Scholarlink

<p>Financial markets are inherently complex and volatile, making accurate predictions essential for early warnings of potential crashes and subsequent recoveries. To address the intricate challenge of accurate time series forecasting in financial markets, various approaches from financial mathematics, machine learning, and advanced deep learning have been explored, each aiming to improve forecasting reliability. In this research, we propose a novel deep learning architecture based on the Encoder-Decoder framework for financial time series forecasting. This architecture consists of the following key components: the Encoder employs a Bidirectional Gated Recurrent Unit (BiGRU) with three layers to capture temporal dependencies and extract deep features from the input data. An Attention Mechanism (AM) is integrated to assign different weights to the encoder’s outputs, highlighting the most relevant features for each decoding step. The Decoder also utilizes BiGRU with three layers, which combines the context vector from the AM with the input at each time step, allowing it to generate informed and contextually relevant outputs. Finally, a fully connected layer is added to learn the relevant features from the decoder’s output and generate the final prediction results. The experiment includes the use of various activation functions and the optimization of hyperparameters. Empirical analysis conducted on three time series from each of the three distinct financial markets: stocks, cryptocurrencies, and commodities, demonstrates that the proposed model consistently outperforms benchmark models in terms of predictive accuracy. The Diebold-Mariano (DM) test further provides statistical evidence supporting the superiority of the proposed model. Furthermore, the model exhibited robust forecasting performance during the extreme market volatility induced by the COVID-19 pandemic, maintaining superior accuracy relative to alternative approaches. These comparative results underscore the proposed model’s exceptional predictive capabilities and adaptability, positioning it as a highly effective tool for financial time series forecasting applications.</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

Encoder-Decoder BiGRU Framework with Attention for Time Series Forecasting in Financial Markets

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Muhammad Zubair,
  • Zhensheng Huang

摘要

Financial markets are inherently complex and volatile, making accurate predictions essential for early warnings of potential crashes and subsequent recoveries. To address the intricate challenge of accurate time series forecasting in financial markets, various approaches from financial mathematics, machine learning, and advanced deep learning have been explored, each aiming to improve forecasting reliability. In this research, we propose a novel deep learning architecture based on the Encoder-Decoder framework for financial time series forecasting. This architecture consists of the following key components: the Encoder employs a Bidirectional Gated Recurrent Unit (BiGRU) with three layers to capture temporal dependencies and extract deep features from the input data. An Attention Mechanism (AM) is integrated to assign different weights to the encoder’s outputs, highlighting the most relevant features for each decoding step. The Decoder also utilizes BiGRU with three layers, which combines the context vector from the AM with the input at each time step, allowing it to generate informed and contextually relevant outputs. Finally, a fully connected layer is added to learn the relevant features from the decoder’s output and generate the final prediction results. The experiment includes the use of various activation functions and the optimization of hyperparameters. Empirical analysis conducted on three time series from each of the three distinct financial markets: stocks, cryptocurrencies, and commodities, demonstrates that the proposed model consistently outperforms benchmark models in terms of predictive accuracy. The Diebold-Mariano (DM) test further provides statistical evidence supporting the superiority of the proposed model. Furthermore, the model exhibited robust forecasting performance during the extreme market volatility induced by the COVID-19 pandemic, maintaining superior accuracy relative to alternative approaches. These comparative results underscore the proposed model’s exceptional predictive capabilities and adaptability, positioning it as a highly effective tool for financial time series forecasting applications.