Artificial Intelligence in capital management - Can we solve the balance sheet?

robot hand moving chess piece

Peter Murphy, member of the Artificial Intelligence and Automation Working Party, is currently investigating how AI and automation can be used in various areas of actuarial practice.

The Working Party started by reviewing the current use of Artificial Intelligence in life insurance companies. Focusing on ALM, capital management, and investment, I have tried to find examples of actuaries using AI techniques to “Solve the balance sheet”. After consulting several practitioners in this area, the working party’s view is that currently AI is not being used too widely, if at all, in ALM and investment within life companies. So this blog has two aims. Firstly, to consider what might be stopping actuaries using AI in these areas and secondly, to see if a wider audience has any good examples.

It is not the case the actuaries are just ignoring AI and Automation. There are many cases of use across the industry from pricing and experience analysis to fitting of proxy models for capital modelling. But it seems that actuaries working in Investment and ALM have been slower in their uptake of these new techniques.

While it has been difficult to find widespread adoption, there are a few examples out there, from recently presented work on training machines to assess Matching Adjustment eligibility[1], to some less well-defined efforts to use Machine Learning techniques to optimise Matching Adjustment portfolios. Also, there are papers looking at the use of reinforcement learning techniques to optimise hedging strategies[2](not in a specific insurance context), but again these problems tend to have well-defined objectives.

But these sporadic examples don’t change the fact that most actuaries in these areas are making very little use of AI techniques. Why might this be the case?

The most famous triumphs of Artificial Intelligence have been problems where there is a clear objective and either, vast quantities of well-labelled data, or where data can be easily simulated. AI is good at chess because there is a very clear goal (winning) and it is possible to simulate a huge number of games to train a system. Similarly, AI is good at picture recognition because companies have generated vast data sets of labelled pictures – as we all know from hours spent trying to identify traffic lights every time we want to buy something.

So why might this not work so well for ALM and capital management? Insurance companies are often faced with complex optimisation problems balancing a wide range of different objectives. For example, the investment strategy for an annuity portfolio will be aiming to maximise the investment return subject to a very wide range of constraints. The company will want close matching of liabilities, limited impact on capital requirements, high levels of diversification and to meet ESG requirements. Developing a quantitative framework that balances all of these is hard, before trying to clearly articulate a company’s appetite for risk.

Further, getting complete information about the problem can be difficult. It is expensive to derive impacts and capital requirements across thousands of different strategies.  To deal with this actuaries tend to automate where they can and then use skill and judgement where this gets too hard. In the example above it is likely the ALM matching would be automated using a simple set of heuristics, but that the more complex decisions around the selection of individual assets, and mix of rating by duration would be more manual processes, relying on the judgement and experience of the actuary.

But the purpose of this Working Party is to consider how the actuarial process needs to evolve – and the direction of travel is clearly to more automation and more use of advanced techniques.

If companies want to use Artificial Intelligence to solve these problems they would need a very clearly defined, quantitative capital management framework, and the required tools need to be in place to allow this to be estimated across wide ranges of different strategies both now and projected into the future. While it is straightforward to write down a requirement to recalculate the balance sheet under myriad different investment strategies, it is likely to be very difficult for all but the simplest business.

There are also difficulties in defining a solely quantitative capital management framework. Insurance companies tend to be large, complex with interactions across business lines and products sold over decades. Modelling all of these interactions is not always possible. Also there might be wider issues that need to be considered around fairness or the purpose of the insurer.

I have tried to outline some of the reasons why solving these problems using AI is difficult, without further developments in companies’ modelling capability.  If you’re an ALM/investment actuary and have experience in using AI techniques in this area, the working party would be keen to hear from you. Please feel free to reach out to me at peter.murphy@royallondon.com.


[2] E.g. “Deep Hedging of Derivatives Using Reinforcement Learning”,  Jay Cao, Jacky Chen, John Hull, Zissis Poulos