对金融资产的价格趋势进行预测一直以来都是一项具有挑战性的任务。2010年股指期货的问世使得做空机制和风险对冲成为了可能，也意味着我国量化交易正式启程。经过十几年的发展，量化行业已经进入了由迅猛发展的“田园时代”到竞争更加激烈的“精细化发展时代”转变的关键节点。在这个过程中，随着高性能计算、低延迟网络等信息技术的持续进步，高频交易在市场中扮演着越来越重要的角色。然而，传统线性模型难以捕捉复杂的非线性关系和多变的市场动态，在处理高频的金融量价数据时往往效果有限。而近年来，深度学习作为一种强大的机器学习工具，在处理大规模数据和复杂模式识别方面取得了显著的进展。 在此背景下，本文将深度学习模型应用于股指期货市场，研究其对超短期价格趋势进行预测的能力，以期在传统模型的基础上进一步提升预测效果，具有重要的现实意义和较高的应用价值。研究选取中国金融期货交易所沪深300股指期货合约的1分钟高频数据作为样本，以2023年1月1日至2024年2月29日作为样本区间。 首先，本文研究和梳理了几种含义明确且金融逻辑稳健的因子构建思路，结合市场特点和金融实践经验，选择了多个高频量价因子进行构建，并对其中一些因子的结构进行了原创性改进。之后，为了验证因子有效性，我们基于每个因子分别构建单因子策略进行了回测分析，运用多种评价指标衡量策略的盈利能力。 最后，本文搭建了多种神经网络结构进行预测并对其性能进行了评估，这些网络模型分别是长短期记忆（LSTM）模型、门控循环单元（GRU）模型以及LSTM-GRU混合模型。研究结果显示，深度学习模型与传统模型相比，在预测准确性方面的表现均有较大幅度的提升，其中门控循环单元模型在所有模型中的综合表现最好，从而验证了深度学习在股指期货市场高频预测方面的有效性及优越性。
Forecasting asset price trends in financial markets has always been a challenging task. The advent of stock index futures in 2010 made short mechanism and risk hedging possible, and also meant the official start of quantitative trading in China. After more than ten years of development, the quantitative industry has entered a key turning point from the rapid development stage to a more competitive fine development stage. In this process, with the continuous advancement of technology, high-frequency trading plays a more important role in the market. Nonetheless, traditional linear models often struggle to capture complex nonlinear relationships and fluctuating market dynamics, thereby limiting their effectiveness when it comes to handling high-frequency financial data. However, the last few years have witnessed deep learning - a formidable tool in the machine learning suite - make substantial strides in the processing of vast data sets and recognition of complex patterns. Under such circumstances, this thesis harnesses the power of deep learning models to explore their predictive capabilities for ultra-short-term price trends in the stock index futures market. The intent is to augment traditional model-based predictions, thus offering important practical significance and substantial application value. The study selects 1-minute high-frequency data of CSI 300 stock index futures contracts on China Financial Futures Exchange as the sample, and takes January 1, 2023 to February 29, 2024 as the sample interval. At first, this thesis reviewed and sorted out several factor construction ideas that have clear meanings and robust financial logic. By combining market characteristics and financial practice experience, multiple high-frequency price-volume factors were selected and built. Moreover, some original improvements were implemented on the structure of certain factors. Following this, to confirm the validity of these constructed factors, we conducted tests and analyses for single-factor strategies based on each individual factor. A range of evaluation metrics were employed to assess the profit potential of these strategies. In the final phase of this research, a variety of neural network structures were constructed for prediction purposes and their performance was evaluated. The models utilized in this study include Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and a combined LSTM-GRU structure. The research results showed that compared to traditional models, deep learning models have significantly improved prediction accuracy. Among all the models, the Gated Recurrent Unit model exhibited the best overall performance. In summary, the conclusions above confirmed the effectiveness and superiority of deep learning in high-frequency prediction in the stock index futures market.
