The case for reinforcement learning in quant finance

The technology behind Google’s AlphaGo has been strangely overlooked by quants

When Google’s Deepmind defeated the world’s top Go player in 2016, it was seen as a breakthrough for artificial intelligence. But the technique used to train AlphaGo, known as reinforcement learning (RL), has not gained much traction in finance, despite its ability to handle complex, multi-period decisions.

Igor Halperin, a senior quantitative researcher at Fidelity Investments, thinks it’s time for that to change: “RL is the best and most natural solution to most of the problems we have in quantitative finance,” he says.

He argues that nearly all problems in quantitative finance – including options pricing, dynamic portfolio optimisation and dynamic wealth management – can be solved with RL or inverse RL, or a combination of the two.

RL techniques work sequentially. At each stage, the algorithm observes the reward obtained in previous stages and proceeds accordingly, trying as many combinations of actions as possible to maximise a given reward function.

Halperin and Matthew Dixon, assistant professor at the Illinois Institute of Technology in Chicago, have published a research paper on the application of RL to dynamic wealth management.

They spotlight two techniques, which can be used either individually or in combination. The first is G-learning, a probabilistic extension of the Q-learning approach popularised by Deepmind. The advantage of G-learning – which is relatively new to finance, despite being well established in other fields – is that it can handle noisy environments and high dimensionality, which Q-learning struggles with.

For this reason, a previous effort by Gordon Ritter to apply Q-learning to dynamic portfolio optimisation was limited to a small number of assets.

“[Q-learning] couldn’t manage a portfolio of 500 stocks and it doesn’t cope well with noisy environments such as financial markets,” says Halperin.

RL is better than Black-Scholes and risk-neutral pricing in general, which makes more harm than good

Igor Halperin, Fidelity Investments

G-learning does not suffer from this problem. Given a reward function – in this case, the maximisation of wealth in a given time horizon – it can find the optimal combination of actions to reach a target outcome using the available historical data.

The second technique, which Halperin and Dixon introduce for the first time in their paper, is called generative inverse reinforcement learning, or GIRL. This works the opposite way to G-learning. GIRL takes the outcomes of strategies – the holdings and returns of a portfolio – and works backwards to infer what investment strategy the manager followed.

Halperin says the tools can be combined to create a robo-advisory solution. GIRL can be used to learn existing strategies, which G-learning can then optimally replicate for clients. The adviser can then potentially tailor solutions to clients’ objectives and level of risk aversion.

Other potential applications include minimising market impact in trade execution. The Royal Bank of Canada’s research centre, Borealis AI, has already used RL to develop a new trade execution system for the bank, called Aiden.

Halperin is also convinced RL can be successfully applied to price derivatives. “RL is better than Black-Scholes and risk-neutral pricing in general, which makes more harm than good,” he says. “Option pricing is all about managing risk, but the main assumption of risk-neutral formulation is that there is no risk, which is self-contradictory.”

Halperin and Dixon’s research is still in the experimental phase and has not been tested in practice, but the authors are confident about its effectiveness.

So why is RL missing from most quants’ existing toolkits? Matthew Taylor, associate professor of computer science at the University of Alberta, reckons it might be down to a scarcity of expertise. “In general, RL is not used much in finance, at least publicly,” he says. “There is a barrier to entry for financial institutions and there aren’t enough reinforcement learning professionals, or enough experts for all the potential.”

The work of Halperin, Dixon and others may fuel wider efforts to apply RL in finance.