HKU forex AI experiment: Alibaba's Qwen tops profit chart, DeepSeek posts biggest losses

Hong Kong University Business School recently released the results of a notable AI forex trading experiment that connected 10 major large language models to real foreign exchange markets as autonomous trading agents. The six-week experiment was designed to test the practical performance of AI models in financial markets.

Results showed that Alibaba's Qwen achieved the highest profitability among all participating models, while DeepSeek posted the biggest losses over the six-week trading period. The findings have sparked widespread discussion about AI's financial application capabilities.

The experiment revealed two key findings. First, trading activity does not correlate with returns. Models like DeepSeek V3.2 and Claude Opus 4.6 executed over 1,000 trades in six weeks but suffered greater losses due to excessive trading frequency. In contrast, top performers Qwen and Kimi kept their trade counts between 500 and 800, demonstrating better market timing.

Second, high-leverage strategies carry significant risks. Some models employed aggressive leverage that amplified losses during market volatility. Gemini 3.1 Pro Preview also showed substantial losses due to similar issues.

These results provide valuable reference data for AI applications in financial trading. They suggest that while large models can understand market information, knowing more does not automatically translate to better trading. Successful AI trading requires superior risk control, trading discipline, and timing capabilities.

It should be noted that the results may vary under different experimental conditions, including initial capital allocation and trading windows. Different parameters could yield different rankings.

Watch for whether model developers use these insights to improve their models' financial decision-making capabilities, and whether financial institutions move to translate these experimental results into actual trading strategies.