Scholarlink

<p>This study proposes and implements a novel deep learning-based model, the multi-label stock classifier (MLSC), which combines a multi-output classification method with several components: a categorical input embedding, a multi-head attention block, residual connections, convolutional neural networks (CNNs), and gated recurrent units (GRUs). This combination enables the model to simultaneously predict stock trend movements across multiple time horizons (e.g., 1-day, 3-days, 7-days, and 10-days). This multi-label allows traders to simultaneously predict price trends over different time horizons, effectively supporting trading decisions by identifying consistent trend movements over a specified period rather than relying solely on the next-day trends. The architecture of the proposed model effectively addresses three key challenges: imbalanced target labels across different horizons where each timeframe predicts either of three classes: <i>buy</i>, <i>hold</i>, or <i>sell</i> for the corresponding stock, accurate multi-label output learning, and efficient capture of both short and long-term dependencies within the data. The classification performance of the MLSC model has been evaluated across nine different indices using a time-series cross-validation fold (where <i>k</i> = 3, 5, and 7) for various time steps (5, 10, and 15). The best results used the 5-time-step, achieving an average accuracy of 78.01% and a return on investment (ROI) of 83.22% across the nine indices. The model’s accuracy decreased as the time step increased, which may be attributed to the fluctuating price levels of many of the included stocks. The experiments demonstrate the impact of time steps on classification performance that can be leveraged in trading applications, catering to short and long-term traders based on their strategies. Finally, a real-time trading simulation showed that the MLSC model generates significantly higher financial returns than benchmark models across all considered indices.</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

MLSC: A Multi-label Stock Classifier for Multi-horizon Stock Trend Prediction

  • 期刊论文

  • 发表日期：

    2025年4月17日

查看全文
  • Ibanga Kpereobong Friday,
  • Sarada Prasanna Pati,
  • Debahuti Mishra

摘要

This study proposes and implements a novel deep learning-based model, the multi-label stock classifier (MLSC), which combines a multi-output classification method with several components: a categorical input embedding, a multi-head attention block, residual connections, convolutional neural networks (CNNs), and gated recurrent units (GRUs). This combination enables the model to simultaneously predict stock trend movements across multiple time horizons (e.g., 1-day, 3-days, 7-days, and 10-days). This multi-label allows traders to simultaneously predict price trends over different time horizons, effectively supporting trading decisions by identifying consistent trend movements over a specified period rather than relying solely on the next-day trends. The architecture of the proposed model effectively addresses three key challenges: imbalanced target labels across different horizons where each timeframe predicts either of three classes: buy, hold, or sell for the corresponding stock, accurate multi-label output learning, and efficient capture of both short and long-term dependencies within the data. The classification performance of the MLSC model has been evaluated across nine different indices using a time-series cross-validation fold (where k = 3, 5, and 7) for various time steps (5, 10, and 15). The best results used the 5-time-step, achieving an average accuracy of 78.01% and a return on investment (ROI) of 83.22% across the nine indices. The model’s accuracy decreased as the time step increased, which may be attributed to the fluctuating price levels of many of the included stocks. The experiments demonstrate the impact of time steps on classification performance that can be leveraged in trading applications, catering to short and long-term traders based on their strategies. Finally, a real-time trading simulation showed that the MLSC model generates significantly higher financial returns than benchmark models across all considered indices.