Can AI Become a Better Investor Than You?

In M6, the top-performing forecasting teams constructed relatively inefficient portfolios on average, while the top-performing investment teams submitted forecasts of various accuracy levels.

The M6 organisers found most participants were measurably overconfident: they assumed much more investment risk than was justified by the accuracy of their forecasts.

Using headlines published after the model's knowledge cutoff, Lopez-Lira and Tang report that GPT-4 scores significantly predict the subsequent drift in stock returns, especially for small stocks and negative news.

Before trading costs, the annualised Sharpe ratio of Lopez-Lira and Tang's equal-weighted GPT-4 overnight-news long-short strategy declined in each of four subperiods, from 6.54 in 2021Q4 to 3.68 in 2022, 2.33 in 2023 and 1.22 over January-May 2024, which the authors describe as suggestive evidence consistent with LLM adoption improving market efficiency.

At a transaction cost of 20 basis points per round trip, Lopez-Lira and Tang's strategy is unprofitable over their sample period (October 2021 to May 2024); at 10 basis points it still earns over 100% cumulatively.

Across 97 variables shown to predict cross-sectional stock returns, McLean and Pontiff found portfolio returns 26% lower out of sample and 58% lower post-publication, and conclude that investors learn about mispricing from academic publications.

A 2025 working paper by Lopez-Lira, Tang and Zhu finds that LLMs recall exact economic and financial values from before their training cutoff and that instructions to respect historical boundaries do not prevent it.

The arXiv version of Kim, Muhn and Nikolaev's "Financial Statement Analysis with Large Language Models", which claimed GPT-4 outperformed financial analysts at predicting earnings changes, was withdrawn on 2025-02-20 after a co-author identified inconsistencies in the data and analyses.

FINSABER (Li et al., KDD 2026) re-tested two previously reported LLM trading agents (FinMem and FinAgent) over 2004-2024 and 100+ symbols with transaction costs and found previously reported LLM advantages deteriorate significantly.

On 2026-07-16 the Forecasting Research Institute reported that AI models have likely reached parity with superforecasters on ForecastBench, with several models statistically indistinguishable from superforecaster-level accuracy; its superforecaster predictions were last elicited in 2024 and a fresh round is planned for fall 2026.

Cao, Jiang, Wang and Yang (Journal of Financial Economics, 2024) report that an AI analyst trained on corporate disclosures, industry trends and macroeconomic indicators surpasses most analysts in stock return predictions, by a narrow margin per prediction.

In AIMA's September 2025 survey of 150 hedge fund managers running about $788 billion, 95% said they use generative AI in their work.

Cao et al. find the AI analyst wins when information is transparent but voluminous, while human analysts win where institutional knowledge is crucial, such as intangible assets and financial distress.

Cao et al. find that combining human analysts with the AI adds significant value and substantially reduces extreme errors.

Cao et al. find analysts caught up with the machine after alternative data arrived only where their employers had built AI capabilities.

S&P Dow Jones Indices' SPIVA U.S. Scorecard reports that 79% of all active large-cap U.S. equity funds underperformed the S&P 500 in 2025.

Rossi and Utkus (Journal of Financial Economics, 2024), studying previously self-directed Vanguard investors who joined its hybrid Personal Advisor Services in 2015-2017, find that the service increased indexing and reduced home bias, the number of assets held, and fees.

Morningstar's 2026 Mind the Gap study estimates the average dollar in US mutual funds and ETFs earned about 1.2 percentage points a year less than the funds themselves over the ten years to 2025.

Fulkerson, Jordan, Riley and Yan (Financial Analysts Journal, 2026), using Morningstar's 2015-2024 sample, estimate that poor timing costs mutual fund investors about 0.10% a year.

In its April 1999 report on the 1998 crisis, the President's Working Group on Financial Markets concluded that the size, persistence and pervasiveness of the widening of risk spreads confounded the risk management models employed by LTCM and other participants.

The Bank of England's Financial Policy Committee describes correlated AI-driven trading as a potential future risk, and the Financial Stability Board's October 2025 monitoring report says there is still little empirical evidence that AI-driven market correlations affect market outcomes.

ForecastBench's operators extrapolated AI-superforecaster parity to November 2026 on 2026-01-29 and to May 2027 on 2026-03-04.

Of the 38 M6 Global teams that forecast more accurately than the benchmark, only 11 also constructed a better portfolio than the benchmark.

The M6 organisers found that a limited number of teams actually chose to exploit their forecasts to define the weights of their investments.

Man Group describes its AlphaGPT signal-research system in detail but discloses neither how many AI-generated signals trade live nor any returns attributable to them.