We developed EmTract, an open-source tool designed to extract investor emotions from financial social media data. We contribute a novel dataset of 10,000 financial social media messages annotated with emotion labels and improve the DistilBERT model by incorporating 4861 tokens, including emojis and emoticons. This augmentation improved the model’s accuracy by over 3 percentage points compared to the standard BERT model, while providing faster inference and reduced computational requirements. Our models and datasets are publicly available to promote broader adoption and further research in financial sentiment analysis. We validated EmTract during the 2021 “meme stock” rally, where it accurately captured spikes in emotions such as anger and disgust following trading restrictions, demonstrating the model’s practical applicability in real-world events. Additionally, heterogeneity tests show that emotions have a stronger impact on smaller, more volatile, and heavily shorted stocks, aligning with established behavioral finance theories. These findings underscore the importance of integrating emotional dynamics into market analysis, particularly for speculative assets. Our contributions represent a significant step forward in understanding the psychological drivers of financial markets and offer practical tools for future research and industry applications.

中文翻译：

---

EmTract：从社交媒体中提取情感


我们开发了 EmTract，这是一种开源工具，旨在从金融社交媒体数据中提取投资者的情绪。我们提供了一个包含 10,000 条带有情感标签注释的金融社交媒体消息的新数据集，并通过整合 4861 个标记（包括表情符号和表情符号）来改进 DistilBERT 模型。与标准 BERT 模型相比，这种增强将模型的准确率提高了 3 个百分点以上，同时提供了更快的推理并降低了计算要求。我们的模型和数据集是公开的，以促进金融情绪分析的更广泛采用和进一步研究。我们在 2021 年的“模因股票”反弹期间验证了 EmTract，它准确地捕捉了交易限制后愤怒和厌恶等情绪的峰值，证明了该模型在现实世界事件中的实际适用性。此外，异质性测试表明，情绪对规模较小、波动较大和严重做空的股票有更大的影响，这与已建立的行为金融学理论一致。这些发现强调了将情绪动态纳入市场分析的重要性，尤其是对于投机性资产。我们的贡献代表了在理解金融市场的心理驱动因素方面向前迈出的重要一步，并为未来的研究和行业应用提供了实用工具。