Scholarlink

<p>This paper presents a hybrid deep learning framework consisting of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Local Interpretable Model-Agnostic Explanations (LIME) to analyze changes in the price of stocks with greater predictive accuracy and interpretive transparency. The framework makes use of the nonlinear, volatile, and time dependent characteristics of stock market data aided by spatial feature extraction using CNN and long-term temporal modeling using LSTM. LIME is added as an explainability layer to understand and assign credit to the logic behind every prediction to ensure trust in model results. Multi-domain financial data, which consists of Yahoo Finance, NASDAQ Composite Index and S&amp;P 500 Constituents, are tested on the system in different market regimes. The proposed framework performs better than all the selected baselines i.e. the average RMSE for the model is 0.0287, which is better than Transformer, GRU (gated recurrent unit), and LSTM baselines. Also, the proposed framework attains the lowest Average Mean Absolute Percentage Error (MAPE) (2.58). These findings confirm that this hybrid model not only solidifies predictive strength when the market is volatile but also allows transparent and interpretable forecasting, which plays a significant role in financial decision support systems.</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

A hybrid deep learning framework to analyze fluctuations in stock prices using CNN, LSTM and LIME

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Aanchal Kaul,
  • Rajinder Kaur

摘要

This paper presents a hybrid deep learning framework consisting of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Local Interpretable Model-Agnostic Explanations (LIME) to analyze changes in the price of stocks with greater predictive accuracy and interpretive transparency. The framework makes use of the nonlinear, volatile, and time dependent characteristics of stock market data aided by spatial feature extraction using CNN and long-term temporal modeling using LSTM. LIME is added as an explainability layer to understand and assign credit to the logic behind every prediction to ensure trust in model results. Multi-domain financial data, which consists of Yahoo Finance, NASDAQ Composite Index and S&P 500 Constituents, are tested on the system in different market regimes. The proposed framework performs better than all the selected baselines i.e. the average RMSE for the model is 0.0287, which is better than Transformer, GRU (gated recurrent unit), and LSTM baselines. Also, the proposed framework attains the lowest Average Mean Absolute Percentage Error (MAPE) (2.58). These findings confirm that this hybrid model not only solidifies predictive strength when the market is volatile but also allows transparent and interpretable forecasting, which plays a significant role in financial decision support systems.