Insight & Analysis

Disruptive influences: the power behind real treasury innovation (part two)

AI robot working with visual technology

Disruptive technology can take many forms for treasurers. Artificial intelligence receives a lot of coverage; can treasurers really find a use for it?

Artificial intelligence (AI) is the financial technology theme of the day. And yet how many treasurers have a clear idea of how it can be used professionally? A straw poll at the recent ACT annual UK conference suggested few saw a definite purpose for it.

Paul Higdon, Chief Product Officer at ION Treasury, speaking at the event, admitted that it’s hard to predict exactly what value AI will deliver in the future. However, he is quietly confident that current AI developments will impact treasury, urging treasury preparedness. Using the development of chess computers as an example, he explained why.

The best technology has been beating chess grandmasters since 1997, when IBM’s Deep Blue defeated Gary Kasparov using sheer brute force processing power. From this experience came the notion of the potential of joining forces with technology, rather than battling against it.

A freestyle chess championship, created in 2005, allowed any grouping of humans and technology, including the best players and the highest processing power. The first victorious team, ZackS, consisted of two amateur players armed with three relatively humble chess-software-loaded PCs. Their power, Higdon explains, existed in understanding those programmes intimately. “The symbiosis of man and machine was able to outperform the best players and the best computers in the world.”

Fast-forward to today, and Higdon believes humans should accept that, at some point, computers will be better than us at performing many more tasks, particularly in the work setting. “We should leverage that technology, learning from it and working with it, to elevate our performance.”

Algorithmic learning

The source of excitement around AI that has emerged over the last decade is largely around progress in the field of machine learning (ML), particularly in deep reinforcement learning.

In 2017, Stockfish 8 was the ultimate chess machine. It had an all human chess strategy built into it. It was declared unbeatable. That same year, the generalised deep-learning algorithm, AlphaZero arrived. It was not programmed in the same way as traditional AI, being taught the rules of the game then given time to ‘learn’ how to play. In just four hours it played against itself nine million times, exploring the permutations and self-optimising its understanding of game play at each step.

AlphaZero was pitted against the ‘unbeatable’ Stockfish. In 100 games, 72 were drawn, AlphaZero won the remaining 28. Within four hours, a deep-learning algorithm had surpassed the entire human history of chess strategy and had created entirely new moves.

“The human chess-player approach is not to give up but to try to understand the new strategies revealed by deep learning,” says Higdon. This requires an understanding of the value of the position of the remaining pieces, rather than simply assessing the value of the pieces themselves. “We expect the same kind of process to happen throughout the business environment, including treasury, to reveal completely new strategies.”

Beyond AI

Classical AI – where programmers tell the software what it should do based on what is required – has a role to play in treasury, and will do so for some time to come, says Higdon. More advanced is MLl. This is not programmed in the traditional sense but uses an algorithm to analyse and learn from much larger data sets. This is where the technology becomes interesting for treasury purposes.

“Treasury is about managing risk and controlling outcomes. Having neural networks running treasury and not knowing why they are taking decisions is not going to happen in treasury anytime soon,” Higdon assures. “Instead, we should leverage classical AI to automate more and more and then gradually build on our understanding from ML and deep learning technologies.”

However, a subset of ML is deep reinforcement learning. Inspired by the way organic brains and nervous systems connect and learn, this uses neural networks, reinforcing positive outcomes to derive better ways of doing things. This is an area of exploration being treated with caution, says Higdon. “I believe that we need all three of these types of AI – classical, ML and reinforcement learning – to really take us forward.”

Vendor approach

ION is sponsoring a doctorate-level research programme at the University of Graz in Austria. This is exploring inputs from a number of treasury teams, to see what aspects of ML and associated algorithms could have the most impact on treasuries and their wider organisations.

Drawn from a list of 20 treasury use cases for ML, the top five, as defined by the academic research to date are:

  • Cash flow forecasting.

  • Classification of liquidity groups (preparing for accounting or matching).

  • Matching.

  • Optimising cash investments based on policy, positions and market rates.

  • Identifying anomalies in payments.

“The team has moved on from identification into building a number of models, including neural networks,” say Higdon. In forecasting, for example, this approach is being compared with linear regression modelling and manual methods, testing outputs for relative accuracy.

Neural networks require training data to help fine tune the algorithm towards more accurate outputs. The system being developed by the university currently uses historic actual information. It will soon be able to absorb economic factors too, says Higdon. This will take outputs to new levels of commercial usability.

A business influenced by weather trends, for example, will be able to consume related data through the algorithm, revealing the impact of weather events on cash forecasting. The aim is for the system to move towards self-optimisation, through recognition of patterns and atypical seasonality that otherwise may not be noticeable with the ‘naked eye’.

Despite its apparent advantages, Higdon discloses that the forecasting algorithm – and others that follow – is not initially expected to fully replace a well-structured process already implemented in a large treasury.

The expectation is that it should provide an additional benchmark when comparing forecasts, he explains. “In particular, it will allow treasurers to focus on areas where the algorithm is predicting something quite different from their traditional methodologies, either helping them to understand an issue with existing processes, or reasons why the algorithm predicts a markedly different cash flow.”

AI has a place in many fields of endeavour, including treasury. As Higdon says, the secret is not to try to resist or defeat it, but to work alongside it. It may be a while before deep learning algorithms become part of the everyday treasury experience, but they are on their way.

Want to read more?

Register to continue reading this article.

Please only use letters.
Please only use letters.
Please complete this field.
Please select an answer.