The parallels between artificial intelligence and pension scheme investing
The field of Artificial Intelligence (AI) seems to operate in a different world to pensions. But there’s a science to thinking about pension scheme asset and liability models that takes a remarkably similar approach to an AI program’s approach to winning a game of Go.
Google’s DeepMind, based in London, is at the forefront of spectacular progress in AI over the past few years. This includes impressive advances in the ancient Chinese board game of Go in the space of two years. Their program is called AlphaGo.
What makes AlphaGo so good? A huge amount of processing power and some very clever programming obviously have an awful lot to do with it. The success of AlphaGo over other AI programs lies largely in optimisations and subtleties of execution that, in another career, I would love to study in detail. But its advantage, and many other AI programs advantage, over humans can be largely understood in terms of its basic approach:
- Objective assessments. It weighs up options objectively and without prejudice (using a Monte Carlo tree search algorithm). Indeed, AlphaGo Zero only learns from playing against itself, so is not biased by intuitive shortcuts and heuristics that humans must resort to
- Thinking far ahead. It simulates very far – all the way to the end of the game – rather than just a few moves ahead, even if it can’t possibly test every possible scenario. In contrast, humans find it hard to think that far ahead
- Focusing on winning. It is not overly greedy and instead focuses on a higher chance of winning, even if not by much. Its resultant playing style strongly favours a greater probability of winning by fewer points over lesser probability of winning by more points.
Understanding the approach leading to AlphaGo’s success can help trustees win their own very serious game of ensuring all pensions can be paid
I won’t pretend that asset and liability models are nearly as complicated as those used by AlphaGo. However, an approach I prefer to tackling the problem of setting long-term investment strategy for pension schemes, shares some remarkable similarities to how, fundamentally, AlphaGo tries to win:
- It uses a Monte Carlo engine to assess the attractiveness of a given strategy
- It simulates all the way to the end – here, this is payment of the last pension, buyout with an insurance or entry of the scheme into the Pension Protection Fund (PPF)
- It defines success in terms of trustees’ ultimate aim to pay all pensions instead of taking unnecessary risk trying to generate more return than is needed to pay all pensions.
Possible metrics to judge the attractiveness of an investment strategy are numbers such as:
- the chance that all pensions can be paid
- the expected proportion of benefits that can be met from a given investment strategy (which also allows for the extent of any shortfall)
Like AlphaGo, models can be set up to come up with some counterintuitive solutions. For example, in some circumstances allowing for covenant risk should not necessarily lead to taking less investment risk.
Go and financial markets are two very different beasts. The latter is far more uncertain and complex, which perversely can mean that overly complicated approaches can obfuscate more than they add value, and a large degree of human oversight is needed (so hopefully the robots won’t be replacing me just yet). However, understanding the approach that led to AlphaGo’s success can help trustees win their own very serious game – ensuring all pensions can be paid as they fall due.