DeepTriangle: A Deep Learning Approach to Loss Reserving

Kevin Kuo

doi:10.3390/risks7030097

Abstract

We propose a novel approach for loss reserving based on deep neural networks. The approach allows for joint modeling of paid losses and claims outstanding, and incorporation of heterogeneous inputs. We validate the models on loss reserving data across lines of business, and show that they improve on the predictive accuracy of existing stochastic methods. The models require minimal feature engineering and expert input, and can be automated to produce forecasts more frequently than manual workflows.

Highlights

In the loss reserving exercise for property and casualty insurers, actuaries are concerned with forecasting future payments due to claims
While Caruana (1997) contains detailed discourse on the specific mechanisms of multi-task learning, we provide some heuristics on why it may improve predictions: by using the response data for claims outstanding, we are effectively increasing the training data size since we are providing more signals to the learning algorithm; there may be hidden features, useful for predicting paid losses, that are more learned by trying to predict claims outstanding; by trying to predict claims outstanding during training, we are imposing a bias towards neural network weight configurations which perform that task well, which lessens the likelihood of arriving at a model that overfits to random noise
We see that DeepTriangle improves the performance of the popular chain ladder and ODP models, common machine learning models, and Bayesian stochastic models

Summary

Introduction

In the loss reserving exercise for property and casualty insurers, actuaries are concerned with forecasting future payments due to claims. Estimating these payments is important from the perspectives of various stakeholders in the insurance industry. Loss reserves, and transactions related to them, are essential components in the balance sheet and income statement of the insurer. For the regulators, accurate loss reserves are needed to appropriately understand the financial soundness of the insurer. There can be time lags both for reporting of claims, where the insurer is not notified of a loss until long after it has occurred, and for final development of claims, where payments continue long after the loss has been reported. The amounts of claims are uncertain before they have fully developed

Objectives

Results

Conclusion