Advancing the Use of Deep Learning in Loss Reserving: A Generalized DeepTriangle Approach

Linear Models 20 years ago, to Gradient Boosting Machines ten years ago, to neural networks in the last few years. The benefits of enhanced pricing accuracy and competitive advantage to early adopters have been widely recognised.

While popular in pricing, these techniques have been slower to be adopted in reserving. Non-parametric individual claim reserving using decision trees was only explored in 2017, ten years after comparable advances in pricing. More recently, DeepTriangle has explored an aggregated projection approach for predicting claims payments in incremental loss triangles.

However, these approaches all have their limitations. Individual claims projections are volatile and present bias in prediction when aggregated. The estimates from DeepTriangle, a neural network, can be hard to explain, and the original paper predicts loss ratios instead of claims payments.

In many regulatory environments, reserving has more oversight than pricing which makes the application of ML techniques more difficult. Not only does the result need to be accurate, but it also needs to be explainable and stable. Both these can be difficult for ML approaches known for their lack of explainability and complexity.

In this paper we explore machine learning approaches for reserving and address the limitations of existing approaches. We then compare the approaches in terms of accuracy, interpretability, and transferability. Finally, we develop a simpler, more stable, and more explainable approach to reserving based on learnings from existing approaches. Through this, we want to demonstrate that machine learning approaches can solve reserving problems whilst addressing oversight challenges, and should be considered for regulatory approval. Actuaries should be innovators, making use of the latest analytics solutions, which should extend to that which underpins our profession – actuarial reserving.