Scholarlink

This study compares the performance of Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and a novel hybrid CNN-LSTM approach for predicting stock prices of Indian public sector undertakings (PSUs): BPCL, NHPC, NMDC, and SBIN, representing diverse sectors such as energy, power, mining, and banking. Using historical stock data from January 1, 2000, to January 31, 2025, along with technical indicators, models were trained and evaluated through time series forecasting techniques. Results indicate that LSTM outperforms other models for BPCL, NHPC, and NMDC, capturing long-term dependencies with the highest R2 values (up to 0.9942) and lowest errors. For SBIN, CNN achieves better accuracy (R2 = 0.9401), excelling in short-term trend detection. Surprisingly, the hybrid CNN-LSTM model does not surpass standalone models, suggesting that individual architectures may be more effective. Future research should explore ensemble techniques to enhance predictive performance.

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

Comparison of LSTM, CNN, and Novel Hybrid Models for Indian PSU Stock Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pallavi Mallikarjun Kanaki,
  • Shaikh Zakir Mujeeb

摘要

This study compares the performance of Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and a novel hybrid CNN-LSTM approach for predicting stock prices of Indian public sector undertakings (PSUs): BPCL, NHPC, NMDC, and SBIN, representing diverse sectors such as energy, power, mining, and banking. Using historical stock data from January 1, 2000, to January 31, 2025, along with technical indicators, models were trained and evaluated through time series forecasting techniques. Results indicate that LSTM outperforms other models for BPCL, NHPC, and NMDC, capturing long-term dependencies with the highest R2 values (up to 0.9942) and lowest errors. For SBIN, CNN achieves better accuracy (R2 = 0.9401), excelling in short-term trend detection. Surprisingly, the hybrid CNN-LSTM model does not surpass standalone models, suggesting that individual architectures may be more effective. Future research should explore ensemble techniques to enhance predictive performance.