Evaluating Machine Learning Methods for IBNR Claims Reserving on Real Claims Data
Despite the growing body of research and adoption of machine learning in the field of insurance, the practicality of such methods in claims reserving is still unclear. We are increasingly seeing different machine learning models and methods being explored as alternative ways of estimating claims reserves. However, a lot of the literature on applying machine learning to claims reserving has been evaluated on synthetic data or on old yearly data. Hence, testing the efficacy of such methods with up-to-date claims data and comparing them against estimates made by actuaries is important if we are to see practitioners take advantage of existing research. Apart from focusing on high-frequency and short-tailed products, machine learning can give the biggest improvement in the first few development months since they contain the highest level of uncertainty. With this in mind, we present a machine learning approach to monthly IBNR reserving -- based on the work of Caesar Balona and Ronald Richman -- that we validate on real monthly claims data from Gjensidige. Actual versus Expected analyses demonstrate that this approach can in many instances outperform the reserving done by an actuary. Potential benefits of such methods include increased automatization and accuracy while removing some of the subjective evaluation when performing claims reserving.
Find the Q&A here: Q&A on 'General Insurance Reserving'