[논문 리뷰] BERT-based Financial Sentiment Index and LSTM-based Stock Return Predictability

Joshua Zoen Git Hiew, Xin Huang외 4명|arXiv (Cornell University)|2019. 06. 21.

Stock Market Forecasting Methods참고 문헌 18인용 수 40

한 줄 요약

본 논문은 Weibo에서 수집한 중국어 텍스트를 바탕으로 세 개의 HKSE 주식에 대한 BERT 기반 텍스트형 감정 지수를 구축하고, 이를 옵션- 및 시장 암시 감정과 비교하며, 다중 감정 채널을 사용할 때 LSTM이 VAR보다 주가 수익률을 더 잘 예측할 수 있음을 보여주되, 더 긴 기간에는 ML 모델이 과적합될 수 있다.

ABSTRACT

Traditional sentiment construction in finance relies heavily on the dictionary-based approach, with a few exceptions using simple machine learning techniques such as Naive Bayes classifier. While the current literature has not yet invoked the rapid advancement in the natural language processing, we construct in this research a textual-based sentiment index using a well-known pre-trained model BERT developed by Google, especially for three actively trading individual stocks in Hong Kong market with at the same time the hot discussion on Weibo.com. On the one hand, we demonstrate a significant enhancement of applying BERT in financial sentiment analysis when compared with the existing models. On the other hand, by combining with the other two commonly-used methods when it comes to building the sentiment index in the financial literature, i.e., the option-implied and the market-implied approaches, we propose a more general and comprehensive framework for the financial sentiment analysis, and further provide convincing outcomes for the predictability of individual stock return by combining LSTM (with a feature of a nonlinear mapping). It is significantly distinct with the dominating econometric methods in sentiment influence analysis which are all of a nature of linear regression.

연구 동기 및 목표

• Construct a textual financial sentiment index for three HKSE stocks using BERT fine-tuned on Chinese text from Weibo.
• Compare BERT with other NLP models for sentiment classification in finance.
• Develop a three-channel sentiment framework (BERT-based, option-implied, market-implied) to study their joint impact.
• Assess stock return predictability using sentiment indices with LSTM and compare to VAR.
• Evaluate year-by-year vs. full-period predictive performance and discuss potential overfitting.

제안 방법

• Pre-process and manually label Weibo posts for three HKSE stocks with a three-class polarity (positive, neutral, negative) using a voting consensus from experts.
• Fine-tune Chinese BERT on labeled data and compare against Transformer with attention, PMI-enhanced BiLSTM, PMI-enhanced Multichannel CNN, and FastText; report micro-averaged precision, recall, and F1.
• Compute BSIt,i = (Post,i − Negt,i) / (Post,i + Neut,i + Negt,i) for each stock i and trading day t to form the BERT-based sentiment index.
• Construct option-implied sentiment OSIi via risk-neutral skewness of option data (per Han, 2008).
• Construct market-implied sentiment MSIi via PCA-based or market-data-based approach following Chong et al. (2014) for daily data.
• Model stock return predictability using VAR (linear) and LSTM (nonlinear) with inputs including sentiment indices and eight risk factors (r1, r3−r1, Inf, UMD, HKD, rm−r1, SMB, HML).

실험 결과

연구 질문

• RQ1 Can a BERT-based sentiment index built from Chinese social media improve sentiment measurement in finance compared to existing NLP models?
• RQ2 How do three sentiment channels (textual, option-implied, market-implied) interact in explaining stock return movements?
• RQ3 Is LSTM able to outperform VAR in predicting individual stock returns when incorporating sentiment indices, and under what time horizons?
• RQ4 Does combining multiple sentiment channels (Mixture) yield better predictive accuracy than using a single channel?

주요 결과

• BERT outperforms other models in sentiment classification on Chinese financial text (higher precision, recall, and F1, with notably higher recall).
• The BERT-based sentiment index (BSI) plus two additional channels (OSI, MSI) form a three-channel framework for financial sentiment analysis.
• Across yearly predictions, LSTM with sentiment indices generally achieves lower MSE than VAR, indicating nonlinear relationships in sentiment-impact on returns.
• The strongest predictive performance frequently occurs when using a Mixture of all three sentiment channels, rather than any single channel, under the LSTM model.
• Over the full period 2016–2018, LSTM does not consistently outperform VAR, suggesting possible overfitting for longer horizons.
• Adding sentiment indices significantly improves prediction when using LSTM compared to the No-SI baseline.