Use and Value of Unusual Data in Actuarial Science

A Joint Research Initiative between Université de Québec à Montréal and AXA Group Actuarial Function

Satellite Images

The weak correlation between indices and insurance losses has, until now, hindered their development. Professor Charpentier will explore how to extract relevant but sometimes noisy information from satellite images, such as the presence of water or burnt areas, to set up insurance indices in relation to the actual loss experience. The use of data with greater granularity could allow insurers to generate indices of much better quality, therefore achieving much faster index coverage. Professor Charpentier will also investigate how pictures can be translated into a mathematical or automated fashion for the pricing or claims assessment process, such as after a natural catastrophe or to form a crop yield index for crop insurance. He will ultimately develop an algorithm that can populate missing information at the time of submission, such as the construction type of a building, the number of floors of a building, or its year of construction.

Network data

Peer-to-peer insurance is based on reciprocity between the parties involved. It involves networking or using an existing network of insured persons by creating communities that share the risk within an acceptable limit. The insured people in the network exchange a promise to provide financial compensation up to a predefined amount in the occurrence of a claim for one of the members. A promise of a few dozen euros can be converted into a few hundred euros if the network is relatively dense, thus offering a simple technique for buying back the deductible. Experiments in micro-credit have shown that these "sharing economy" type mechanisms strengthen solidarity between members, reduce fraud, and offer an alternative to prevention.

Using health exposure and claims data, Professor Charpentier will develop an algorithm that shows how individuals that are connected in some way (via the same company, friends/family, or location) exhibit similar or dissimilar risk features, such as mortality correlations among siblings or cancer incidence by location (perhaps due to being close to an electrical substation or a main road). 

Text data 

Tweets following the occurrence of a disaster can be used to understand spatial diffusion, which can be useful for the management of claims and crises. Natural catastrophes and the spread of communicable diseases studied via tweets, including the recent COVID-19 pandemic, can be used to support customers such as targeted health apps. Using publicly available tweets via Twitter, Professor Charpentier will develop an algorithm to show how social media posts can be used to form an index. This index can then be used for marketing, claims response, or complaint management. 

The three research strands above will be investigated separately, as one type of data is not necessarily related to another. For example, pictures of car damage for claims assessment may not be related to understanding text data. However, Professor Charpentier will explore potential correlations where applicable, for example in pictures that contains text.

February 2022