Ultimate Loss Reserve Forecasting Using Bidirectional LSTMs

This paper aims to demonstrate how deep learning (a subset of machine learning) can be used to forecast the ultimate losses of a sample group of Property and Casualty insurance companies. The paper initially explores the concept of loss development - how losses incurred by an insurance company mature across time. These losses then reach a final amount, known as the ultimate loss. The paper also looks at some already existing methods of forecasting the ultimate loss. The paper then introduces a novel method of forecasting losses, one which involves the use of deep learning neural networks. This new method uses Long Short-Term Memory (LSTM) - an advanced form of a deep learning architecture which specializes in finding patterns in temporal data. The findings of this method are then compared to a currently existing Python package which can also be used to predict ultimate losses. The paper also goes to critique some shortcomings of the model that is presented.