Scholarlink

In the domain of financial time series forecasting, deep learning architectures have demonstrated significant efficacy in modeling intricate temporal dependencies. This study presents a comparative analysis of two prominent recurrent neural network frameworks-Long ShortTerm Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) - with a specific focus on their predictive accuracy as quantified by the Mean Absolute Percentage Error (MAPE). Using historical stock price data of Apple Inc. (AAPL) obtained from Yahoo Finance, both models are trained and evaluated under identical experimental conditions. The empirical findings reveal that the Bi-LSTM architecture exhibits superior generalization capability, achieving a MAPE of \(2.4{{\% }}\) on the training set and \(2.1{{\% }}\) on the testing set. In contrast, the traditional LSTM model records higher MAPE values of \(4.79{{\% }}\) and \(3.9{{\% }}\) , respectively. These results underscore the enhanced predictive robustness of Bi-LSTM, affirming its advantage in capturing bidirectional temporal dynamics. Consequently, the Bi-LSTM model emerges as a more reliable computational tool for stock price forecasting, offering improved accuracy in support of data-driven financial decision-making.

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

Evaluating the Predictive Performance of LSTM and Bi-LSTM Models in Financial Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dinh-Cuong Do,
  • Quang-Quy Tran,
  • Ha-Cong Ly Nguyen,
  • Thi-Ngoc Linh Tran

摘要

In the domain of financial time series forecasting, deep learning architectures have demonstrated significant efficacy in modeling intricate temporal dependencies. This study presents a comparative analysis of two prominent recurrent neural network frameworks-Long ShortTerm Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) - with a specific focus on their predictive accuracy as quantified by the Mean Absolute Percentage Error (MAPE). Using historical stock price data of Apple Inc. (AAPL) obtained from Yahoo Finance, both models are trained and evaluated under identical experimental conditions. The empirical findings reveal that the Bi-LSTM architecture exhibits superior generalization capability, achieving a MAPE of \(2.4{{\% }}\) on the training set and \(2.1{{\% }}\) on the testing set. In contrast, the traditional LSTM model records higher MAPE values of \(4.79{{\% }}\) and \(3.9{{\% }}\) , respectively. These results underscore the enhanced predictive robustness of Bi-LSTM, affirming its advantage in capturing bidirectional temporal dynamics. Consequently, the Bi-LSTM model emerges as a more reliable computational tool for stock price forecasting, offering improved accuracy in support of data-driven financial decision-making.