基于分解-重构-集成框架的股指预测及行业轮动策略

Stock Index Prediction and Industry Rotation Strategy Based on a Decomposition-Reconstruction-Integration Framework

股指数据受到多重因素的影响,呈现出非线性、非平稳、高复杂度、高波动的特点,因此单一模型很难完整刻画其数据特征.本文提出了一个基于分解-重构-集成框架的股指收益率复合预测模型,利用变分模态分解(VMD)将原始的高复杂度股指时间序列进行分解,以复合多尺度熵(CMSE)作为重构指标,将股指数据分量重构为长期趋势项、中期影响项、短期扰动项,根据其数据特征分别利用ARIMA、BPNN、LSTM模型进行预测,最后将各频率项预测进行集成得到最终预测结果.本文将提出的方法应用于八个重要行业股指预测,并与以Fine to Coarse(FTC)、样本熵(SE)、模糊熵(FE)、多尺度排列熵(MSPE)作为重构方法的模型进行比较.随后,本文提出了两种行业轮动策略——等权投资与动态权重投资,从保守和激进两种角度验证所提出模型在实际交易中的性能.实证结果表明了在股指预测中,以复合多尺度熵作为重构指标优于其余重构指标.相较于基准模型,本文提出的复合模型能获得较低的预测误差及较高的方向精度,并且我们提出的行业轮动策略在风险及收益方面表现优异.

Stock index data is influenced by multiple factors,exhibiting nonlinear,non-stationary,high complexity,and high volatility characteristics.Therefore,it is difficult for a single model to fully capture its data features.This paper proposes a hybrid forecasting model for stock index returns based on a decompositionreconstruction-integration framework.Utilizing variational mode decomposition(VMD),the original high-complexity stock index time series is decomposed.The composite multiscale entropy(CMSE)is employed as a reconstruction indicator to reorganize the stock index data components into long-term trend terms,medium-term impact terms,and short-term disturbance terms.ARIMA,BPNN,and LSTM models are adopted for forecasting based on their respective data characteristics.Finally,the predictions of various frequency components are integrated to obtain the ultimate forecasting results.The proposed method is applied to forecasting eight significant industry stock indices and is compared with models utilizing fine to coarse(FTC),sample entropy(SE),fuzzy entropy(FE),and multiscale permutation entropy(MSPE)as reconstruction methods.Furthermore,two industry rotation strategies,equal-weight investment and dynamic-weight investment,are proposed to validate the performance of the proposed model in practical trading from both conservative and aggressive perspectives.The empirical results demonstrate that CMSE outperforms other reconstruction indicators in stock index forecasting.Compared to benchmark models,the hybrid model presented in this paper achieves lower forecasting errors and higher directional accuracy,and the proposed industry rotation strategies exhibit excellent performance in terms of risk and return.