Elm Wealth投资公司的Jerry Bell、Victor Haghani和James White最近进行的一项研究强调了宏观交易的极端难度，宏观交易涉及根据政治和经济趋势对资产价格走势进行投注。即使模拟中的参与者获得了提前预览次日《华尔街日报》头版的独特优势，人类和人工智能模型也难以避免破产并获得利润。在Elm Wealth的更新实验中（该实验源于2023年一项针对研究生水平金融专业学生的研究），志愿者对标普500指数和30年期国债进行交易。大约有一半的参与者亏损，六分之一的人破产，平均收益仅为3.2%。随后在其网站上举办的在线版游戏（拥有更大的样本量）产生的财务结果甚至更糟。

该实验还测试了领先的人工智能模型，包括ChatGPT、Claude、Gemini和Grok，要求它们扮演管理100万美元虚拟赌注的中年美国投资者。只有Claude和ChatGPT成功实现了盈利，最终的平均资金分别达到260万美元和150万美元，而Grok和Gemini则遭受了亏损。人类和机器面临这些失败的主要原因是根本无法合理地确定下注规模。没有任何AI模型预测股票和债券方向的准确率超过60%，但它们应用的平均杠杆却高达7到12倍。考虑到美国股市的历史波动性，这种过度的杠杆使它们面临灾难性的资本损失风险。

人类交易员也表现出类似的缺陷，无论新闻使价格预测变得多么容易，他们经常在将近三分之一的交易日中采用超过20倍的杠杆，从而承担了过高的风险。相比之下，被邀请参与相同模拟的五位专业宏观交易员全部实现盈利，平均回报率达到130%。虽然他们63%的预测准确率仅略高于AI模型，但他们最关键的区别在于动态的仓位控制。他们根据信心显著调整下注规模，在没有把握时甚至完全不交易。这表明，虽然在拥有远见的情况下预测市场方向很困难，但通过仓位管理控制风险才是宏观交易中更大的挑战。

A recent study by Jerry Bell, Victor Haghani, and James White of Elm Wealth highlights the extreme difficulty of macro trading, which involves betting on asset prices based on political and economic trends. Even when participants in a simulation were given the unique advantage of previewing the next day's front page of the Wall Street Journal, both humans and artificial intelligence models struggled to avoid ruin and secure profits. In Elm Wealth's updated experiment, which evolved from a 2023 study involving graduate finance students, volunteers traded on the S&P 500 index and 30-year Treasury bonds. Roughly half of the participants lost money and one in six went bust, yielding an average gain of just 3.2%. A subsequent online version of the game with a larger sample size yielded even worse financial outcomes.

The experiment also tested leading artificial intelligence models, including ChatGPT, Claude, Gemini, and Grok, by prompting them to manage an imaginary $1 million stake as a middle-aged American investor. Only Claude and ChatGPT managed to turn a profit, finishing with average pots of $2.6 million and $1.5 million respectively, while Grok and Gemini suffered losses. The primary reason for these failures across both humans and machines is a fundamental inability to properly size bets. None of the AI models correctly predicted the direction of stocks and bonds more than 60% of the time, yet they applied high leverage between seven and 12 times. This excessive leverage exposed them to catastrophic capital losses, especially given the historical volatility of the American stock market.

Human traders exhibited similar flaws, frequently taking on too much risk by using leverage exceeding 20 times on nearly a third of the trading days, regardless of how predictable the news made the price movements. In contrast, five expert macro traders invited to play the same simulation all finished in the black, achieving an average return of 130%. While their prediction rate of 63% was only marginally better than that of the AI models, their crucial differentiator was dynamic position sizing. They significantly varied their bets based on confidence, sometimes abstaining entirely. This demonstrates that while predicting market direction with foresight is difficult, managing risk through position sizing is the ultimate challenge in macro trading.

Source: Why macro trading is hard

Subtitle: More difficult than knowing what to buy is how much

Dateline: 6月 25, 2026 03:28 上午