可解释机器学习与股指收益率预测的经济价值---来自中国市场的证据

Interpretable machine learning and the economic value of stock index return forecasts: Evidence from Chinese stock market

摘要

本文针对上证综指和深证成指,搭建了一个同时具备预测能力和可解释性的多模态混合框架.输入端汇集了宏观经济、市场交易、企业盈利和投资者情绪4类信息.建模时,先用CEEMDAN把收益率序列按尺度分解,由LightGBM挑出各频段的关键特征,再分别用Attention-LSTM、Attention-GRU和LSTM拟合高、中、低频分量,最后由SVR做非线性集成.在样本外检验中,混合模型在两个市场都取得了正的调整后样本外R²---上证0.48%,深证1.23%,均优于线性基准和单一深度学习模型.SHAP分析显示,真正驱动预测的主要是宏观货币与价格类变量.择时方面,在long-cash策略、0.1%交易成本的设定下,混合模型相对买入持有改善了风险调整后绩效.这些结果为多模态机器学习用于股指收益率预测、并兼顾可解释性与经济价值,提供了来自中国市场的经验证据.

Abstract

This paper builds a multimodal hybrid framework for the Shanghai Composite index and the Shenzhen Component index that aims to be both predictive and interpretable. On the input side it draws on four kinds of information-macroeconomic conditions, market trading, corporate earnings, and investor sentiment. The modeling proceeds in stages: CEEMDAN first decomposes the return series by scale, LightGBM selects the key features within each frequency band, Attention-LSTM, Attention-GRU, and LSTM then fit the high-, medium-, and low-frequency components respectively, and SVR carries out the nonlinear ensemble. The hybrid model attains a positive adjusted external examination R ² in both markets-0.48% for Shanghai and 1.23% for Shenzhen-outperforming linear benchmarks and a single deep-learning model. SHAP analysis shows that the predictions are driven mainly by monetary and price-related macroeconomic variables. On the timing side, under a long-cash strategy with a 0.1% transaction cost, the hybrid model improves risk-adjusted performance relative to buy-and-hold. Taken together, these results offer evidence from the Chinese market on the interpretability and economic value of multimodal machine learning for stock index return forecasting.

关键词

股指收益率预测 / 多模态数据 / 可解释机器学习 / SHAP / 市场择时

Key words

stock index return forecasting / multimodal data / interpretable machine learning / SHAP / market timing

参考文献

[1] Gu S H, Kelly B, Xiu D C . Empirical asset pricing via machine learning[J]. The Review of Financial Studies, 2020, 33(5): 2223-2273.

[2] Fama E F . Efficient capital markets: II[J]. The Journal of Finance, 1991, 46(5): 1575-1617.

[3] Lo A W . The adaptive markets hypothesis: market efficiency from an evolutionary perspective[J]. Journal of Portfolio Management, 2004, 30(5): 15-29.

[4] 刘莉, 郝显峰, 王玉东. 基于时变稳健加权最小二乘法的股市收益率预测[J]. 管理科学学报, 2024, 27(1): 141-158.

[5] Liu L, Hao X F, Wang Y D. Forecasting stock returns: a time-varying robust weighted least squares approach[J]. Journal of Management Sciences in China, 2024, 27(1): 141-158.

[6] Masini R P, Medeiros M C, Mendes E F . Machine learning advances for time series forecasting[J]. Journal of Economic Surveys, 2023, 37(1): 76-111.

[7] Engle R F, Ghysels E, Sohn B . Stock market volatility and macroeconomic fundamentals[J]. Review of Economics and Statistics, 2013, 95(3): 776-797.

[8] Sezer O B, Gudelek M U, Ozbayoglu A M . Financial time series forecasting with deep learning: a systematic literature review: 2005-2019[J]. Applied Soft Computing, 2020, 90: 106181.

[9] Chen L Y, Pelger M, Zhu J . Deep learning in asset pricing[J]. Management Science, 2024, 70(2): 714-750.

[10] 李斌, 龙真. 中国股票市场可预测性研究:基于机器学习的视角[J]. 管理科学学报, 2023, 26(10): 138-158.

[11] Li B, Long Z. Return predictability in the Chinese stock markets: a machine learning perspective[J]. Journal of Management Sciences in China, 2023, 26(10): 138-158.

[12] 李斌, 屠雪永. 基于机器学习和资产特征的投资组合选择研究[J]. 系统工程理论与实践, 2024, 44(1): 338-359.

[13] Li B, Tu X Y. Portfolio selection based on machine learning and asset characteristics[J]. Systems Engineering-Theory & Practice, 2024, 44(1): 338-359.

[14] Tang Y J, Song Z Y, Zhu Y L, et al. A survey on machine learning models for financial time series forecasting[J]. Neurocomputing, 2022, 512: 363-380.

[15] 范小云, 王业东, 王道平, 等 . 不同来源金融文本信息含量的异质性分析:基于混合式文本情绪测度方法[J]. 管理世界, 2022, 38(10): 78-95.

[16] Fan X Y, Wang Y D, Wang D P, et al. Heterogeneity analysis of information content for financial text from different sources: a hybrid text sentiment measurement method[J]. Journal of Management World, 2022, 38(10): 78-95.

[17] 姜富伟, 刘雨旻, 孟令超. 大语言模型、文本情绪与金融市场[J]. 管理世界, 2024, 40(8): 42-59.

[18] Jiang F W, Liu Y M, Meng L C. Large language model and textual sentiment analysis in Chinese stock markets[J]. Management World, 2024, 40(8): 42-59.

[19] 张宗新, 吴钊颖. 媒体情绪传染与分析师乐观偏差:基于机器学习文本分析方法的经验证据[J]. 管理世界, 2021, 37(1): 170-185.

[20] Zhang Z X, Wu Z Y. Media sentiment contagion and analysts' optimistic bias: empirical evidence based on machine learning text analysis[J]. Management World, 2021, 37(1): 170-185.

[21] 姚加权, 冯绪, 王赞钧, 等 . 语调、情绪及市场影响:基于金融情绪词典[J]. 管理科学学报, 2021, 24(5): 26-46.

[22] Yao J Q, Feng X, Wang Z J, et al. Tone, sentiment and market impacts: The construction of Chinese sentiment dictionary in finance[J]. Journal of Management Sciences in China, 2021, 24(5): 26-46.

[23] Devlin J, Chang M W, Lee K, et al. BERT: pre-training of deep bidirectional transformers for language understanding[C]// Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (long and Short Papers). [S. l.: s. n.], 2019: 4171-4186.

[24] Gan B Q, Alexeev V, Bird R, et al. Sensitivity to sentiment: News vs social media[J]. International Review of Financial Analysis, 2020, 67: 101390.

[25] Lin Y, Yan Y, Xu J L, et al. Forecasting stock index price using the CEEMDAN-LSTM model[J]. The North American Journal of Economics and Finance, 2021, 57: 101421.

[26] Cao J, Li Z, Li J . Financial time series forecasting model based on CEEMDAN and LSTM[J]. Physica A: Statistical Mechanics and its Applications, 2019, 519: 127-139.

[27] Li Y R, Zhu Z F, Kong D Q, et al. EA-LSTM: Evolutionary attention-based LSTM for time series prediction[J]. Knowledge-Based Systems, 2019, 181: 104785.

[28] 胡青渝, 陈其安. 基于先验前馈神经网络的股票市场收益预测[J]. 系统工程理论与实践, 2025, 45(10): 3287-3303.

[29] Hu Q Y, Chen Q A. Stock market return prediction based on prior feedforward neural networks[J]. Systems Engineering-Theory & Practice, 2025, 45(10): 3287-3303.

[30] 戴星宇, 王群伟. 融合谱分解和傅里叶系数的多元区间值时序模型及其预测应用[J]. 系统工程理论与实践, 2025, 45(7): 2385-2404.

[31] Dai X Y, Wang Q W. A multi-variate ITS model with spectral decomposition and Fourier-coefficient[J]. Systems Engineering-Theory & Practice, 2025, 45(7): 2385-2404.

[32] 陈凯杰, 唐振鹏, 吴俊传, 等 . 基于分解-集成和混频数据采样的中国股票市场预测研究[J]. 系统工程理论与实践, 2022, 42(11): 3105-3120.

[33] Chen K J, Tang Z P, Wu J C, et al. Forecasting China's stock index: a hybrid method based on decomposition-integrated and mixed-frequency data[J]. Systems Engineering-Theory & Practice, 2022, 42(11): 3105-3120.

[34] 林昱, 常晋源, 黄雁勇. 融合经验模态分解与深度时序模型的股价预测[J]. 系统工程理论与实践, 2022, 42(6): 1663-1677.

[35] Lin Y, Chang J Y, Huang Y Y. On the prediction of the stock price based on empirical mode decomposition and deep time series model[J]. Systems Engineering-Theory & Practice, 2022, 42(6): 1663-1677.

[36] Lundberg S M, Lee S I . A unified approach to interpreting model predictions[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. [S. l.]: ACM, 2017: 4768-4777.

[37] Hassija V, Chamola V, Mahapatra A, et al. Interpreting black-box models: a review on explainable artificial intelligence[J]. Cognitive Computation, 2024, 16(1): 45-74.

[38] Bandt C, Pompe B . Permutation entropy: a natural complexity measure for time series[J]. Physical Review Letters, 2002, 88(17): 174102.

基金资助

国家社会科学基金(23BJY014)