金融学院知识库

随着我国A股市场的不断完善与发展，传统的投资策略如基本面研究、技术面研究等已经很难使投资者在市场上获得超额收益。近年来，量化投资的概念进入人们视野，量化投资基金的整体规模逐渐增大，市场中量化交易的占比也越来越高。在未来这种新型投资方式将成为市场上主流的交易方式，同时也会成为金融机构争夺客户资源的主要工具。因此，研究量化投资策略对于我国量化投资行业的发展格外重要。

本文将遗传算法引入XGBoost模型的超参数调优过程，提高了模型的调参效率和预测准确率，在对策略进行历史数据回测后，获得了超过沪深300指数的收益。在选股模型的构建过程中，本文采用设置标签的方式对股票进行分类，设置得分评价方案对预测结果进行排序，同时使用滚动训练的方式来保证模型训练时使用到的是最新数据。在对遗传算法的设计上，使用F1-score作为算法的适应度函数。

通过因子重要性分析，本文发现在所使用的31个技术指标之中，波动率类指标如归一化的波动幅度均值（NATR）的分类表现最强，其次是重叠类指标如双指数移动平均线（DEMA）、考夫曼的自适应移动平均线（KAMA），周期类指标、成交量类指标等其他指标的分类效果不明显。

With the continuous improvement and development of China's A-share market, traditional investment strategies such as fundamental research and technical research have made it difficult for investors to obtain excess returns in the market. In recent years, the overall scale of quantitative investment funds has gradually increased, and the proportion of quantitative trading in the market has also increased. In the future, quantitative investment this new investment method will become the mainstream trading method in the market, and become the main tool for financial institutions to compete for customer resources. Therefore, the study of new quantitative investment strategies is particularly important for the development of quantitative investment in China.

In this paper, for the first time in the field of financial quantification, the genetic algorithm is introduced into the hyperparameter tuning process of the XGBoost model, the stock selection model based on the XGBoost algorithm is optimized, and the return exceeding the CSI 300 index is finally obtained through backtesting of historical data. Compared with the traditional quantitative stock selection model, our model can significantly improve the computational efficiency. At the same time, due to the selection of suitable parameters, the prediction accuracy of the model and the return rate of backtesting of historical data have been improved to varying degrees.

This paper also describes the construction process of the stock selection model in detail, and explains the setting of various aspects of the model. We use the method of setting labels to classify stocks, set score evaluation scheme to sort the prediction results, and select the rolling training method to ensure that the data used in model training is relatively new. In the design of genetic algorithm, F1-score is used as the fitness function of the algorithm.

Through factor importance analysis, we found that among the 31 technical indicators used, volatility indicators such as the normalized mean volatility (NATR) performed the strongest in the classification task, followed by overlapping indicators such as double-exponential moving average (DEMA), Kaufman's adaptive moving average (KAMA), and other indicators such as cycle indicators, volume indicators have less obvious classification effect.