Literature Review

Collect published research on Machine Learning in reserving techniques

The Literature Review Workstream collects published research on Machine Learning (ML) in reserving techniques. The literature collected so far discuss the following methods: Neural Networks (NN), Gradient Boosting Machines (GBM) and random forests (RF). By summer 2020, the workstream reviewed a total of 69 papers and continues to scan for new relevant papers to enrich the lessons learned.

There is a growing body of literature on application of ML techniques to general insurance reserving. The majority use individual claims data, and this possibly could be extended to cover micro-level data points in the region surrounding the claim event to include: transaction dates, reopened claims, closing dates etc. There is still clearly a lot to learn and a compelling case for use of simulated data to close gaps in data is clearly present and should be considered.

Reviewed Papers

A compilation of the full list of papers reviewed by the working party as at summer 2020 can be found here. We are continuing to review papers and will update this page from time to time.

Papers of Interest

We have flagged papers of particular interest in the table below.

Title	Model	Notes
Ahlgren M (2018). Claims reserving using gradient boosting and Generalized Linear Models	GBM	Compares GBMs with GLMs. Conclusion is that the performance of the two is similar.GLM is then preferred on the ground of interpretability.
Duval F and Pigeon M (2019). Individual loss reserving using a gradient boosting-based approach	GBM	Uses GBM to boost structured GLM micro-models.
Taylor, G. (2019). Loss reserving models: Granular and machine learning forms	NN	Surveys recent developments in granular models and machine learning models for loss reserving, and assesses their potential for future development.
Gabrielli, A., Richman, R., and Wüthrich, M. V. (2020). Neural network embedding of the over-dispersed Poisson reserving model	NN	This paper embeds a classical actuarial regression model into a neural network architecture. Such models can be fitted efficiently which allows us to explore bootstrap methods for prediction uncertainty.
Kuo, K. (2020). Individual claims forecasting with Bayesian mixture density networks	NN	We introduce an individual claims forecasting framework utilizing Bayesian mixture density networks that can be used for claims analytics tasks such as case reserving and triaging.
Baudry M and Robert C (2017). Non-parametric individual claim reserving in insurance	RF	Model results in reasonable forecasts with a prediction error much reduced relative to Chain-Ladder.
De Felice M and Moriconi F (2019). Claim Watching and Individual Claims Reserving Using Classification and Regression Trees	RF	This model forecasts not only the ultimate claim cost but also the claim status at various points in time.

Emerging Literature Gaps

The papers reviewed showed us some gaps in the literature that need to be followed up. For example, model classifications of interest to the working party are:

• Bayesian Network Classifiers
• Ensemble models
• Mixed Density Function

We will be collecting papers for the above models and reviewing them in the coming months so that summaries and learnings can be published.

Posts

Talks