Scholarlink

<p>Stock market prediction remains one of the most intricate challenges in financial data science due to its non-linearity, volatility, and susceptibility to human sentiment. Traditional time-series forecasting techniques—relying solely on historical price indicators—often fall short in capturing the psychological and behavioral dimensions of market dynamics. To address these limitations, this study introduces SABER (Sentiment-Aware BiLSTM Ensemble with RO-BERTa), a novel multimodal architecture that fuses structured financial indicators with deep contextual sentiment embeddings derived from Robustly Optimized Bidirectional Encoder Representations from Transformers (RO-BERTa), temporal modeling via Bidirectional Long Short-Term Memory (BiLSTM), and data efficiency through Deep Batch Active Learning (DBAL). This framework effectively bridges the gap between qualitative sentiment and quantitative financial data, enabling more robust stock predictions. The model was trained and evaluated on a combination of three high-quality datasets: Twitter-based financial sentiment (TSA), large-scale tweet network data (Twitter7), and historical NASDAQ stock prices (2020–2024). Empirical results on major tech stocks (AAPL, MSFT, AMZN, GOOGL) reveal SABER’s superior performance across key metrics—achieving RMSE of 0.92, R² of 0.89, and directional accuracy of 88.2%—outperforming baseline and advanced models alike.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SABER: a multimodal sentiment aware regression framework using ROBERTa BiLSTM and deep batch active learning for stock market prediction

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Chetanpal Singh,
  • Jatinder Warraich,
  • Rahul Thakkar,
  • Nguyen Vo

摘要

Stock market prediction remains one of the most intricate challenges in financial data science due to its non-linearity, volatility, and susceptibility to human sentiment. Traditional time-series forecasting techniques—relying solely on historical price indicators—often fall short in capturing the psychological and behavioral dimensions of market dynamics. To address these limitations, this study introduces SABER (Sentiment-Aware BiLSTM Ensemble with RO-BERTa), a novel multimodal architecture that fuses structured financial indicators with deep contextual sentiment embeddings derived from Robustly Optimized Bidirectional Encoder Representations from Transformers (RO-BERTa), temporal modeling via Bidirectional Long Short-Term Memory (BiLSTM), and data efficiency through Deep Batch Active Learning (DBAL). This framework effectively bridges the gap between qualitative sentiment and quantitative financial data, enabling more robust stock predictions. The model was trained and evaluated on a combination of three high-quality datasets: Twitter-based financial sentiment (TSA), large-scale tweet network data (Twitter7), and historical NASDAQ stock prices (2020–2024). Empirical results on major tech stocks (AAPL, MSFT, AMZN, GOOGL) reveal SABER’s superior performance across key metrics—achieving RMSE of 0.92, R² of 0.89, and directional accuracy of 88.2%—outperforming baseline and advanced models alike.