Robinhood 周三宣布推出新功能，让客户可使用人工智慧聊天机器人进行股票交易，并建立所谓的「agentic trading accounts」。使用者可用 Anthropic 的 Claude Code 与 OpenAI 的 ChatGPT 建构投资组合、把交易策略自动化，并直接下单。这项动作显示券商平台正为散户投资人展开一场新的技术军备竞赛；eToro 早在两个月前就推出 agentic trading，Tiger Global 支持的 Public 也已推出可由投资人提示执行交易的 agents。

自 ChatGPT 于 2022 年推出以来，AI 聊天机器人已被用来影响散户的投资决策，但关于 AI 生成建议究竟改善或恶化报酬的学术研究仍很少。去年 10 月，Nof1 让 6 个领先的大型语言模型各自拿到 10,000 美元，在 Hyperliquid 这个衍生品交易平台上进行交易，目标是在以数小时为决策周期的中低频交易中最大化利润。结果显示，这些模型只在 32 组结果中有 6 次以获利收场，且偏误差异明显，其中一些偏好短线押注，并且对看似微不足道的提示词变化高度敏感。

Stephen Mackintosh 认为，成功的前沿模型还需要 Jane Street、Hudson River Trading 与 Citadel Securities 等交易公司用来训练模型的传统资料集；他警告，现成的 LLM 并不适合直接产生报酬，只是一般性的机率模型。投资人倡议团体 Better Markets 的政策主管 Amanda Fischer 则指出，典型散户并不会因为更多交易而受益，但 Robinhood 的手续费收入显然会受益。Robinhood 也提醒，AI 驱动策略在某些市场条件下可能表现不佳、变动迅速，且可能难以即时监控或停止。

Robinhood said on Wednesday it is launching a feature that lets clients use AI chatbots for share trading through new “agentic trading accounts.” Users will be able to use Anthropic’s Claude Code and OpenAI’s ChatGPT to build portfolios, automate trading strategies and place orders. The move highlights a new arms race among brokerage platforms to offer retail investors more technology; eToro introduced agentic trading two months earlier, and Tiger Global-backed Public has also rolled out agents that investors can prompt to execute trades.

Since ChatGPT launched in 2022, AI chatbots have been used to shape retail investors’ decisions, but there is still little academic work on whether AI-generated recommendations improve or worsen returns. Last October, Nof1 gave 6 leading large language models $10,000 each to trade on Hyperliquid, a derivatives platform, with the goal of maximising profit through mid-to-low frequency trading, where decisions are made over hours. The experiment found that the models finished in profit only 6 times across 32 sets of results, with notable bias differences, including some that favored short-term bets and were highly sensitive to seemingly trivial prompt changes.

Stephen Mackintosh said successful frontier models will still need the traditional datasets used by firms such as Jane Street, Hudson River Trading and Citadel Securities to train their models; he warned that out-of-the-box LLMs are not suited to generating returns and are only generic probabilistic models. Amanda Fischer, policy director at Better Markets, said the typical retail investor does not benefit from more trading, while Robinhood’s fee income certainly does. Robinhood also warned that AI-driven strategies may perform poorly under certain market conditions, move quickly, and be difficult to monitor or stop in real time.