Scholarlink

<p>Stock price prediction is a core challenge in the financial domain, with its complexity stemming from market non-stationarity, high noise characteristics, and dynamic interactions across multiple time scales. Existing deep learning methods face issues in financial time series forecasting, including slow response of traditional decomposition methods and the difficulty of general architectures to simultaneously capture long-term trends and short-term seasonal patterns. This study proposes the EMA-StockPredictor model, an innovative architecture combining Exponential Moving Average (EMA) decomposition with dual-stream neural networks. This model introduces a novel rolling-window EMA decomposition strategy into the deep learning prediction framework, which strictly avoids look-ahead bias while enabling faster response to market changes, enabling faster response to market changes through exponential weighting characteristics. Meanwhile, a specialized dual-stream network is designed, employing a linear stream (MLP) to process smooth trend components and a nonlinear stream (CNN combined with Patching technique) to extract complex seasonal patterns. Experiments on five US technology stocks (AAPL, GOOG, NVDA, ORCL, TSLA) covering approximately 1500 trading days from 2020-2025 demonstrate that, compared to baseline models including ARIMA, LSTM, DLinear, PatchTST, iTransformer, and TimeMixer, EMA-StockPredictor achieves 15.3% reduction in average MAE, 22.7% reduction in MSE, and 62.8% directional prediction accuracy. Trading strategies based on model predictions achieve a Sharpe ratio of 1.67 and 22.5% annualized return, with maximum drawdown of only −7.8%. Ablation studies quantify the contribution of each component: RevIN normalization 19.2%, dual-stream architecture 12.3%, EMA decomposition 8.6%, and Patching 6.7%. This study provides a new methodological paradigm for financial time series forecasting, demonstrating the effectiveness of organically combining classical statistical techniques with modern deep learning architectures.</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 multi-stock market trend prediction model based on exponential decomposition and dual-stream networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Tong Zhang,
  • Zhengwei Wang,
  • Tong Zheng,
  • Yan Guo

摘要

Stock price prediction is a core challenge in the financial domain, with its complexity stemming from market non-stationarity, high noise characteristics, and dynamic interactions across multiple time scales. Existing deep learning methods face issues in financial time series forecasting, including slow response of traditional decomposition methods and the difficulty of general architectures to simultaneously capture long-term trends and short-term seasonal patterns. This study proposes the EMA-StockPredictor model, an innovative architecture combining Exponential Moving Average (EMA) decomposition with dual-stream neural networks. This model introduces a novel rolling-window EMA decomposition strategy into the deep learning prediction framework, which strictly avoids look-ahead bias while enabling faster response to market changes, enabling faster response to market changes through exponential weighting characteristics. Meanwhile, a specialized dual-stream network is designed, employing a linear stream (MLP) to process smooth trend components and a nonlinear stream (CNN combined with Patching technique) to extract complex seasonal patterns. Experiments on five US technology stocks (AAPL, GOOG, NVDA, ORCL, TSLA) covering approximately 1500 trading days from 2020-2025 demonstrate that, compared to baseline models including ARIMA, LSTM, DLinear, PatchTST, iTransformer, and TimeMixer, EMA-StockPredictor achieves 15.3% reduction in average MAE, 22.7% reduction in MSE, and 62.8% directional prediction accuracy. Trading strategies based on model predictions achieve a Sharpe ratio of 1.67 and 22.5% annualized return, with maximum drawdown of only −7.8%. Ablation studies quantify the contribution of each component: RevIN normalization 19.2%, dual-stream architecture 12.3%, EMA decomposition 8.6%, and Patching 6.7%. This study provides a new methodological paradigm for financial time series forecasting, demonstrating the effectiveness of organically combining classical statistical techniques with modern deep learning architectures.