In kevinykuo/deeptriangle: DeepTriangle

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)
library(deeptriangle)
library(tidyverse)

DeepTriangle: A Deep Learning Approach to Loss Reserving

This is the companion repository to the DeepTriangle paper.

A preview of the current draft can be found at https://kasaai.github.io/deeptriangle/manuscript.pdf.

A citeable preprint can be found at https://arxiv.org/abs/1804.09253.

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)

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.

library(cowplot)

p1 <- dt_plot_predictions(predictions, "1767", "commercial_auto", "paid_loss") + xlab("")
p2 <- dt_plot_predictions(predictions, "1767", "commercial_auto", "claims_outstanding")
p12 <- plot_grid(
  p1 + theme(legend.position = "none"),
  p2 + theme(legend.position = "none"),
  align = "v",
  ncol = 1
)
legend <- get_legend(p1)
plot_grid(p12, legend, rel_widths = c(1, 0.2), nrow = 1)

Company 337 workers' compensation.

library(cowplot)

p1 <- dt_plot_predictions(predictions, "337", "workers_compensation", "paid_loss") + xlab("")
p2 <- dt_plot_predictions(predictions, "337", "workers_compensation", "claims_outstanding")
p12 <- plot_grid(
  p1 + theme(legend.position = "none"),
  p2 + theme(legend.position = "none"),
  align = "v",
  ncol = 1
)
legend <- get_legend(p1)
plot_grid(p12, legend, rel_widths = c(1, 0.2), nrow = 1)

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/.

kevinykuo/deeptriangle documentation built on Aug. 8, 2019, 3:30 p.m.