Advanced Umbrella Insurance Risk Aggregation Using Machine Learning

Abstract

Weather events and natural catastrophes can provide insurance companies with risk levels and behavior over large geographical areas through a single metric. Nevertheless, due to a small number of observations and high dimensional data, developing risk models from modellers founded on scientific knowledge has traditionally been problematic. To overcome the problem of data shortages, other approaches based on machine learning and artificial intelligence, which can automatically uncover dependencies in data, have been explored. However, competition remains tough due to a growing empirical literature that uses new modeling techniques and combinations of old ones. In particular, randomized forests and multi-level regression trees type algorithms are on the front line for practical and easy access war- stages.
Machine Learning approaches outperform basic and widely applied equivalent experimental techniques, such as generalized linear models with a Gaussian distribution for continuous outcomes, Kriging with a linear trend, pure Kriging, and square covariance function. Aggregated rainfall insurance contracts are affected by several sources of uncertainty, the most important being the number of observations in the risk scenario to be simulated. They also compare the various inputs generated from the initial spatially continuous process and demonstrate that by estimating the model parameters using inputs derived from an RF-based model, re-insurers can substantially augment profit functions when incorporated into weather risk models.
Ultimately, they argue that ML models, such as RFs, have the capability to reproduce scientific models, plus they contend that the architectures employed could be modified to accommodate a larger area of study. They believe that the models herein presented will be useful tools at the service of the insurance industry. Weather index-based insurance schemes offer insurance companies the opportunity to effectively hedge against climate events that could have costlier expenses. Given the premises, they will focus entirely on the umbrella industry. They will not discuss first loss contracts, one-sided accounts, and non-anonymity contracts. Because of issues in the observed experienced loss data, they target events of exogenous random origin.