Financial sentiment analysis using FinBERT with application in predicting stock movement
Financial sentiment analysis using FinBERT with application in predicting stock movement
This paper presents a hybrid model for stock movement prediction that combines FinBERT, a pre-trained language model specialized for financial text, with LSTM networks. The authors hypothesize that news sentiment contains predictive information for market movements, grounded in the Efficient Market Hypothesis. They process a large dataset of financial news titles using FinBERT to extract daily sentiment scores, which are then concatenated with historical market features (open, high, low, close, volume) and fed into a two-layer LSTM with 100 hidden units each. The model is trained to minimize MSE on next-day closing price prediction, using a temporal 90/10 train-test split and early stopping.
Experiments on a combined dataset of over 1 million news articles show that the proposed FinBERT+LSTM model achieves the lowest validation loss (3.1975e-04), outperforming ARIMA (0.00975), LSTM without sentiment (1.1937e-03), BERT+LSTM (3.3132e-04), and FinBERT+DNN (3.4177e-04). The results confirm that sentiment analysis enhances predictive accuracy, though the improvement over base BERT is modest, suggesting room for further FinBERT fine-tuning. The paper also introduces a Numerical Sentiment Index (NSI) based on price movements for potential use in fine-tuning.
Limitations include small dataset sizes for individual stocks (e.g., MSFT reduced to 1440 rows), reliance on news titles only, and a tendency to underpredict long-term trends. Future work is proposed to address these issues, such as hybrid ARIMA-LSTM modeling, incorporating full news articles or social media data, fine-tuning FinBERT on task-specific data, and exploring advanced time-series architectures like TEANet. Overall, the study demonstrates the value of integrating domain-specific NLP with deep learning for financial forecasting.
Highlights
- 1Integrates FinBERT sentiment analysis with LSTM networks for stock movement prediction.
- 2Demonstrates that incorporating sentiment analysis significantly improves prediction accuracy over traditional models.
- 3Proposes a Numerical Sentiment Index (NSI) for market-based sentiment quantification.
- 4Outperforms ARIMA, LSTM without sentiment, and BERT+LSTM baselines on MSE.
- 5Identifies limitations and future directions including hybrid ARIMA-LSTM and fine-tuning FinBERT.
Methods
- MFinBERT for financial sentiment extraction from news titles, producing daily aggregated sentiment scores.
- MLSTM-based time-series forecasting module with two LSTM layers (100 units each) and dense layers.
- MData preprocessing: merging news datasets, Min-Max scaling, temporal train-test split (90/10), handling non-trading days.
- MBaseline models: ARIMA(1,1,2), LSTM on numerical data only, FinBERT+DNN, and BERT+LSTM.
Results
- RFinBERT+LSTM achieved the lowest validation loss (3.1975e-04) among all models.
- RARIMA had higher MSE (0.00975) but captured long-term trends well.
- RLSTM without sentiment had validation loss 1.1937e-03, outperformed by sentiment-enhanced models.
- RFinBERT+LSTM slightly outperformed BERT+LSTM (validation loss 3.3132e-04), indicating marginal benefit from domain-specific pre-training.
- RModel tends to underpredict long-term trends, suggesting potential for hybrid ARIMA-LSTM approach.
Analyze Paper
Generate insights from "Financial sentiment analysis using FinBERT with application ...".