Stock Movement Prediction with Financial News using Contextualized Embedding from BERT
Stock Movement Prediction with Financial News using Contextualized Embedding from BERT
This paper presents FT-CE-RNN, a novel model for predicting short-term stock price movements following financial news events using only news headlines. The model leverages contextualized embeddings from a fine-tuned BERT model, which captures domain-specific financial knowledge, and feeds them into a recurrent neural network (RNN) for binary classification (positive or negative movement). The authors address limitations of prior work that used static embeddings (e.g., Word2Vec, GloVe) by employing BERT's contextualized representations, and they fine-tune BERT on financial news to incorporate domain expertise. The model is evaluated on a large dataset of Bloomberg News headlines for STOXX Europe 600 stocks, using market-adjusted returns with different time windows for news during and outside trading hours. Results show that FT-CE-RNN achieves state-of-the-art accuracy and Matthews Correlation Coefficient compared to baselines including Naive Bayes, SSESTM, Bloomberg's proprietary sentiment, BERT, and FinBERT. Additionally, trading simulations with two strategies (selecting top/bottom scored stocks and a weighted portfolio) demonstrate superior annualized returns and Sharpe ratios, confirming practical profitability. The paper also introduces a new evaluation metric that focuses on extreme prediction scores, aligning with investor interest in market-moving news rather than neutral ones. Ablation studies confirm the importance of fine-tuning and the choice of embedding layer, with the last layer of fine-tuned BERT and LSTM cells yielding the best performance.
Highlights
- 1Introduces FT-CE-RNN, a novel model combining fine-tuned BERT contextualized embeddings with RNN for stock movement prediction from news headlines.
- 2Demonstrates state-of-the-art performance on Bloomberg News data, outperforming static embedding and other baseline models in accuracy and trading simulations.
- 3Proposes a new evaluation metric focusing on extreme prediction scores to better reflect investor interests in market-moving news.
- 4Shows significant improvement in trading simulations using millions of headlines, validating real-world applicability.
Methods
- MFine-tuned BERT model to generate contextualized embeddings with financial domain knowledge.
- MRecurrent Neural Network (RNN) with LSTM cells to process sequential embeddings for binary classification.
- MData labeling using market-adjusted returns with different time windows for trading and non-trading hours, selecting top/bottom 15% for training.
- MTrading simulations with two strategies (top/bottom stocks and weighted portfolio) to evaluate practical performance.
Results
- RFT-CE-RNN achieves higher accuracy and Matthews Correlation Coefficient (MCC) compared to baselines (NBC, SSESTM, Bloomberg, BERT, FinBERT) on test sets.
- RTrading simulations show superior annualized return and Sharpe ratio, demonstrating profitability in real scenarios.
- RContextualized embeddings from fine-tuned BERT outperform static embeddings and base BERT embeddings in prediction tasks.
- RThe model performs best when using embeddings from the last layer of fine-tuned BERT and LSTM cells in the RNN.
Analyze Paper
Generate insights from "Stock Movement Prediction with Financial News using Contextu...".