主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2015, Vol. 23 ›› Issue (10): 47-56.doi: 10.16381/j.cnki.issn1003-207x.2015.10.006

• 论文 • 上一篇    下一篇

基于变量选择和遗传网络规划的期货高频交易策略研究

陈艳1, 王宣承2   

  1. 1. 上海财经大学统计与管理学院, 上海 200433;
    2. 深圳市福田区发展研究中心, 广东 深圳 518048
  • 收稿日期:2014-07-30 修回日期:2015-06-18 出版日期:2015-10-20 发布日期:2015-10-24
  • 作者简介:陈艳(1982-),女(汉族),四川泸州人,上海财经大学统计与管理学院,副教授,博士,研究方向:金融风险管理、商务智能.
  • 基金资助:

    国家自然科学基金资助项目(71271128,71331006,71571113);长江学者和创新团队发展计划(上海财经大学,IRT13077);上海财经大学创新团队支持计划

A Study on High-Frequency Futures Trading Strategy Based on Variable Selection and Genetic Network Programming

CHEN Yan1, WANG Xuan-cheng2   

  1. 1. School of Statistics and Management, Shanghai University of Finance and Economics, Shanghai 200433, China;
    2. Development Research Center of Futian District, Shenzhen 518048, China
  • Received:2014-07-30 Revised:2015-06-18 Online:2015-10-20 Published:2015-10-24

摘要: 高频交易在当前国际金融市场上炙手可热,股指期货的推出、融资融券和转融通业务的开通,使得我国高频交易市场初现端倪。本文立足于我国金融衍生品市场的现状提出了基于LASSO变量选择方法和遗传网络规划的期货高频交易策略。该策略首先使用LASSO从众多技术指标中,选出极少数最有效的指标作为判断函数,然后通过一种进化算法遗传网络规划来搜索合适的买点和买点,从而构建交易策略,并以黄金、铝和橡胶期货的5分钟高频交易数据为例进行回测检验。结果显示:第一,与最优子集法相比,LASSO方法在不降低预测精度的情况下,选出的指标数量最少,且均集中在趋势指标和震荡指标中。第二,通过结合遗传网络规划模型与Q强化学习法,搜索效率得到了显著提高,构建出适合于衍生品市场的简洁有效的交易策略,且在不同品种的期货交易中均超越了"买入并持有"策略,并获取超额收益,在量化投资领域充分体现了实践价值。

关键词: 遗传网络规划, 变量选择, Q强化学习, 智能计算, 高频交易

Abstract: In this paper, high-frequency futures trading strategy has been built by LASSO variable selection method and the genetic network programming (GNP). The proposed strategy uses LASSO as a variable selection method, which is able to select the most effective variables from a large number of technical indicators. Then, the selected indicators will be treated as the judgment functions in GNP to determine the buying and selling points. The 5 minutes high-frequency futures trading data of gold, aluminum and rubber is also used as an example for backtesting. The results show that: First, when compared with the optimal subset method, LASSO method selects the least number of indicators, while it shows almost the same prediction accuracy and better robustness. These indicators are mainly applied to describe the trend and shock. Second, GNP gets higher search efficiency and builds a simple and effective trading strategy after combining with Q reinforcement learning method. The proposed method outperforms the "Buy & Hold" strategy in different futures contracts and finally obtains the excess returns. It shows its practical value in the field of quantitative trading.

Key words: genetic network programming, variable selection, Q reinforcement learning, intelligent computing, high-frequency trading

中图分类号: