AllianceBernstein ESG AI Yields 25% Excess Return Shorting Bad News

联博ESG AI通过做空负面消息斩获25%超额收益 基于分层标注对开源语言模型进行微调，其表现超越GPT-4，并通过在出现高置信度负面ESG争议后做空相关股票，实现了1.04的信息比率。 Guan指出，零样本大语言模型提示在处理复杂金融任务时极为脆弱，并提出了基于SASB标准的结构化三阶段分层标注框架。通过在这一专家数据集上微调Llama 3 8B等开源模型，团队实现了77%的准确率，超越了GPT-4等通用模型。针对标普500指数成分股的量化事件研究和回测表明，负面ESG新闻能够持续预示股票表现不佳；针对高置信度负面情绪的做空策略实现了超过25%的年化超额收益，信息比率达1.04。最后，Guan概述了后续利用强化学习在不同持有周期内动态调整情绪信号的研究规划。 现成大语言模型的零样本提示对于关键任务型量化金融来说过于脆弱，因为微小的空格或格式改动就可能导致分类准确率剧烈波动。 (2:20) 精准的金融自然语言处理需要结构化的三步递进流程：识别SASB ESG主题相关性、确立对公司的直接影响，以及判定情绪走向。 (2:20) 该研究使用了包含508家标普500指数成分公司的12万篇Perigon新闻报道（共21.4万对实体-文章配对）的基础数据集，其中5,250篇文章由接受过SASB培训的人工标注员进行了标注。 (6:23) ESG新闻呈现出结构性不对称，负面事件与正面事件的比例约为2比1，因为正面公告多为泛化的公关宣传，而负面新闻则涉及具体的争议事件。 (13:40) 在领域专业标注数据上微调如Llama 3 8B Instruct等开源模型可达到77%的准确率，超越了包括GPT-4在内的规模更大的专有模型。 (13:40)

Che Guan of AllianceBernstein demonstrates that off-the-shelf large language models and zero-shot prompting are too brittle for ESG-driven quantitative investing, whereas fine-tuning open-source models on hierarchically structured annotations produces statistically significant alpha signals. Prompt Fragility and Financial NLP Complexities (2:20) Off-the-shelf models struggle with financial domain nuances, particularly distinguishing incidental entity mentions from direct operational impacts. In an interactive test using an article referencing Liberty Mutual and Allstate, adding a boilerplate mention of an unrelated firm caused Gemini to misclassify neutral context as an Environmental, Social, and Governance (ESG) positive impact. Guan stresses that real-world financial text evaluation demands nested logic: verifying whether an article addresses a relevant ESG topic, establishing if the event directly impacts the target company, and assessing the specific tone of that impact. Relying on prompt engineering to navigate these layers proves unreliable. Citing empirical benchmarks on Llama 2 across demographic and professional classifications, Guan notes that altering minor punctuation—such as changing prompt formatting from a colon to a space, or adding a single trailing space—caused classification accuracy to swing wildly from 0.36% to 50% and 75%. Such extreme sensitivity makes zero-shot prompting unsuitable for quantitative investment workflows. Hierarchical SASB Annotation and Skewed Distributions (9:23) To establish consistent ground-truth data, the research team analyzed 120,000 Perigon news articles published between September 2022 and September 2023 across 508 S&P 500 companies, capping coverage at five firms per article to yield 214,000 evaluation pairs. Human annotators labeled a sample of 5,250 articles using a three-step sequential framework derived from Sustainability Accounting Standards Board (SASB) standards: determining SASB relevance, identifying direct firm impact, and categorizing sentiment polarity. The resulting distributions revealed a structural asymmetry: negative ESG news appeared at double the volume of positive news across both broad and strict signal definitions. Under the broader signal, 69% of articles were neutral, 15% negative, and 8% positive. In the strict core signal, 84% were irrelevant, 9% negative, and 4% positive. Guan explains this imbalance by noting that positive corporate releases are frequently vague public-relations boilerplate, whereas negative ESG events recount concrete controversies that yield clearer sentiment classifications. Fine-Tuned Open Models Outperform Proprietary Benchmarks (13:40) Rather than relying on closed-source foundation models via API, the team fine-tuned open-source architectures on their expert-annotated ESG dataset. Testing encompassed Mistral 7B and Llama 3 8B across both base and instruct variants, compared against zero- and few-shot implementations of GPT-3.5 Turbo, GPT-4o mini, and GPT-4. Fine-tuning delivered substantial gains in macro F1 scores, with the fine-tuned Llama 3 8B Instruct model reaching a peak overall accuracy of 77%. According to Guan, domain-specific supervised fine-tuning allowed the 8-billion-parameter open model to surpass proprietary giants like GPT-4 in both overall accuracy and scoring consistency. Event Studies and Alpha Signal Performance (16:46) Using the fine-tuned Llama 3 8B Instruct model, the team evaluated 1-day forward excess returns relative to the S&P 500 across the full 214,000-article dataset. News articles classified as negative generated an average 1-day excess return of -5 basis points per day, while an extreme negative cohort—isolated using high model confidence—produced -8 basis points per day, with both demonstrating statistical significance. Aggregating daily company-level sentiment through majority voting, a backtest of a strategy shorting the extreme negative cohort achieved over 25% excess returns over the one-year evaluation period from an initial $1 capital base. The strategy achieved an Information Ratio of 0.65 for the standard negative cohort and 1.04 for the extreme negative cohort, exceeding the typical 0.50 quantitative threshold for standalone investment signals. To resolve the issue of temporal decay—where human-assigned sentiment can generate positive short-term returns that dissipate over longer holding periods, such as ExxonMobil's March 2026 headquarters relocation—Guan outlines upcoming work using Reinforcement Learning. By using multi-month forward returns as a reward function, the team aims to dynamically calibrate sentiment models to align with specific multi-month investment horizons.

联博（AllianceBernstein）的Che Guan指出，通用大语言模型直接用于ESG新闻情绪分析时极易因提示词脆弱性而失效；通过专业框架构建层次化标注并微调开源小模型，能够在量化回测中产生具有统计显著性的超额收益。 提示工程脆弱性与嵌套逻辑缺陷 (2:20) 在金融和ESG分析场景中，直接使用通用大语言模型（如Gemini）进行零样本提示往往会产生错误分类。Guan指出，在新闻末尾仅附带提及某家公司时，模型会错误地将非核心提及归类为积极情绪，忽视了其实际应为中性。现实中的金融自然语言处理包含多层嵌套依赖：首先必须判断文本是否属于相关ESG议题，其次确认该议题是否对特定公司造成实质影响，最后才能评估该影响的情绪倾向。 通用模型的零样本提示对格式细微变动极为敏感。Guan援引Llama 2在四分类任务中的实验表明，将提示格式从“Passage: <text>\nAnswer: <class>”（冒号后无空格）改为替换冒号为空格，准确率从0.36%大幅跃升至50%；而仅在冒号后补入一个空格，准确率便升至75%。这种剧烈波动表明，单纯依靠提示工程无法满足量化投资对稳定性和鲁棒性的高标准要求。 SASB框架下的三阶段层次化标注 (9:23) 为建立稳健的训练基准，Guan的团队联合范德堡大学数据科学研究所，基于Perigon新闻数据流抓取了2022年9月至2023年9月期间覆盖标普500指数中508家公司的12万篇新闻。由于单篇报道常涉及多家公司，研究将单篇关联上限设为5家，构建了包含214,000个“文章-公司”对的评估集。新闻覆盖呈长尾分布：约200家企业全年报道少于50篇，而约100家核心企业年报道达250篇左右。 针对抽取的5,250篇文章，团队依据可持续发展会计准则委员会（SASB）的五大核心维度（环境、社会资本、人力资本、商业模式与创新、领导力与治理），实施了严格的三阶段顺序标注流程。标注员必须依次判定：文章是否主要聚焦于SASB议题、该议题是否对目标企业构成主要或次要的实质运营影响，以及最终的影响极性。基于此流程，团队构建了广义与核心两种ESG信号，有效避免了将公关陈词与附带提及混淆为有效投资信号的问题。 数据非对称性与开源模型微调优势 (13:40) 人工标注的数据集呈现出显著的结构性失衡：负面新闻的体量达到正面新闻的两倍。在广义信号中，中性新闻占69%，负面占15%，正面仅占8%，无关占8%；在更为严格的核心信号中，无关占84%，负面占9%，正面仅占4%。Guan分析指出，企业正面公告多属于宽泛的公关宣发，缺乏具体增量细节，而负面ESG事件通常涉及具体的争议事件，标注员和模型均能以更高置信度进行识别。 在模型评测中，针对特定领域微调后的开源模型表现超越了闭源通用大模型。经指令微调的Llama 3 8B Instruct模型在测试集上达到了77%的综合准确率。微调后的8B参数开源模型在宏观F1值和准确率上击败了采用零样本或少样本提示的GPT-3.5 Turbo、GPT-4o mini以及GPT-4，证明在受限任务中领域专项微调比盲目调用庞大通用模型更具性价比和可复现性。 事件研究与量化回测Alpha表现 (16:46) 研究团队使用微调后的Llama 3 8B Instruct对全量214,000个样本对进行推理，并通过事件研究法测试次日开盘买入、后日平仓的1日超额收益（相对标普500基准）。在文章级别，被判定为负面ESG新闻的标的平均产生每天-5个基点的超额收益，而模型高置信度的“极端负面”组超额收益达到每天-8个基点，且统计检验显著。 在将单日多家媒体提及通过多数投票聚合为公司级日常信号后，团队进行了为期一年的多空模拟组合回测。以1美元初始资金做空极端负面组，一年内产生的累计超额收益超过25%。从风险调整收益来看，普通负面组的信息比率（Information Ratio, IR）为0.65，而极端负面组的IR达到1.04，大幅超过量化投资界通常采用的0.50独立有效策略门槛。 Guan指出静态人工标注的核心局限在于存在时间衰减：适合次日交易的正面情绪信号，在扩展至一个月或三个月持有期时可能完全失效甚至反转（如埃克森美孚迁址案例）。联博团队未来的研究方向是引入强化学习（RL），将多月前瞻收益作为奖励函数，动态调整情绪分布，使大语言模型输出能够自适应匹配不同投资期限。

## Background 该视频由频道 Ai4 于 2026 年 8 月 19 日在 YouTube 平台发布，记录了联博资产管理公司（AllianceBernstein）首席数据科学家 Che Guan 的演讲。视频简介指出，研究团队利用大语言模型与自然语言处理技术从实时新闻中提取 ESG 洞见，并通过专家标注数据微调模型、开展事件研究与量化回测，证实了 ESG 负面情绪信号在预测股票跑输大盘和增强 Alpha 生成方面的显著价值。 ## TL;DR 演讲首先通过生动的实验揭示了单纯依赖提示词工程在复杂多层 ESG 语义分析中的脆弱性，进而介绍了团队基于 SASB 框架与 5,250 篇专家三步标注样本构建的高质量数据集。实验表明，经微调后的 8B 开源模型（如 LLaMA-3 8B Instruct）在 ESG 分类准确率和情绪抽取上大幅跃升，并成功超越了 GPT-4 等闭源巨头。基于标普 500 成分股的单日交易回测显示，负面及极度负面 ESG 情绪与股票未来跑输大盘高度相关，极度负面组合的一年超额收益达到 -25%（信息比率达 1.04）。最后，主讲人展望了通过强化学习以未来实际收益作为奖励函数，来解决多周期信号衰减与标签对齐的进阶方案。 ## Atmosphere & style 整场演讲氛围严谨专业且不失幽默接地气，主讲人开场以“挡在大家吃午餐路上”的自嘲迅速破冰并抓牢全场注意力。现场互动中展示的 Prompt 脆弱性案例设计巧妙，引发听众对大模型落地金融业务深层挑战的共鸣与思考。叙述节奏紧凑明快，将硬核的数据工程、模型评测指标与金融量化回测曲线有机融合，充满了顶级资管机构探索前沿 AI 生产力时务实求真的技术感染力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=4s) 联博（AllianceBernstein）与范德堡大学合作，探索通过微调大语言模型（LLM）从实时新闻中提取 ESG 情绪信号以赋能投资实践并获取 Alpha。 - [00:52](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=52s) ESG 情绪分析是一个包含“ESG 相关性”、“企业影响度”与“情绪倾向”的多层嵌套任务，仅依赖提示词工程（Prompt Engineering）极其脆弱且不可靠。 - [07:08](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=428s) 研究基于 SASB（可持续发展会计准则委员会）的五大核心支柱制定行业标准，构建了覆盖标普 500 指数 508 家成分股的 21.4 万组新闻样本库。 - [09:23](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=563s) 设计了严谨的三步递进式专家标注流程（ESG 相关性判定 → 企业影响度评估 → 情绪导向划分），完成了 5,250 篇高质量基准数据集的标注。 - [09:23](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=563s) 标注数据显示负面 ESG 新闻数量约为正面新闻的 2 倍，原因在于负面违规/诉讼事件通常非常具体明确，而正面 PR 宣传往往偏向泛化的宏观表述。 - [14:37](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=877s) 经过 5,000 条专家数据微调后，8B 参数级的轻量开源模型（如 LLaMA-3 8B Instruct 达 77% 准确率）在 ESG 情绪识别任务上全面超越了 GPT-4 等通用闭源大模型。 - [17:26](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1046s) 事件研究表明负面 ESG 新闻具有显著预测力：次日普通负面组超额收益为 -5 bps，极度负面组恶化至 -8 bps，模型置信度越高负向预警效果越强。 - [19:22](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1162s) 按公司多数投票法聚合后进行一年期回测，极度负面组累计录得 -25% 超额收益（融券做空可获利 25%），信息比率（IR）高达 1.04，展现出优异的量化因子价值。 - [22:41](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1361s) 新闻情绪标签存在跨时间周期衰减问题，单日短线有效的正面情绪在中长期（如 3 个月）可能完全回归中性，导致固定的人工标注无法直接复用。 - [22:41](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1361s) 团队提出利用强化学习（RL）框架，将不同持仓周期的实际超额回报作为奖励函数反向微调大模型，从而无需重复高昂的人工标注即可实现跨周期的信号对齐与动态校准。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=4s) 演讲开场与研究背景 主讲人 Che Guan（联博资产管理公司 AllianceBernstein 首席数据科学家）在开场时幽默地调侃道，自己深知自己正处在大家和午餐之间，“我压力很大，因为我知道我正挡在大家吃午餐的路上”，因此会严格控制时间，争取准时结束分享，不让大家排长队。随后，他切入正题，介绍了本次关于如何为可持续投资（ESG）优化大语言模型（LLM）的研究。他坦言这项工作并没有堆砌极其复杂的模型架构，但其中提炼出的洞见能够切实赋能投资实践。该项目是他与同事 Andrew Chen 以及来自范德堡大学数据科学研究所（Data Science Institute of Vanderbilt University）的三位学生共同完成的合作成果。 ```mermaid flowchart LR accTitle: ESG大语言模型投资研究整体框架 subgraph Data["数据准备与标注"] D1["新闻数据收集<br/>(Paragon 21.4万样本)"] --> D2["SASB 框架定义"] D2 --> D3["专家三步序列标注<br/>(5,250 篇样本)"] end subgraph Modeling["模型微调与评测"] D3 --> M1["构建宽泛/核心 ESG 信号"] M1 --> M2["开源与闭源模型对比<br/>(LLaMA-3 8B / GPT-4)"] M2 --> M3["领域微调显著超越基座"] end subgraph Backtest["投资回测"] M3 --> B1["对21.4万篇新闻打分"] B1 --> B2["次日交易策略回测<br/>(负面情绪与跑输强相关)"] end ``` --- ### [00:52](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=52s) 现场互动与 Prompt 脆弱性陷阱 为了直观展现大模型在 ESG 语义分析中的挑战，Che Guan 在现场做了一个互动测试。他展示了一篇 2023 年 7 月 13 日的新闻标题，并现场询问观众对该新闻情绪（Positive / Negative / Neutral）的直觉判断。随后，他针对新闻中提及的两家保险公司（Liberty Mutual 与 Allstate）提问，询问该新闻对这两家公司的影响是正面还是负面。在调用大模型（如 Gemini）进行解析时，模型判定两者均为负面影响。 紧接着，Che Guan 展示了一个巧妙的测试技巧：他在文章末尾随手添加了一行“来自 XYZ 公司”。当他再次询问这篇新闻对 XYZ 公司在 ESG 层面是否有影响时，直接使用 Gemini 会错误地得出“正面影响”的判断；而现场观众的判断则是“中性/无关”。Che Guan 指出：“我在文章里加上我的名字，并不代表这篇新闻在 ESG 层面会对我有影响。”这个现象揭示了 ESG 情绪分析的核心本质是一个**多层嵌套的复杂问题**： 1. **ESG 相关性**：新闻是否涉及所关注的 ESG 主题？ 2. **企业相关性**：若涉及该主题，是否对目标企业产生实质影响？ 3. **情绪导向**：这种影响在 ESG 视角下究竟是正面还是负面？ Che Guan 进一步引用了一篇几年前基于开源模型 LLaMA 2 进行四分类任务（性别、职业、种族、宗教）的学术文献，以说明纯粹依赖 Prompt Engineering（提示词工程）的极端脆弱性。在该实验中： - 若输入格式为 `passage:\nanswer:`，LLaMA 2 的分类准确率仅有微不足道的 0.0036（0.36%）； - 仅仅把冒号替换为空格 `passage \n`，分类准确率瞬间飙升到 0.50（50%）； - 若在冒号后面单纯增加一个空格 `passage: \n`，准确率进一步跃升至 0.75（75%）。 这种现象通常源自模型开发商（如 Meta）在自身微调数据时所采用的模板格式。Che Guan 总结道，提示词在大多数情况下虽然可用，但绝无法在所有场景下提供稳健、可靠的输出。为此，联博团队确立了本项研究的四大核心贡献： - 建立了涵盖 ESG 相关性、公司影响度和情绪倾向的专家标注数据集； - 微调了多个大语言模型以生成精准的 ESG 情绪信号； - 将微调模型生成的 ESG 情绪与真实金融市场数据进行历史回测； - 证实了负面 ESG 新闻情绪与股票未来跑输大盘（Underperformance）之间存在强烈的负相关性。 --- ### [07:08](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=428s) 数据集构建与 SASB ESG 框架 在数据准备阶段，研究团队从数据供应商 Paragon 抓取了 2022 年 9 月至 2023 年 9 月整整一年的海量新闻，涵盖 120,000 篇独立新闻报道，覆盖了标普 500 指数（S&P 500）中的 508 家成分股公司。由于单篇新闻可能涉及多家企业，团队将单篇新闻关联的目标公司数量限制在最多 5 家，最终构建了包含 214,000 个“新闻-公司”配对的基准分析样本库。 从新闻分布特征来看： - 约有 200 家公司在这一年内的相关报道少于 50 篇； - 头部约 100 家大型企业的相关报道达到 250 篇左右，平均下来几乎每个交易日就有一篇报道，展示出良好的数据覆盖度。 为了让后续标注具备严格的金融与行业标准，团队在标注前安排专家深入研读了**可持续发展会计准则委员会（SASB）**的五大核心 ESG 支柱： 1. **环境（Environment）** 2. **社会资本（Social Capital）** 3. **人力资本（Human Capital）** 4. **商业模式与创新（Business Model & Innovation）** 5. **领导力与治理（Leadership & Governance）** --- ### [09:23](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=563s) 专家三步序列标注与两大情绪信号定义 为了高效且高质量地完成人工标注，团队从数据集中跨 79 个细分行业随机抽取了 5,250 篇新闻样本，并设计了一套**三步递进式标注流程（Sequential 3-step Annotation）**： ```mermaid flowchart TD accTitle: 三步递进式标注流程 Step1["第1步：ESG 相关性判定<br/>(Major / Minor / No)"] --> Cond1{"是否与 ESG 相关？"} Cond1 -- "No" --> EndIrr["标记为 Irrelevant（无关）"] Cond1 -- "Major / Minor" --> Step2["第2步：企业相关性/影响度<br/>(Major / Minor / No)"] Step2 --> Cond2{"是否对企业有实质影响？"} Cond2 -- "No" --> EndIrr Cond2 -- "Major / Minor" --> Step3["第3步：ESG 情绪语调<br/>(Positive / Negative / Neutral)"] ``` - **第一步（ESG 相关性）**：判断文章是否聚焦于 SASB 五大主题之一。若文章核心探讨 ESG 主题则标为 Major；若仅略微涉及则标为 Minor；完全无关则标为 No。 - **第二步（企业影响度）**：强依赖于第一步。若第一步为 No，直接标记为 Irrelevant；若第一步为 Major/Minor，则评估新闻对该特定企业是否具有重大影响或强相关性，分为 Major、Minor 和 No。 - **第三步（情绪导向）**：强依赖于前两步。只要前两步中任一步被判定为不相关，直接标为 Irrelevant；只有当新闻既符合 ESG 主题又对企业具有实质影响时，才评估其情绪偏向（Positive、Negative 或 Neutral）。 基于这三层标签，团队衍生构建出两类颗粒度不同的投资情绪信号： | 信号名称 | 正面（Positive）判定条件 | 负面（Negative）判定条件 | 中性（Neutral）判定条件 | |:--- |:--- |:--- |:--- | | **宽泛信号（All ESG News with Firm Impact）** | ESG $\in$ {Major, Minor}<br/>Firm $\in$ {Major, Minor}<br/>Sentiment = Positive | ESG $\in$ {Major, Minor}<br/>Firm $\in$ {Major, Minor}<br/>Sentiment = Negative | ESG $\in$ {Major, Minor}<br/>Firm $\in$ {Major, Minor}<br/>Sentiment = Neutral | | **核心信号（Core ESG News with Firm Impact）** | ESG = Major<br/>Firm = Major<br/>Sentiment = Positive | ESG = Major<br/>Firm = Major<br/>Sentiment = Negative | ESG = Major<br/>Firm = Major<br/>Sentiment = Neutral | 从 5,250 篇标注样本的分布特征来看： - 在**宽泛信号**中：中性（Neutral）占 69%，负面（Negative）占 15%，正面（Positive）占 8%； - 在**核心信号**中：无关（Irrelevant）占 84%，负面占 9%，正面占 4%； - **关键洞察**：在两种信号定义下，**负面新闻样本量均为正面新闻的 2 倍左右**。Che Guan 解释道，这是因为企业的正面公关宣传往往偏向泛泛而谈的宏观表述（Generic），而负面事件（如违规、污染、诉讼）通常非常具体（Specific），更容易被识别和标注。 --- ### [14:37](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=877s) 模型微调效果对比与回测机制设置 在模型训练与验证实验中，研究团队横向对比了 4 款开源模型与 3 款主流闭源商业模型： - **开源模型**：Mistral 7B Instruct、Mistral 7B General、LLaMA-3 8B Instruct、LLaMA-3 8B General； - **闭源商业模型**：GPT-3.5 Turbo、GPT-4o mini、GPT-4。 评测对比了开源模型在**仅预训练（Pre-trained / Zero-shot 推理）**与**使用 5,000 条专家标注数据微调（Fine-tuned）**后的表现，评测指标涵盖各类别的 F1 Score（Negative、Neutral、Positive、Irrelevant）及整体准确率（Overall Accuracy）。 实验得出了两项关键结论： 1. **微调带来质的飞跃**：以 LLaMA-3 8B Instruct 为例，微调后模型的整体准确率跃升至 **77%**，显著优于未经微调的基座模型； 2. **小模型超越大巨头**：经过专业标注数据微调后，8B 参数量级的轻量级开源模型在 ESG 任务上的表现**直接超越了 GPT-4 等闭源大模型**。该结论在宽泛信号和核心信号两套体系上均得到了一致复现。 最后，Che Guan 阐述了实证回测的策略构建方式：团队使用微调后的 LLaMA-3 8B Instruct 模型对全量 214,000 篇新闻进行批量打分。一旦模型识别出针对某家公司的情绪信号，策略便会在次日开盘执行对该公司的相应交易，以检验情绪信号在二级市场的真实 Alpha 生成能力与预测有效性。 ### [17:26](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1046s) 文章级情绪信号与单日超额收益测试 为了评估大语言模型提取的情绪信号对股票走势的预测能力，团队设计了一套基于事件研究的测试框架。具体做法是根据新闻发布的时间，在次日建仓并在后天平仓，以此计算每篇新闻对应标的相对于标准普尔500指数（S&P 500）的单日超额收益（Excess Return）。接着，团队将所有样本划分为五个不同的组别（Cohorts），包括被标注为 ESG 相关的样本组、正面情绪组、负面情绪组等；其中还特别从负面情绪中剥离并构建了一个置信度更高的“极度负面”（Extreme Negative）子组。 通过对各组别超额收益的均值进行统计检验，结果展现出了极其显著的不对称性： - **普通负面组**：标的股票在新闻发布后次日的平均超额收益约为 **-5 个基点（bps）**。 - **极度负面组**：单日平均超额收益进一步恶化至 **-8 个基点（bps）**，且基于 t 统计量的 p 值表现出高度的统计显著性。 - **结合公司重大影响的 ESG 负面新闻**：当大模型同时判断负面 ESG 事件对公司具有实质性影响时，负面组和极度负面组同样录得了非常显著的负向超额收益。 ```mermaid flowchart TD accTitle: ESG情绪信号生成与投资回测流程 subgraph S1["数据处理与聚合"] A["实时新闻流与大模型分析"] --> B["文章级 ESG / 情绪分类"] B --> C["每日公司级多数投票聚合"] end subgraph S2["策略构建与回测"] C --> D["单日领先交易策略 (T+1 建仓 / T+2 平仓)"] D --> E{"情绪分组 (Cohorts)"} E -->|"极度负面组"| F["1年超额收益: -25% (做空获利 +25%)<br/>信息比率 (IR) = 1.04"] E -->|"普通负面组"| G["1年稳健下行<br/>信息比率 (IR) = 0.65"] end ``` --- ### [19:22](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1162s) 公司级情绪信号聚合与回测表现 在实际交易中，单一公司在同一天内往往会被多家媒体报道，无法直接针对单篇文章频繁交易。因此，研究团队通过**多数投票机制（Majority Vote）**将单日内与某家公司相关的所有 ESG 新闻聚合为唯一的公司级每日情绪得分。随后，团队基于这一整合信号开展了回测：若某公司当天被评定为特定情绪，系统将于次日买入并在随后一个交易日平仓。 以 1 美元初始资金进行为期一年的回测，结果令人瞩目。图表中代表极度负面组的浅蓝色曲线呈现持续单边下跌态势，一年下来累计录得 **-25% 的超额收益**。这表明，"如果能够对该组合进行融券做空，投资者将能从中直接斩获超过 25% 的额外超额收益"。 从量化评价指标来看，这些负面情绪信号展现了极高的投资价值： - 在第一个情绪信号体系中，普通负面组的信息比率（Information Ratio, IR）达到了 **0.65**，极度负面组的信息比率更是高达 **1.04**。 - 在业界量化实践中，信息比率一旦超过 0.5 便被视为极具alpha潜力的优质因子，而本研究中两套信号体系的负面与极度负面组均稳稳跨过了 0.5 的门槛。 - 这一实证研究清晰地表明，经过领域专家标注数据微调的大语言模型不仅能精准捕捉语境与细微语义，其生成的负面 ESG 情绪信号更对股票未来的跑输风险具有强有力的预警作用。 --- ### [22:41](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1361s) 跨时间周期的挑战与强化学习对齐展望 由于高质量的人工专家标注成本极其昂贵，研究团队进一步思考：当投资周期从“单日领先（1 day ahead）”拉长到“一个月（1 month ahead）”甚至“一个季度（1 quarter ahead）”时，原有的人工情绪标签是否依然有效？ 以 2026 年 3 月 10 日埃克森美孚（ExxonMobil）宣布将法定总部迁往德克萨斯州（Texas）为例： - 在消息公布的次日，其股价跳涨了 **2%**； - 消息公布后一个月内，累计涨幅扩大至 **7%**； - 然而放眼到三个月的维度，该事件带来的超额涨幅基本全部回吐，回归中性。 这表明，人工标注的“正面”标签在单日交易中非常有效，但若直接套用于三个月的中长期持仓，该标签就会彻底失效。为了在不重复进行昂贵人工标注的前提下复用这些数据，团队提出了基于**强化学习（Reinforcement Learning, RL）**的对齐方案： - **核心思路**：将选定时间周期（如三个月后）的实际超额回报作为强化学习的奖励函数（Reward），反向微调大语言模型的输出概率分布。 - **动态修正**：当模型针对某类事件预测为“正面”，但中长期收益回归中性时，算法会利用奖励信号施加惩罚，自动将模型在三个月预测目标下的输出校准为“中性”。 演讲者最后表示，关于利用强化学习对齐跨周期投资信号的最新研究成果，将在来年的大会上做进一步分享。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=4s) 我压力很大，因为我知道我正挡在大家吃午餐的路上 — Che Guan > [00:52](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=52s) 我在文章里加上我的名字，并不代表这篇新闻在 ESG 层面会对我有影响。 — Che Guan > [19:22](https://www.youtube.com/watch?v=A1FLWhmEJeM&t=1162s) 如果能够对该组合进行融券做空，投资者将能从中直接斩获超过 25% 的额外超额收益 — Che Guan ## Mentioned Che Guan, Andrew Chen, AllianceBernstein, Ai4, Data Science Institute of Vanderbilt University, Paragon, Liberty Mutual, Allstate, ExxonMobil, SASB, LLaMA 2, LLaMA-3 8B Instruct, LLaMA-3 8B General, Mistral 7B Instruct, Mistral 7B General, GPT-3.5 Turbo, GPT-4o mini, GPT-4, Gemini, S&P 500 ## Video description Che Guan - Principal Data Scientist, AllianceBernstein We use Large Language Models (LLMs) and Natural Language Processing (NLP) to extract Environmental, Social and Governance (ESG) insights from real-time news, creating an expert-annotated dataset to evaluate ESG classification, firm relevance, and sentiment. Our fine-tuned models outperform pre-trained ones in ESG detection, firm impact, and sentiment analysis. Furthermore, in-context learning does not improve performance, indicating optimal tuning. Event studies and backtests show that our sentiment signals predict underperforming stocks, with higher model confidence in negative sentiment correlating to worse outcomes. These findings emphasize the value of fine-tuning models with expert-annotated data, and leveraging ESG sentiment signals to generate investment insights from qualitative data to enhance alpha generation. [Source](https://www.youtube.com/watch?v=A1FLWhmEJeM)

[0:04] Che Guan: 我压力很大，因为我知道我正挡在大家吃午餐的路上 [0:52] Che Guan: 我在文章里加上我的名字，并不代表这篇新闻在 ESG 层面会对我有影响。 [19:22] Che Guan: 如果能够对该组合进行融券做空，投资者将能从中直接斩获超过 25% 的额外超额收益

[0:04] Che Guan: I have a pressure because I know that I'm speaking between you and lunch. So I try to finish my presentation on time so that you don't have to stay in the long line. [2:20] Passage: <text> Answer: <class> [2:20] Passage <text> Answer <class> [2:20] Passage: <text> Answer: <class>

[0:04] Hello everyone. Um, I have a pressure because I  know that I'm speaking between you and lunch. So [0:09] I try to u finish my presentation on time so that  you don't have to stay in the long line. Today [0:16] I'm going to present our previous research on  optimizing large language models for sustainable [0:24] investors. I don't have many complicated model but  I believe this the insights from this presentation [0:32] can be beneficial for your investment with  that um so this is a joint work between me [0:38] my colleague Andrew chain and three people who are  student from data science institute of vender bill [0:52] now let's start um so let's start with a  question. If you just read this article, [1:00] you don't have to read the contents. Can I ask  you what is your assessment of the sentiment in [1:06] the news article published um in July 13, 2023?  What's your what's your view? Positive, negative, [1:16] neutral? Anybody? Just look at the title. How  many people say positive or negative? Nobody. [1:28] Negative. All right. Um I also highlight two  companies in the news article. Um it is uh um [1:37] Liberty Neutral and Alate. Now my next question  is do you think the news article going to impact [1:45] these two companies? If so, will the impact  be positive or negative? Positive. Positive. [1:55] Any people say negative? Neutral. All right.  Now, let's look at the answer from Gemini. So, [2:08] is negative for the first one and negative for  the second one. All right. This is from Gemini. [2:20] Next, I going to do a little trick.  I'm going to I'm going to add a few [2:26] words in the end of the article. If you see  this article is published, let's say it's [2:32] is from company XYZ. Do you think the news  article going to impact company XYZ? If so, [2:41] will the impact be positive or negative from ESG  environmental, social and governance perspective? [2:53] Any views? Positive, negative, neutral. All  right. All right. Thank you. So now you can [3:03] see that it's little bit the question is a little  bit tricky. So if you read the if you directly use [3:12] large language model like Gemini, Gemini going to  say it's positive but I agree with you is neutral [3:20] because I only add a company XYZ. I can have  my name in it but doesn't mean that it going to [3:26] impact me or the company from ESG perspective. You  see there are bunch of questions are connected. [3:34] First, we care if the news article include  the topic I'm interested. The second is if [3:42] the news article have that content, does it have  a positive or negative impact? So, it's nested. [3:52] Now, you can see I can specify those nuances  with my prompt so that I can have a accurate [3:59] sentiment. But let's look at the prompting.  This is an article published a few years ago [4:05] and in this article the authors use a old open  source model llama 2. All right. So if you look [4:14] at the if you look at the figure so the idea so  what happens here every time we give an article [4:23] or passage to llama 2 and the goal is to let  the llama 2 specify if the article or if the [4:32] contents can be classified into one of the four  categories gender um gender professions race [4:43] and religions. Now, if you look at the box on  the upper left, we can have a format passage, [4:54] next line, answer, right? So, passage column and  I I just copy paste my contents and then answer [5:06] next line answer column answer. If I do this, if  I just run lama 2 to do classification, you can [5:15] see the accuracy is around 0.0036. It's pretty  small. However, if I just slightly replace the [5:25] colon with a space, you see the performance, the  accuracy jump from 0.03 all the way to 0.5. What [5:37] about if I just replace? What about if I just add  a space after the colon performance jump from 0.5 [5:45] to 75? You see that just slightly changing some  words or format for my prompting I got different [5:56] results and this is probably due to how Facebook  fine-tune is open source fine tune it model [6:05] uh using their own data. So prompting may not give  you prompting works most of the time but doesn't [6:13] give you a robust result every time. In this  research we try to tackle all these challenges. [6:23] There are four contributions. First we created  the expert annotated data set labels ESG relevance [6:32] company impact and sentiment. Second, we fine-tune  a multiple large language models to generate ESG [6:41] sentiment and we back test these ESG sentiments  with our with financial data and results shows [6:51] that there is a strong negative correlation  between ESG news and under performance. Now [6:59] I'm going to share the details in next 15 minutes.  We start with the data. So in this research we use [7:08] data from Paragon and we have collected 120,000  news articles from se September 2022 to September [7:20] 2023. It's a one-year data. It covers 508  companies from S&P 500. And because each [7:31] articles can be multi can be relevant to multiple  companies, we train the um we make sure there [7:40] are only up to five comp articles relevant to  individual company. So in total we have two thou [7:47] 214,000 samples for this test. Next we do a simple  analysis for this data. In this simple chart, [8:00] we rank articles by individual company. Then we  rank the company by the unique number of articles [8:08] within the year. And if you look at the chart,  you can see that if you look at the left bar, [8:14] there are around 200 companies um having um  fewer than 50 articles during the period. [8:23] If you look at the bar on the right side, there  are around 100 companies um having around 250 [8:32] um articles during the period. Roughly speaking,  one article per day. So it's a good coverage. [8:42] Next, before we start to label or annotate our  ESG data, we have our experts to carefully read [8:54] um SASBY uh the um the d um the dimension of ESG  from SASBY framework. It covers from environment, [9:04] social capital, human capital, b business model  and innovation and the leadership and governance. [9:12] So after reading these individual  categories carefully, now we have [9:17] the knowledge to label our news articles. [9:23] To start, we randomly sample around 5,250  news articles um crossing 79 industries from [9:34] our research data set. we then to be uh able to  effectively and efficiently annotate individual [9:42] article we conduct the annotation in a sequential  way in three steps uh and and the definition are [9:53] um at least in table one two and three. So  let's look at the first step. If you look at [9:58] table one on the right side here, the goal is  to find out if the article is relevant to ESG [10:07] topic that we are interested. So the definition  for major is if the article majorly focus on [10:16] um one of the five SASB ESG topic is labeled as  major. If the article is um only slightly focused [10:28] on SASBY ESG topic, we label as minor. Otherwise,  we label as no. This is step one. Step two, [10:39] we first check if the article is relevant to  ESG. If no, we mark it as irrelevant. Otherwise, [10:47] we just look at the ESG label articles and we see  if the article going to have a significant impact [10:55] um to the company or is most relevant to the  company. We have major, minor and no. This [11:03] is step two. Step two, table two depends on  table one. Now let's look at the step three. [11:11] We check the first two tables. if either the ESG  relevance um if the news article is not relevant [11:23] to ESG or if the news article is not relevant to  the company we label as irrelevant otherwise we [11:32] measure the tone of the article could be positive  negative and neutral so table three depends on [11:40] table one and table two now we have three tables  three labels in the app from our annotators. [11:54] Now based on the three labels for  individual news articles, so we have 215,000 [12:02] news articles. Sorry. So in total, but we  sample randomly sample around 5,000 articles [12:07] for annotation. For this 5,000 annotated  articles, we can create two signals. The [12:15] first one is called all ESG news with firming  app. So let's look at the definition for this. [12:24] Let's look at the first one positive. If ESG  news, if ESG relevance is major or minor, [12:31] firm relevance is major or minor. If the sentiment  is positive, we label the article as positive. [12:40] Similarly, if ESG relevance is major and minor,  firm relevance major or minor, sentence neutral, [12:47] we label as neutral. And we do the same thing  for inactive. If none of the three are satisfied, [12:54] we label as irrelevant. This is our first signal.  For the second signal, we make it more strict. [13:03] If you just look at the posit positive  definition instead of major or minor, [13:09] if you look at the first um if you  look at the definition this time, [13:13] if ESG relevance is major, firm relevance  is major and sentiment is positive, [13:20] we define as positive. So this is the second  signal is more strict than the first one. So [13:28] now we have derived two sentiment signals. We're  going to back test the the signals shortly. [13:40] Let's look at the distribution of these annotated  samples. For the for the uh war ES news with firm [13:49] impact, it's around 69% for neutral, 50%  for negative and 8% for positive. And you [13:57] can see that the size of negative is double  size of positive. For the second signal core [14:05] ESG news with firm impact, it's around 84% for  irrelevant, 9% for negative, and 4% for positive. [14:15] So the number of negative is still double size  of a positive. The potential reason is positive [14:24] news tend to be generic whereas negative news is  often more specific and easier to be identified. [14:37] Now we're going to use these samples to fine-tune  a model. In this experiments we chose four open [14:47] source models and three closed source models.  For the four open source models, we use Mistro [14:53] 7B instruction model, MHOR 7B general model,  Lama 3 8B um instruction model and the llama [15:02] 3B general models. We also have three closed  source model GPT 3.5 turbo, GPD4.0 mini and GPD4. [15:14] And for the open source model, we can either use  model directly do the inference to classify or we [15:21] can fine-tune the model with our 5,000 annotation  samples to do the inference. If it's done without [15:29] fine-tuning, we say pre-train. We label the  results as pre-trained. Otherwise, we label [15:34] the results as fine-tune. To compare the model  performance, we use F1 score covers negative, [15:44] neutral, positive, irrelevant. And if you look  look at if you look at the last show, we also [15:50] have the overall accuracy that can be easier  to compare different models. Two observations. [15:57] First observation if you look at the llama 3B  instruction model in general the fine-tuned [16:06] model is better than pretraint model. The accuracy  we we can achieve using lama 3b instruction model [16:15] is point is 77%. Second observation, after  fine-tuning a smaller open source model, [16:23] we're able to beat or outperform  the closed source model like GPD4. [16:32] This is for the first signal. For the second  signal, we have the similar observations. So, [16:38] it's u it's not by chance, but lucky, but um  we see the similar observations. Now I'm going [16:46] to back test these two signals with our news  articles. So in this in this experiment we just [16:54] use fine-tune llama 3B uh llama llama 3 8B model  instruction model. We fine tune the model. We use [17:04] a model to score 214,000 news articles. And then  let's say if we if I observe there's a positive [17:17] um sentiment for from that article we trade  the company next day and then close my [17:26] position the day up to next day. So we can  easily calculate the excess return against [17:32] ISM 500. So for each article I can kind  of get the excess return one day ahead. [17:41] And then I take the average of individual cohort.  Here I have five cohorts. If the article is [17:49] labeled as ESG relevant, I put the push article  into that cohort. If the article is labeled as [17:57] positive, I push it to the positive cohort. So  on so forth. Specifically we create a extreme [18:06] um negative cohort which is a subset of negative  cohort. So what we have five cohorts I just need [18:14] to calculate the excess return one day have. So  if you look at if you look at the first signal on [18:21] the left side, if you just look at the negative  and extreme negative cohorts, you can see that [18:28] roughly speaking for the negative cohort, every  day the excess return is negative five basis [18:36] point. If you look at the extreme negative cohort,  every day the excess return is eight basis point. [18:46] If you look at the t uh if you look at the p value  from the t statistics the extreme negative core is [18:55] significant important. If we look at the second  sentiment core ESG news with firm impact we see [19:05] the similar pattern especially for negative cohort  and extreme negative cohort we see significant [19:13] negative return for these two cohort. Okay. Now  this is test is done based on the article level [19:22] but each company can be associated with multiple  articles each day. We cannot trade on individual [19:29] article. Next step for each company we aggregate  individual news relevant news into one signal. We [19:38] do we use majority vote. So every day there's  only one sentence score for for individual [19:47] company. All right. So one signal per company  per day. Now given this consolid signal we do [19:58] back test for the five signals. Again if the  company is labeled as positive today I trade [20:08] the company I buy the company tomorrow and close  my position the day after tomorrow. This is day [20:15] ahead trading. We have I have five goh horse. I  start with $1 in the beginning. After one year, [20:26] if you see if you see the curve uh um in  light blue, this is a performance from [20:36] extreme negative cohort. You can see that if  I start with $1 in the beginning, I lose 25%. [20:48] This is excess return. This mean that  if I I'm able to short that company, [20:55] I can earn additional 25% return. Right? [21:06] For the second signal, we see similar pattern.  If you look at the negative cohort in red in [21:19] red or if you look at the extreme negative cohort  in blue light in blue light blue again I start [21:28] with $1 after a year I can if I can if I'm able  to short those cohort I can earn more than 25% [21:36] revenues. In this page, I'm trying to give you  um kind of like a overall performance of the [21:41] two signals. If you look at the first signal,  the sharp the information ratio for negative [21:48] cohort is 65%. For the extreme negative is 1.04.  We know that if any signal with information ratio [21:57] about 0.5, it means it's a good one. We see about  0.5 for negative and extreme negative from the [22:07] first signal. We see the same thing for the second  signal. This mean that these two sentiment scores [22:13] could be profitable. To sum up, we show that um  large language model can understand the contextual [22:25] sentiment and nuances languages with greater u  depth. fine-tuning gonna improve the performance [22:32] and most significantly we find out negative ESG  news tend to predict stock and the performance. [22:41] I just want to take you one minute because  this research was done two years ago. [22:46] Now we know that human annotations are very  expensive and if I change my time horizon [22:55] let's say I'm not looking into one day hat I'm  looking into one month hat or a quarter hat can [23:01] I still use a sentiment maybe the sentiment  doesn't going to lose it predictive power [23:08] how can we still use this sentiment score from  humans let me give an example in 2026 March 10th [23:17] XM mobile announced that it's going to move its  legal home to taxes on the same day next day [23:25] the stock performance jump by 2%. And then in one  month performance jump by 7%. But if you look at [23:34] the three months ahead it come back to neutral. So  if I have human label this company with positive [23:43] today I can use it for next day trading but  for three months ahead I cannot use this [23:49] sentiment labels right. How can we still leverage  these sentiment labels from humans? We can use [23:58] reinforce learning to align these human annotated  samples to the time horizon we chose. We choose [24:08] idea is we can use three months ahead return  as reward to slightly change the distribution [24:16] from language model. If it predicts positive  using the reward from reinforcement learning, [24:25] we can penalize this prediction and adjust it  to neutral if we target for three months ahead. [24:33] We'll share with you the results next year. With  that, thank you very much for your attention.