One cutting-edge area where AI is used is financial market trading. The banking industry has always been an early adopter of technology, and algorithmic trading is nothing new. As far back as 2019, fully 92% of all foreign exchange trading was done by algorithms rather than humans. For stock trading, the percentage is lower, but still around two-thirds or more of all stock exchange trades in the USA are conducted by algorithms. There is even a draft XML standard for expressing algorithmic order types, called FIX Algorithmic Trading Definition Language (FIXatdl).
I recently interviewed someone at the sharp end of all this, someone who works in production support for a global bank with a major office in London. He explained that his bank used AI in some different ways to help with trading. When an order comes into the bank from a client, it may be to buy or sell a quantity of a certain financial instrument, such as a stock or bond, e.g. “buy 1,000 Google shares” or to sell a particular bond, which will be a code to identify a company (like MSFT for Microsoft) along with the coupon rate and maturity date. Confusingly, there are multiple standards for financial instrument codes, and one AI algorithm analyses the incoming orders and assigns the correct codes to each order. This is a simple classifier, a machine learning algorithm.
The more interesting use of AI in that bank is that of an autonomous trading algorithm, in this case, one based on reinforcement learning. This branch of AI is the one used by Google DeepMind to develop AlphaZero, a chess-playing model, and its relatives AlphaGo (to play the board game Go) and AlphaFold, to predict 3D protein folding, a key area in drug development. In the case of AlphaZero, the AI model was taught only the basic rules of chess, and then left to play against itself millions of times, getting slightly better each time as it learned what approaches and strategies worked and which did not. This program was pitted against the world’s best specialist chess computer program, Stockfish, and across a match of a hundred games, AlphaZero did not lose a single game to Stockfish. This striking demonstration led to the AlphaFold program, which won its inventors the Nobel Prize for Chemistry in 2024.
In the case of this particular trading model, the same approach was used. The model was given access to a live feed of real market data and then made trades, initially just pretend trades, as it learned what trading strategies worked and which did not. Eventually, a couple of years ago, this AI model was turned on and allowed to trade for real. Clients could opt to use either a human-written algorithm or this AI model to execute their trade. The AI is fed with real-time market data and also financial news feeds to help it understand market sentiment. The model can then decide to split a large trade into smaller component trades so as not to alert other traders to what it is doing; other traders, if they knew a large trade was incoming, would be able to take advantage of this and buy or sell stock ahead of the incoming trade. Now over half of this particular bank’s U.S. customers use the AI trader instead of human ones. The AI trading algorithm is not entirely free to make any decisions it likes. There are tight constraints, which ensure that it is compliant with the regulations and does not manipulate the markets in a predatory or risky manner. Examples of that involve risk checks on the model’s orders – for example, if they are too far away from the market price, or their volume is unreasonably large, risking a large impact on the market, its order would not be sent to the exchange.
All this trading happens at blinding speed, and a couple of examples will show how this, in itself, can cause an issue. In one recent case, a particular market feed was briefly down.
It was supposed to receive the average volume available in the opening auction for a given instrument over the last 30 days. The program received incorrect data that presented the historical average volume to be over ten times larger than it actually was. It therefore made the wrong decision and sent far too big a chunk of an order into the opening auction. The model is heavily reliant on its inputs, like current or historical market data, signals (the sentiment information from the media) and therefore if the inputs contain an error, it can have a huge negative impact on the model’s performance. This mistake cost tens of thousands of dollars. A second incident occurred when the exchange was down, and the model did not receive any updates on its orders. In such a case, the model is more unpredictable as suddenly it realises that it is not receiving any fills that it is expecting, which leads it to try to hastily adjust and keep cancelling and resending adjusted orders. This required manual intervention from the bank’s and also the exchange, but there was no impact as the connection to the exchange was down anyway.
There is nothing particularly new here. In 2010, a “flash crash” on US exchanges lasted about half an hour, exacerbated by algorithmic trading. This caused the Dow Jones to lose 9% of its value in a few minutes before recovering. One issue with reinforcement learning and deep learning models is that they are black boxes, meaning that their decision-making process is opaque. The widespread adoption of similar models can cause market participants to act in unison and amplify market volatility, as seen in the 2010 crash. An over-reliance on historical trading data may be an issue for trading models, which can struggle in situations where they are presented with edge cases or novel situations.
Investment banks use a range of AI platforms for their trading. This particular one that I was told about was a proprietary platform developed by the bank itself a few years ago. However, there is a range of commercial AI trading platforms, some marketed at retail investors. Some of these carry out sentiment analysis for stocks, some offer technical trading data and others carry out financial research on companies. Some offer predictive models for stock and foreign exchange trading, market alerts or strategy back-testing. An example of this is the Medallion Fund by the US hedge fund Renaissance Technologies. This quantitative trading fund delivered remarkable returns over an extended period. Between 1988 and 2022 it returned over 39% annually compared to the 10.7% average growth of Standard & Poor‘s 500 index. This compounds to a 90,000 times return to its investors. In other words $1,000 invested in it in 1988 would have returned $90 million. The exact nature of its model is a closely guarded secret, but it uses machine learning and self-learning algorithms, involving statistical pattern recognition and high-frequency trading.
Many hedge funds and investment banks now use AI in various aspects of their operations, from research to sentiment analysis to executing trades, signal generation and portfolio optimisation. Mostly this involves machine learning with neural networks, or reinforcement learning, as in the case of the trading system discussed in my interview, as described above.
Large language models (LLMs) may be used for spotting patterns in unstructured text, such as earnings calls and financial news articles. An example is the BloombergGPT, an LLM trained on finance data, introduced in late 2023, used for producing market reports and trading analysis.
Although there are risks with using AI for trading, as noted, its use is growing and already represents a majority of trading in stocks and the vast majority of trading in foreign exchange. The AI genie cannot be put back in the bottle, so it is up to regulators like the Securities and Exchange Commission to put in place sufficient controls and oversight to minimise the risk of financial destabilisation. Regulations cover areas like risk management, disclosure, fairness, and supervision. However, as more and more trading becomes automated and the speed of execution of large trades is almost instantaneous, it remains to be seen how well these regulations will stand up to extreme market conditions. In the world of trading, milliseconds can mean millions, and those milliseconds belong to machines.
