Using artificial intelligence to better "calculate the future"

Extensive research, quantification, model validation... scientific expertise has always been important in the detection and evaluation of risks, which is central to the insurance profession. While AXA’s experts are in continual contact with the scientific world, connections from outside of the usual circles are less common. To promote such interactions, the AXA Research Fund has organized a series of workshops with some of the researchers who receive grants from the fund. The main speaker of the first meeting was Professor Maurizio Filippone, Director of the AXA Chair in Computational Statistics and Associate Professor at EURECOM. Marine Habart, Group Life, Savings & Health Chief Risk Officer and Madeleine-Sophie Deroche of the P&C Risk Management Department were also present. These three experts responded to our questions below.

The first of these meetings was on “Artificial Intelligence & Risk Modeling”. How does your research relate to this issue?

Maurizio Filippone: Decision-making requires the estimation of probabilities of certain events. And the more important the decision, the more important it is to be accurate. For this reason I specialize in a particular area: qualification of uncertainty. This involves designing tools to enable predictions to be made with a certain degree of confidence, while also quantifying the accompanying uncertainties... The objective here is to make algorithms more reliable and to facilitate decision-making.

What is risk modeling? Why is it essential for the insurance profession?

Marine Habart: The core business of an insurer is to know and anticipate the risks that its clients may face, in order to better protect them. This implies a dual responsibility: we must provide the best possible risk modeling to ensure our insurance products are priced appropriately, and we must remain vigilant to anticipate any potential changes to these risks. Accurate modeling is a major challenge: if a risk is underestimated we are exposed to losses that can lead to bankruptcy; but we become uncompetitive if a risk is overestimated.

Madeleine-Sophie Deroche: I specialize in natural disaster modeling and my risk assessment work has applications with three underlying objectives. The first level, as Marine discussed, concerns ensuring the financial stability of our business and the ability to pay claims quickly. The second involves the adjustment of prices and the detection of exposure accumulation in sensitive areas. The third level is about ensuring prevention to alert our customers of potential natural risks.

Marine Habart: The same goes for health! For example, we have developed a sophisticated global-level prospective model for pandemic risk assessment, which takes into account all virus-related parameters (transmissibility, lethality, emergence zone, etc.) but also the mitigation measures that could be implemented (vaccination, transport reduction, quarantine, etc.). This model incorporates parameters reflecting the current global situation: transport, the location of AXA policyholders around the world, preparedness of countries, temperature, etc.

What is the importance of the qualification of uncertainty in this process?

M-S. D.: It is essential that we are able to base our decisions on a certain degree of confidence in our results. In other words, rather than formulating a single prediction, we link it to high and low estimates, so that it can be associated to the various parameters that can alter the outcome. One of our main challenges today is to design our models in such a way as to guarantee modeling coherence, i.e. that we obtain sufficiently precise data to enable us to effectively fulfill our objectives.

How can these uncertainties be calculated, and how can they be reduced?

M.F.: In practice, calculating and reducing uncertainties involves running an algorithm numerous times to detect and fine-tune parameters. Unfortunately, this process can be very costly in terms of time and energy. Let’s consider an example of an algorithm with a computation time of a week. With each parameter change, you’ll need to wait another week to have a quantifiable result... Thus, the identification and then reduction of uncertainties can take months or even years. The challenge I face in my work is addressing this issue through a probabilistic approach. The idea is to use the language of probability to complement these algorithms and to determine which parameters require modification, to find ways to reach a solution as quickly as possible with the greatest reliability... while minimizing the operational steps needed to get there.

M.H.: What has been particularly inspirational about these workshops is that they’ve provided an opportunity to gain an overview of this issue of calculation time and its impact. Maurizio has even been able to show us one of his current projects that focuses on reducing the environmental impact of artificial intelligence, which takes a new approach that utilizes light-based computing. Indeed, the carbon footprint of this activity is an important concern. We always try to optimize computation time and associated resources when using algorithmic tools, which in turn reduces our carbon footprint. 

M-S. D.: Since 2016, due to the internalization of the artificial intelligence-based models that we use, we have developed expertise and tools that give us greater flexibility and precision. The integration of machine learning algorithms enables us to improve production performance, especially regarding weather-related hazards.  

In general, what was your impression of this meeting?

M.H.: Research is a key element in our work. Our role is to anticipate risks, to “calculate the future” by using our actuarial models, because we don’t have a crystal ball... It is therefore vital that we share our visions with researchers from different fields (demographers, epidemiologists etc.). This meeting provided a great opportunity to do so.

M-S. D.: One of our major challenges is to ensure that our risk modeling integrates the latest scientific and technological advances. Depending on the type of application required, we implement different partnerships, particularly with AXA-supported researchers. For these reasons, this meeting was very rewarding.

M.F.: I have received funding from AXA for seven years. Obtaining this chair has allowed me to build a team and carry out work on a subject for which I have a great passion. As a scientist, contributing to this event seemed very natural, and also essential. We had very good discussions with the AXA teams about their work on epidemic and climate risks... As a researcher, I find these meetings particularly valuable because they help reduce the time that elapses between a scientific discovery and its concrete application.