DeepTriangle

This is the companion repository to the paper DeepTriangle: A Deep Learning Approach to Loss Reserving.

Experiments

To get started, either clone the repo and build the R package, or install with

devtools::install_github("kasaai/deeptriangle")

You will also need the insurance package, which can be installed with

devtools::install_github("kasaai/insurance")

The experiments can be found in analysis/main.R. It is recommended that you use a GPU since many instances of the models are fit.

For convenience, we provide a predictions.feather file in the release.

predictions <- feather::read_feather("datasets/predictions.feather")

model_results <- dt_compute_metrics(predictions) %>%
  bind_rows(stochastic_model_results) %>%
  bind_rows(read_csv("datasets/automl_results.csv")) %>%
  gather(metric, value, mape, rmspe)

dt_tabulate_metrics(model_results, metric = "mape") %>%
  knitr::kable(booktabs = "T", digits = 3)

lob	Mack	ODP	CIT	LIT	AutoML	DeepTriangle
commercial_auto	0.060	0.217	0.052	0.052	0.068	0.043
other_liability	0.134	0.223	0.165	0.152	0.142	0.109
private_passenger_auto	0.038	0.039	0.038	0.040	0.036	0.025
workers_compensation	0.053	0.105	0.054	0.054	0.067	0.046

To create actual vs. predicted plots, use the dt_plot_predictions() function. Here are successful and unsuccessful examples of the model’s forecasting attempts.

Company 1767 commercial auto.

Company 337 workers’ compensation.

Testing different architectures

If you would like to try out different architectures or hyperparameters, you can do so by providing a function that returns a keras model. See the source code of dt_model() for a template.

For more details on the keras R package, visit https://keras.rstudio.com/.