By Pan Yiannakou, CEO of Swarm Technology
Artificial Intelligence (AI) has advanced exponentially in recent years. It is widely regarded as a must-have tool to help traders distil vast sets of data to make the most profitable decisions quickly, but how does this work when taking longer term positions? Chief Executive of Swarm Technologies, Pan Yiannakou, explores the traps to watch out for when deploying AI in slower, positional trading.
Once a decision has been made to take a position in a market, the initiator must quickly consider a vast array of options to achieve the most profitable outcome – the exact timing; execution methodology; instrument; location and/or intermediary, to name just a few.
Take the decision to buy gold, for example. A trade can be executed immediately, after a few seconds, a few minutes or a few hours. A trader can either ‘lift the offer’ or ‘sit on the bid’; they can set a price target (below) or set a price trigger (above). They could also wait for some other event to trigger the trade. A trade could be made in the cash market, futures, options, as a contract for difference (CFD), or through an exchange-traded fund (ETF). The trader can conduct their business on an exchange, through a broker or directly with a counterparty.
Without restrictions, there are more than a million combinations of ways to gain a long exposure in gold. To add another layer of complexity, these variations are not static and anything from updated broker fees to shifting liquidity to changing counterparties has the potential to exact a significant cost impact.
In an environment where speed is critical and the variations are so vast and constantly moving, the case for deploying AI to solve this data-rich, complex, multi-dimensional problems is clear: AI is powerful enough that not only can it perform well in a dynamic, complex environment, it can also identify genuine opportunities hidden in the data that humans could never recognise.
However, high-frequency trading also has its limitations, particularly issues with liquidity. Sometimes, it is not possible to put to work all the funds required by participants in short term trades. The alternative is slower trading: taking positions that run for hours, days, weeks or longer. Berkshire Hathaway has famously held Wells Fargo, Coca-Cola and American Express stocks for over 25 years. But slower, positional trading presents a particular set of challenges for AI which could immediately and directly hit profitability.
Weaknesses of AI
- Biased data
AI is designed to identify biases, and to adapt to new ones. If historical data contains a bias that is suddenly no longer valid, that can result in poor decisions. Slow trading limits the speed at which AI can adapt. If there is a chaotic shock that changes the bias, damage can be difficult to avoid. Consider Covid-19’s impact on global financial markets. There was the initial reaction, which led to an almost instantaneous shift in bias across all asset classes. By the time AI systems pick up the switch in bias, the damage has been done.
- New price action
AI works well when trained on large amounts of historic data, or when sufficient new data is being generated quickly enough. When markets move in a way they never have done before, there is no historic data for AI to base decisions on. Some events are so rare – like the subprime crisis of 2007-08 – that there may not be enough examples in the past.
- Random clusters
AI is very efficient at finding patterns in large amounts of data. But this can be a weakness. Often, large amounts of numeric data such as price contain random patterns – ‘ghosts’ in the data, or random clusters as they are known. There is a simple test to identify this problem: present an AI trading system with a series of genuinely random prices. Will it find patterns and generate trading signals? The answer is usually ‘yes’. AI is so good at finding patterns, it will find random patterns too. Executing trades based on patterns that do not contain a genuine bias will cause eventual losses due to trading costs.
- Feedback
There is a strong feedback loop that degrades the value of patterns that are easily identified by AI. Consider a specific event that leads to a price gain of 5 per cent in a particular market, for example. If that exact same event occurs in the exact same circumstances again, the participants will all know that it led to a 5 per cent rally the previous time. When the pattern repeats, sellers are likely to sell before the 5 per cent target is hit, not wishing to be the last out. Even with exactly the same trigger and exactly the same starting conditions, the market would behave differently. The more participants that identify the same opportunity, the less effective that opportunity is.
Trade like an ant!
AI trading solutions do not have to be limited to fast trade execution. There are ways of using AI that avoid the pitfalls listed above. Swarm Technology, for example, uses the natural principles of swarm intelligence in its AI solution; this is a form of biomimicry. Swarm does not use AI for pattern recognition or to aid trade execution. It generates a Swarm Matrix, that shares information between all the markets and the systems applied to those markets; in a similar way to that of ants in a colony, who share information and adapt their behaviour accordingly.
While AI does have its drawbacks in slower, positional trading, it is not impossible to overcome the pitfalls.
