Martin Neil, co-founder of Agena, explains how the insurance industry is ready to embrace Bayesian analytical methods
In what ways is artificial intelligence and big data unsuitable for systemic risk analysis in the insurance world?
Artificial intelligence (AI) approaches that rely on big data naively assume that all we need is more and larger data sets to compensate for our lack of understanding about what we are interested in — the process that gives rise to the risky events we want to avoid or the opportunities we wish to pursue.
The crowning achievements of science are, in large part, based on theories and models of reality; these have not been discovered using big data and machine learning, but are instead the result of human insight and experience.
For many insurance problems, where the processes and risks are already very well understood, AI and big data can really bring benefits, usually by automating and optimising pre-existing prediction solutions. However, where the processes of interest are complex, rare or novel in a way that gives rise to systemic risk — these, by their very nature, do not have a lot of data attached to them. It is arguable that systemic and complex risks are those that will have the biggest commercial impact, yet they are the least suitable for the application of big data-based AI.
How do you introduce causal modelling into risk analysis and decision making?
Instead of using big data we have pioneered ideas sourced from another branch of AI called Bayesian reasoning.
Two ideas are central to this Bayesian approach: causal modelling of the underlying risk process and the measurement of probability from data, but also using human judgement. By combining these we can exploit human insight and expertise to model variables and how they are connected so that we can explain how data is generated. This is what we call the “smart data” approach.
We help organisations structure their problems in a causal fashion so that they can see what behaviours and practices lead to better decisions and ultimately better outcomes. Identifying causal drivers is a creative exercise driven by conversation, imagination and the challenge of preconceived ideas, which is then followed up by the design of decision and risk models that directly represent reality.
What kinds of new modelling technologies should actuaries look to?
Actuaries and risk analysts should be looking at adopting Bayesian networks, but should not make the mistake of simply adopting a Bayesian statistics tool. At first glance these might look like identical offerings, but Bayesian statistics is simply a flavour of machine learning from data – it does not directly offer any help with causal reasoning nor does it provide any capability to embed human judgement and experience during model building.
The kinds of modelling features that are essential in Bayesian networks software should promote ease of use, high productivity and quick model revision, because most of the model building is done ‘live’ with problem owners, rather than being done separately in an ‘ivory tower’.
Which parts of the insurance decision-making process can Bayesian analysis help with?
Our insurance clients have used Bayesian and causal analysis to help in operational risk, cyber risk, risk tolerance, strategic decision making and investment modelling, among others. Our tools and technology have also been used in many other application areas, which at first glance might look unrelated to insurance problems but share common challenges that risk and decision makers in insurance can leverage.
We are currently completing our cloud version of AgenaRisk which will make these models available throughout a business without the rigmarole of having to buy licences, laptops, machines or write special code.
Is the industry ready to move on from traditional classical statistical analysis and embrace causal and Bayesian methods?
Yes, it is. Some 20 years ago, Bayesian and causal thinking was very much at the periphery of statistical thinking in insurance. But the industry has had to wake up to the fact that data and classical statistics isn’t enough to deal with the kinds of systemic events that regulations are now forcing them to consider.
Inevitably, actuaries and analysts make assumptions all the time and classical statistics encourage these to be ignored or kept hidden. This must change. There is a growing realisation that causal and Bayesian thinking is moving out from AI and computer science and replacing large areas where classical statistics was never a good fit in the first place.