creator Tian, Tian 整合深度学习和创新特征选择以提高期货市场短期价格预测 2024 Spring 2024 Dissertation Illinois Institute of Technology SSB / Stuart School of Business advisor Cooper, Ricky Management Finance Cophenetic Correlation Deep-learning Models(LSTMCNN_LSTMGRU-LSTM) Feature Selection Futures Markets Price Prediction en This study presents a novel approach for predicting short-term price movements in futures markets using advanced deep-learning models, namely LSTM, CNN_LSTM, and GRU_LSTM. By incorporating cophenetic correlation in feature preparation, the study addresses the challenges posed by sudden fluctuations and price spikes while maintaining diversification and utilizing a limited number of variables derived from daily public data. However, the effectiveness of adding features relies on appropriate feature selection, even when employing powerful deep-learning models. To overcome this limitation, an innovative feature selection method is proposed, which combines cophenetic correlation-based hierarchical linkage clustering with the XGBoost importance listing function. This method efficiently identifies and integrates the most relevant features, significantly improving price prediction accuracy. The empirical findings contribute valuable insights into price prediction accuracy and the potential integration of algorithmic and intuitive approaches in futures markets. Moreover, the developed feature preparation method enhances the performance of all deep learning models, including LSTM, CNN_LSTM, and GRU_LSTM. This study contributes to the advancement of price prediction techniques by demonstrating the potential of integrating deep learning models with innovative feature selection methods. Traders and investors can leverage this approach to enhance their decision-making processes and optimize trading strategies in dynamic and complex futures markets.这项研究提出了一种新颖的方法，使用先进的深度学习模型（即LSTM、CNN_LSTM和GRU_LSTM）预测期货市场的短期价格走势。通过在特征准备中纳入cophenetic相关性，该研究解决了由突然波动和价格尖峰带来的挑战，同时保持多样化并利用来自日常公开数据的有限数量变量。然而，即使使用强大的深度学习模型，增加特征的效果也取决于适当的特征选择。为了克服这一局限性，提出了一种创新的特征选择方法，它结合了基于cophenetic相关性的层次链接聚类和XGBoost的重要性列表功能。这种方法有效地识别和整合了最相关的特征，显著提高了价格预测的准确性。实证研究结果为价格预测准确性和算法与直觉方法在期货市场中的潜在整合提供了宝贵的见解。此外，开发的特征准备方法提高了所有深度学习模型（包括LSTM、CNN_LSTM和GRU_LSTM）的性能。这项研究通过展示深度学习模型与创新特征选择方法的整合潜力，为价格预测技术的进步做出了贡献。交易员和投资者可以利用这种方法来增强他们的决策过程，并优化动态和复杂的期货市场交易策略。 born digital application/pdf In Copyright http://rightsstatements.org/page/InC/1.0/ Restricted Access http://hdl.handle.net/10560/islandora:1025166