Investor interest in AI has so far focused on the big returns yielded by tech stocks. But might the large language models (LLMs) made famous by ChatGPT also help with stock selection? Can AI-powered technology give active investors an edge, and if so what are the implications for passive investors?
Waves of technology have of course repeatedly transformed the financial industry. Machine-driven analysis of historic trading data has helped funds identify undervalued securities and replicable trading patterns since it was pioneered by Jim Simons back in the 1980s.
Sophisticated stock screening apps are readily available to private investors and index funds themselves, from simple trackers to the most exotic thematic ETFs, were made possible by computing power.
But LLMs bring something new: The capacity to simulate human intelligence through the synthesis and interpretation of vast reams of data. Wealth managers are rolling out the first generation of AI systems able to process disparate information to identify winning trades, including financial statements, economic statistics, political developments, market sentiment, container ship movements, corporate website traffic, legal cases and even weather forecasts.
JP Morgan’s custom ‘LLM Suite’ AI tool helps research analysts with writing, idea generation and summarising documents through access to third-party models.
‘Truffle Sniffer’, an AI tool developed by litigation finance specialist Legalist, analyses civil suits to identify attractive investment targets, scanning court records, judges, litigation classes and pre-trial rulings to flag stocks likely to benefit from certain outcomes.
Some firms are introducing ‘agentic’ AI systems that pick stocks according to specified investment styles. Rather than following Warren Buffett in spending five to six hours each day reading financial documents, investors can let AI take the strain of analysing data 24/7 to spot the kinds of ‘diamonds in the rough’ with strong fundamentals and growth potential that the great man laboured to find.
The AI-powered LQAI ETF developed by SoftBank-backed Qraft Technologies can already pick stocks and produce monthly holdings reports autonomously.
Powerful AI systems are not only available to deep-pocketed fund houses. Many private investors now use ChatGPT to find stocks and conduct due diligence. Prompted with questions about a company’s history, current activities, financials and press coverage, the free tool can produce investment narratives for specified companies and indicate how their stocks might perform.
Indeed, right now private investors may be in a position to adopt AI more readily than financial services companies wary of using consumer chatbots running on third-party servers that may compromise client data regulations.
Putting AI to the test
A University of Chicago study, ‘Financial Statement Analysis with Large Language Models’, published earlier this year, offered intriguing academic evidence for ChatGPT’s capacity to beat professional analysts.
The researchers fed the system balance sheets and income statements for more than 15,000 companies dating from 1968 to 2021. Each sheet and statement was stripped of dates and company names, and included only the standard two years of data, with no information given about the longer-term history of the company.
The tool was then prompted to write economic narratives predicting changes in company earnings. Even with this scarce information, the model’s accuracy was 60% compared to 57% for human predictions, implying, as the paper put it, that without “any narrative or industry-specific information, the LLM outperforms financial analysts in its ability to predict earnings changes”.
The research builds on previous studies indicating that computer models can outperform the average analyst when predicting near-term earnings. Extrapolation over relatively short periods plays to the machine’s strengths with its capacity to follow rules and regressions and ignore the biases encouraged or confirmed by the richer information to which humans have access.
More surprising, perhaps, was the model’s capacity to construct alpha-generating long and short model portfolios based on the companies for which it forecast significant changes in earnings, its test portfolios outperforming the broad stock market by 37 basis points (bps) a month on a capitalisation-weighted basis, and 84bps a month on an equal-weighted basis.
The preliminary results indicated that “trading strategies based on GPT's predictions yield a higher Sharpe ratio and alphas than strategies based on other models”. The strong equal-weighted performance showed the model was a competent value investor, adept at forecasting the earnings of smaller stocks.
It is important to note the model was focused on short-term earnings forecasts. Understanding long-term earnings trajectories has more to do with fine judgements about nebulous matters that cannot be simply extrapolated - the structural advantages of businesses against changing economic, cultural, technological and political conditions, the weighing of risk versus reward over long horizons.
AI’s capacity to do this kind of thinking will be a true test of its claims to emulate human intelligence. Concerns about the limitations of LLMs have already been highlighted by their tendency to ‘hallucinate’, to generate false content by overlooking information that would have been spotted by a human.
A recent Stanford University study of the responses generated by three state-of-the-art generative AI models to 200,000 legal queries found hallucinations were “pervasive and disturbing”. When asked specific, verifiable questions about random federal court cases ChatGPT 3.5 hallucinated 69% of the time.
Perhaps such tendencies are inherent features of the technology itself rather than bugs that will simply be ironed out. Generative AI models are probabilistic machines trained to give the most statistically likely response. Like humans, they generalise from the particular, but without human contingencies such as common sense, context, nuance or reasoning. Trained on legacy data it is unclear how models might anticipate and respond to black swan events.
The hard logic of passive investing
Whatever its ultimate capacity, AI is another valuable sword in the experienced active investor’s armoury. It is unclear, though, what implications it has for their passive counterparts.
The logic of index investing holds. The index represents the return of the average investor. By definition, not every fund manager can beat the market. Fund managers compete against each other for limited alpha whether they use AI or not.
As the use of AI becomes pervasive the stocks it highlights will trigger buying orders that will push up their prices and erase the advantage. New technology simply makes the pricing of assets more accurate, making it harder to beat the index.
There may be a welcome consequence for passive investors. As the market becomes ever more efficient funds will rely more on lower fees for differentiation.
Investors prepared to put in the work that successful active investors have always been prepared to do will reap rewards by making intelligent use of AI. For the rest of us, the fundamentals driving the continued success of index investing remain.
Justin Reynolds is a freelance journalist and editor of The Patient Investor blog
