Taxi drivers, stockbrokers and the bloke propping up the bar at your local have all, at one time or another, served as a source of share tips. Now there is a fresh seam of supposed wisdom for retail investors to mine: chatbots such as ChatGPT and Claude.
These artificial intelligence tools, known in the trade as large language models (LLMs), are increasingly being pressed into service by amateur and professional investors alike to generate investment ideas. Yet for all the awe AI has inspired, the jury is still out on whether the machines are actually any good at making money.
Back in 1973, the academic Burton Malkiel argued in his now-famous book that a blindfolded monkey throwing darts at the financial pages of a newspaper could pick a portfolio just as profitable as one chosen by highly paid professionals. His point, the bedrock of the efficient market hypothesis, was that returns on the stock market are essentially random and unpredictable, and that nobody can hold a lasting edge over anyone else.
The notion that LLMs might be superior stock pickers to humans would, of course, blow a hole in that theory. A clutch of start-ups has already set AI to work trading and investing, with markedly mixed results.
According to a recent test run by the US research lab Nof1, six of the eight most popular AI models lost money investing in American technology shares. Anthropic’s Claude Sonnet shed almost 60 per cent of its initial $10,000 (£7,500) stake, while Google’s Gemini gave up more than $5,000. Only two came out ahead: ChatGPT, which made nearly $900, and Elon Musk’s Grok, which roughly broke even.
To the technology’s believers, however, it is only a matter of time before LLMs start besting the very best of Wall Street.
Faizan Ahmad, a former Meta engineer, is co-founder of Rallies, a start-up that uses AI to help people choose shares. His own experiments have thrown up some eyebrow-raising results, with the machines displaying a flash of ingenuity in navigating choppy markets.
Claude, for instance, deftly handled the fallout from the conflict with Iran by rotating out of growth shares and into defence stocks. ChatGPT, meanwhile, plumped for Credo Technology Group, a high-speed connectivity firm, as a likely beneficiary of the global build-out of internet infrastructure around seven months ago. The shares have since climbed by more than 75 per cent.
“No one had heard about that stock, and I hadn’t,” says Ahmad. “That stock was starting to show very early signs of becoming core to Nvidia’s and other players’ infrastructure.
“These models get access to all of the research and can go through entire SEC [US Securities and Exchange Commission] filings. The ability to parse a plethora of information and then find a stock that actually went up quite a lot was amazing.”
Rallies has launched its AI portfolios on a service that lets retail traders copy its trades. It now has $10m of retail money shadowing ChatGPT’s picks and $14m across all of its AI portfolios. And it is not only the small investor taking an interest.
Far from the living rooms and box bedrooms of ordinary punters, the gleaming towers of the City and the professional money men are dabbling too. Algorithmic trading has long been a feature of institutional investing, a subject Michael Lewis brought to wider attention with his 2014 book Flash Boys, but the arrival of LLMs has now piqued the interest of hedge funds.
At Man Group, the world’s largest listed hedge fund, LLMs have already been behind a number of profitable trade ideas.
“We have, right now, several examples where we’ve had an idea proposed by an LLM and have passed it through our diligence process before ultimately being accepted by the investment committee,” says Tushara Fernando, head of data and AI at the firm. “If it can come up with an accepted proposal, that’s fantastic, and then the resulting code goes into production and can trade real money.”
Unlike some of the start-ups letting AI loose unsupervised, Man uses the technology to generate ideas that still require a human stamp of approval. Even so, Fernando says the sheer speed of LLMs lets fund managers kick around far more ideas than they otherwise could.
“[A fund manager] might previously have tested two to three investment ideas a day, for example, modelling different scenarios with the aim of proving out new trades,” he says. “Now they’re able to explore and backtest hundreds in a very short space of time. While they’ve gone for a coffee, the agent’s running a backtest, which uses historical data to evaluate how a potential trade would likely perform under differing conditions.”
Many of the largest hedge funds were early adopters of AI and machine learning long before LLMs went mainstream. Bridgewater Associates, the US fund founded by Ray Dalio, launched a vehicle using machine learning as the primary basis of its decision-making two years ago. In 2018, Two Sigma poached Mike Schuster, an AI specialist, from Google’s Brain team to spearhead its efforts.
Balyasny Asset Management, one of the biggest hedge funds in the US, said recently that 95 per cent of its investment teams were using OpenAI. Agents are set to work analysing and synthesising tens of thousands of documents, from company filings to research notes and earnings reports. The firm has also used AI to monitor and update the probability of mergers and acquisitions completing, and to dissect speeches by central bankers. Balyasny said the technology had slashed the time taken to work out the economic implications of those speeches from two days to 30 minutes.
Unsurprisingly, simply having access to the latest and greatest models is not enough. It is proprietary data that gives the hedge funds their edge. Anthropic announced a partnership with Man Group in February to deploy its Claude model across the firm’s investment process, both to surface new insights from data and to speed up coding tasks.
That, though, presents its own headache for firms such as Man, which must bolt cutting-edge LLMs onto their own highly sophisticated technology stacks.
“LLMs are fantastic at using public knowledge. They may know the last 12 songs on the Taylor Swift album and how to solve a Rubik’s cube, but they don’t know Man Group: our strategies, how much we trade, or our databases or execution platform,” says Gary Collier, chief technology officer at Man Group.
For the largest and most established hedge funds, then, people remain firmly at the controls. Ahmad, who intends to launch a hedge fund through Rallies in due course, is convinced that fully AI-managed funds are not far off.
“We are very bullish that eventually, three or two years down the line, there are going to be hedge funds that are entirely run by AI, that are provided with data and anything the models need, which then go ahead and trade,” he says.
Forget the cabbie’s hot tip. The AI chatbot, it seems, may yet become the next font of all stock-picking wisdom. Whether it proves any more reliable than Malkiel’s dart-throwing monkey is, for now, anyone’s guess.
With AI now driving the lion’s share of global trading volume, the technology’s grip on markets is only tightening. For context on the wider boom, see how Britain’s AI investment hit a record £8.3bn and why Big Short investor Michael Burry is betting $1.1bn against AI stocks.