大语言模型交易决策趋同特征：A股情境实验

Convergence Patterns in LLM Trading Decisions: An A-Share Scenario Experiment

摘要: 目的：考察统一政策与账户设定下，通用大语言模型产品的A股调仓输出是否趋同。方法：选取2026年沪深交易规则修订中的两项规定，设置实施首日和标记为实施21天后的两个情境。2026年7月30日对ChatGPT、Claude和DeepSeek在每个条件下各开启5次全新对话，取得30份首次JSON回答，比较方向、目标仓位和配置距离。结果：29份回答通过权重与资金平衡校验。条件A、B分别有14/15、10/15份回答减持提示词中的ST及*ST资产类别，其他三类风险资产均未出现卖出。有效向量的产品内与跨产品平均距离分别为6.0和7.7个百分点，分组表现并不一致。结论：样本呈现风险警示资产类别上的局部趋同，配置幅度和条件反应仍有产品差异。JSON示例可能造成选择性锚定，结果只适用于给定提示框架。

Abstract: Objective: To examine whether general-purpose LLM products converge in A-share reallocation outputs under a common policy and account setting. Methods: Two prompt conditions represented the implementation day and a date labelled as 21 days after implementation of two selected provisions from the 2026 revision of the Shanghai and Shenzhen trading rules. On 30 July 2026, ChatGPT, Claude, and DeepSeek were queried in five fresh conversations per condition, producing 30 first-pass JSON responses. Direction, target weights, and allocation distances were compared. Results: Twenty-nine responses passed the weight and cash-balance checks. In Conditions A and B, 14/15 and 10/15 responses, respectively, reduced the prompt-defined ST and *ST category. None sold the other three risky-asset categories. Mean within-product and cross-product distances were 6.0 and 7.7 percentage points, but subgroup patterns varied. Conclusion: The sample shows local convergence in the risk-warning category, alongside product differences in allocation magnitude and responses to scenario framing. Because the JSON example may have induced selective anchoring, the findings apply only to the specified prompt framework.