Scholarlink

<p>Recently, the use of deep learning network concepts in the field of financial time series has attracted the attention of many researchers due to the noise and non-stationary characteristics of these data. This study presents a new deep learning framework in which the appropriate features of Tehran company’s stock data are extracted using a stack auto-encoder and based on the information of the past 20 days of each stock, its future 10-day status is predicted. In first step, the generated dataset sent to a variable autoencoder to remove noise. In second step, a convolutional autoencoder is used to efficiently generate high-level features. In third step, high-level noise free features are fed into the hybrid model based on Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) to forecast the next ten day price labeling. The results of this research have been evaluated in four different case studies. When the Hybrid 4-Classifier Stacked Autoencoder (H4CSAE) model was used to reduce noise and extract features, the precision is equal to 77.78%, while in other methods such as XGboost, LSTM, GRU and BiLSTM are respectively equal to 52.38%, 47.37%, 63.33% and 65.63%. Based on the results of the research, one of the strengths of the proposed method is reducing the rate of false positives in the final prediction.</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 Model for Stock Market Prediction with Stacked Autoencoder for Feature Engineering

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Maryam Sadat Hashemipour,
  • Morteza Zahedi,
  • Mansoor Fateh

摘要

Recently, the use of deep learning network concepts in the field of financial time series has attracted the attention of many researchers due to the noise and non-stationary characteristics of these data. This study presents a new deep learning framework in which the appropriate features of Tehran company’s stock data are extracted using a stack auto-encoder and based on the information of the past 20 days of each stock, its future 10-day status is predicted. In first step, the generated dataset sent to a variable autoencoder to remove noise. In second step, a convolutional autoencoder is used to efficiently generate high-level features. In third step, high-level noise free features are fed into the hybrid model based on Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) to forecast the next ten day price labeling. The results of this research have been evaluated in four different case studies. When the Hybrid 4-Classifier Stacked Autoencoder (H4CSAE) model was used to reduce noise and extract features, the precision is equal to 77.78%, while in other methods such as XGboost, LSTM, GRU and BiLSTM are respectively equal to 52.38%, 47.37%, 63.33% and 65.63%. Based on the results of the research, one of the strengths of the proposed method is reducing the rate of false positives in the final prediction.