Grainne McGuire and Jacky Poon from the Machine Learning in Reserving working party (MLR WP) will take you through their practical example of applying different machine learning (“ML”) models to a reserving data set using R. The goal is to illustrate a possible work flow in:
- setting up a data set for machine learning
- applying different ML models to this data
- tuning the ML hyper-parameters
- comparing and contrasting performance for the past fitted values and the future predictions.
The data set used is a simulated traditional aggregate data set of a 40x40 triangle of incremental quarterly payments over 10 years. The features are accident, development and calendar periods.
The presentation is introductory level, accompanied by a blog outlining the steps in detail, introducing the mlr3 package in R and providing code to allow the audience to recreate the examples in their own time.
It should be noted that the focus of the session is to illustrate the use of the different models rather than try to find the best model for a specific ML method. The methods show-cased are:
- Decision trees
- Random forests
- XGBoost
- Lasso.