AI cannot pick stocks
Yesterday, I wrote a bit about the hopes for productivity gains from AI.
A team from Tsinghua University tested various AI models to see if they could outperform the market by trading Chinese stocks from the beginning of 2024 to mid-April 2026. The purpose was not to make profits, but to evaluate the models' ability to utilize public information for stock selection and portfolio management. However, when AI models are trained on all available public information during their training window, they inadvertently gain knowledge of future events.
This means that when evaluating their performance outside their training period, their results appear significantly better than in reality. To address this issue, the researchers created a technique to remove identifying information from the Chinese stocks and eliminate training window effects and other potential distortions. After applying this method, the results were as follows: the AI-managed portfolio outperformed the CSI 300 benchmark index by 85.3%, with Qwen 3.6 Plus achieving an impressive 85.3% return compared to the 36.9% return of the CSI 300 index.
While five out of ten models outperformed the benchmark, the remaining five underperformed. A closer examination of the AI portfolios revealed that the stock picking alpha, or the ability to select winning stocks, was disappointingly low, with only Claude Opus 4.7 managing a meager 0.2% cumulative alpha over the 16-month period.
In contrast, other models demonstrated negative or even catastrophic negative alphas, such as Gemini, DeepSeek, GLM, and Kimi.
Written by urgent.news from Klement on Investing's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
