Scholarlink

<p>Accurate long-term forecasting of stock prices, especially using high-frequency data, remains a significant challenge due to the inherent non-stationarity, noise, and complex long-term dependencies present in such series. To address this, we propose StockCI, a novel hybrid model that strategically integrates Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and the Informer architecture. The model is designed with a two-stage workflow: first, the CEEMDAN module adaptively decomposes the original complex price series into a set of simpler, more stationary sub-series (IMFs), effectively filtering out noise and mitigating the problem of mode mixing. Subsequently, the Informer module, renowned for its efficiency in long-sequence forecasting, predicts future values of each decomposed component by leveraging its ProbSparse self-attention mechanism. The final forecast is obtained by reconstructing the predicted components. We evaluated StockCI on minute-level high-frequency data from China’s A-share market, including major indices and individual stocks. Empirical results demonstrate that StockCI significantly outperforms a comprehensive suite of benchmarks, including traditional models (ARIMA), deep learning models (RNN, LSTM, GRU), and state-of-the-art transformers (Autoformer, FEDformer, PatchTST). For instance, StockCI achieved an average reduction in MAE of over 15% compared to the standalone Informer model. Empirical findings confirm StockCI’s superior performance in accurately forecasting long-term stock prices, demonstrating significant improvements over established academic benchmarks including ARIMA, RNN, LSTM, and Informer.</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

StockCI: a hybrid model integrating CEEMDAN and informer for enhanced long-term stock price forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月18日

查看全文
  • Mo-Ce Gao

摘要

Accurate long-term forecasting of stock prices, especially using high-frequency data, remains a significant challenge due to the inherent non-stationarity, noise, and complex long-term dependencies present in such series. To address this, we propose StockCI, a novel hybrid model that strategically integrates Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and the Informer architecture. The model is designed with a two-stage workflow: first, the CEEMDAN module adaptively decomposes the original complex price series into a set of simpler, more stationary sub-series (IMFs), effectively filtering out noise and mitigating the problem of mode mixing. Subsequently, the Informer module, renowned for its efficiency in long-sequence forecasting, predicts future values of each decomposed component by leveraging its ProbSparse self-attention mechanism. The final forecast is obtained by reconstructing the predicted components. We evaluated StockCI on minute-level high-frequency data from China’s A-share market, including major indices and individual stocks. Empirical results demonstrate that StockCI significantly outperforms a comprehensive suite of benchmarks, including traditional models (ARIMA), deep learning models (RNN, LSTM, GRU), and state-of-the-art transformers (Autoformer, FEDformer, PatchTST). For instance, StockCI achieved an average reduction in MAE of over 15% compared to the standalone Informer model. Empirical findings confirm StockCI’s superior performance in accurately forecasting long-term stock prices, demonstrating significant improvements over established academic benchmarks including ARIMA, RNN, LSTM, and Informer.