使用人工智能领域的遗传规划算法构建交易策略是当前投资实务界和学术界日趋活跃的重要研究问题.现有遗传规划策略的适应度都是静态的单目标, 并未准确刻画自然界和金融市场真实的环境特点, 文章提出一种随机性多目标适应度策略.首先将文献中单目标策略在沪深300指数上进行测评, 然后以夏普比率(Sharpe)、信号正确率(Accuracy)、收益回撤比(refdown) 三种指 标为基础, 在权重中引入随机系数项构建随机性多目标适应度, 并基于强类型遗传规划和滚动窗口交易模式, 构建出一种{STGP-RMO-Rolling}策略.在对沪深300的实证分析中, 其样本外收益率较单目标策略平均提高10\%以上, 夏普比率提高2.0以上, 对上证50、中证500的测试也取得类似效果.进一步在沪深300股指期货上进行了长达7年的测试分析, {STGP-RMO-Rolling}获得了84.47\%的超额收益率, 尤其在熊市中表现出色, 不过在牛市难以战胜``买入-持有''策略.对滚动窗口的训练、预测长度进行敏感性分析, 发现最佳参数组合为125/20, 且最优参数附近呈现出可信的参数平原特征.

Using genetic programming, an artificial intelligent algorithm, to find trading strategies has become an active research in theoretical and practical circles. The existing genetic programming strategies are all studied based on single target fitness. This paper proposes a multi-objective strategy. We construct the random multi-objective based on three basic fitness, Sharpe ratio, accuracy and sterling ratio, and introduce the random constant into the weigh. Apart from these, strongly typed genetic programming algorithm and rolling window are also used to construct the STGP-RMO-Rolling strategy. In the empirical analysis on CSI 300, the out-of-sample excess return has been improved by more than 10\%, and Sharpe ratio improved by more than 2.0, which is similar to the performance on SSE 50 and CSI 500 index. We make a further test on IF300 futures for a seven-year period, and obtain an excess yield of 84.47\%, and find it performs well during bear and volatile market. During the sensitivity analysis for the training and prediction length of the rolling window, the best parameters are 125/20, and present the reliable plain for the parameter.