精算科学和保险行业需要超越 GLM 等传统建模方法。