The dominant underwriting approach is a mix between prescriptive rule-based engines and traditional underwriting. Applications are first assessed by automated prescriptive rule-based engines which typically are capable of processing only more standard and simple applications. The applications that cannot be processed by the prescriptive rule-based engines are then reviewed by underwriters or referred to the reinsurers. Building on previous work, this research aims to achieve high prediction accuracy with updated predictive machine learning models for complicated applications that cannot be processed by rule-based engines. Techniques such as natural language processing and clustering analysis are used to process free-text data such as descriptions of impairments and occupations. Various feature selection methods such as mutual information and recursive feature elimination are used to improve prediction accuracies. Machine learning algorithms such as Extreme Gradient Boost and Random Forest are used to predict underwriting decisions. The improved Extreme Gradient Boost model is the best performer with approximately 100% accuracy on the training set and 80% accuracy on the testing set. Additional in-depth analyses are carried out on the modelling outputs for the best performing model, such as the accuracy by risk class, analyses of differences between predicted and actual underwriting decisions. Other underwriting insights, as well as adaptions for popular product types, are discussed. Finally, LIME and SHAP models are used to explain the model predictions so the models can be implemented in practice and be explained to various non-technical stakeholders such as applicants and brokers.

Find the Q&A here: Q&A on 'Data Driven ERM'