Scholarlink

<p>Financial time series prediction requires researchers to detect dynamic market indicator relationships while also predicting short-term market trends that are either positive or negative. Conventional LSTM (cLSTM) models face limitations when identifying multiple complex relationships between their input data points. Our work adds Price-to-Earnings (PE) and Price-to-Book (PB), Volatility Index (VIX), and Sentiment Score to the standard Open, Close, Volume features to enhance prediction accuracy levels. We introduce two enhanced Extended Long Short-Term Memory (xLSTM) architectures: (i) xLSTMcg, which implements cross-gating, and (ii) xLSTMeg, which employs exponential gating to improve information flow between feature channels. We compare categorical versus feature-based loss functions and apply ensemble techniques—bagging, boosting, and stacking—to optimize our bidirectional fused models. Furthermore, our study also evaluates the stability of models on validation dataset. The bidirectional configurations produced superior results compared to unidirectional configuration by 2% to 3%. Through the experiment, xLSTMeg model achieved 86.2% accuracy, but when combined within a stacking ensemble, yielded an accuracy of 91.2% with area-under-the-curve (AUC) of 0.95. The obtained results of xLSTMeg showed better performance than both cLSTM and xLSTMcg models. Additionally, our study meets regulatory standards by maintaining less than a 5% performance difference between experimental and external validation dataset. Integrating advanced gating mechanisms with ensemble fusion significantly enhances LSTM-based financial forecasts, delivering robust performance under varied market conditions. This framework lays a solid foundation for future market-forecasting research and supports more effective trading decisions and risk-management strategies.</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

StockAI 3.0: Ensemble fusion paradigms using novel gating mechanism in long short-term memory architectures for forecasting sentiment-based stock trends

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Yuvraj Sharma,
  • Siddharth Gupta,
  • Neha Gupta,
  • Ekta Tiwari,
  • Rajesh Singh,
  • Narendra N. Khanna,
  • Mustafa Al-Maini,
  • Vijay Rathore,
  • Puneet Ahluwalia,
  • Vandana Kumari,
  • Subbaram Naidu,
  • Luca Saba,
  • Jasjit S. Suri

摘要

Financial time series prediction requires researchers to detect dynamic market indicator relationships while also predicting short-term market trends that are either positive or negative. Conventional LSTM (cLSTM) models face limitations when identifying multiple complex relationships between their input data points. Our work adds Price-to-Earnings (PE) and Price-to-Book (PB), Volatility Index (VIX), and Sentiment Score to the standard Open, Close, Volume features to enhance prediction accuracy levels. We introduce two enhanced Extended Long Short-Term Memory (xLSTM) architectures: (i) xLSTMcg, which implements cross-gating, and (ii) xLSTMeg, which employs exponential gating to improve information flow between feature channels. We compare categorical versus feature-based loss functions and apply ensemble techniques—bagging, boosting, and stacking—to optimize our bidirectional fused models. Furthermore, our study also evaluates the stability of models on validation dataset. The bidirectional configurations produced superior results compared to unidirectional configuration by 2% to 3%. Through the experiment, xLSTMeg model achieved 86.2% accuracy, but when combined within a stacking ensemble, yielded an accuracy of 91.2% with area-under-the-curve (AUC) of 0.95. The obtained results of xLSTMeg showed better performance than both cLSTM and xLSTMcg models. Additionally, our study meets regulatory standards by maintaining less than a 5% performance difference between experimental and external validation dataset. Integrating advanced gating mechanisms with ensemble fusion significantly enhances LSTM-based financial forecasts, delivering robust performance under varied market conditions. This framework lays a solid foundation for future market-forecasting research and supports more effective trading decisions and risk-management strategies.