Abstract:
Actuaries depend primarily on simulations to build catastrophe (cat) models. By applying
image processing techniques such as the novel convolutional neural network (CNN), it is
possible to use both numeric and map data to improve modeling. To this end, we
illustrated applying CNN to calibrate a cat model, using the more efficient U-Net
architecture, which has been shown to perform well with limited data because of its
localized predictive ability. We evaluated our CNN model using real-life data obtained
from the National Oceanic and Atmospheric Administration. We also used these data to
build a more traditional generalized linear model and compared the results with those
from the CNN model. We found that even with limited data, the CNN model performed
well
Description:
Actuaries depend primarily on simulations to build catastrophe (cat) models. By applying
image processing techniques such as the novel convolutional neural network (CNN), it is
possible to use both numeric and map data to improve modeling. To this end, we
illustrated applying CNN to calibrate a cat model, using the more efficient U-Net
architecture, which has been shown to perform well with limited data because of its
localized predictive ability. We evaluated our CNN model using real-life data obtained
from the National Oceanic and Atmospheric Administration. We also used these data to
build a more traditional generalized linear model and compared the results with those
from the CNN model. We found that even with limited data, the CNN model performed
well
