统计与数据科学学院知识库

  近年来，随着人工智能和文本挖掘技术的崛起，时间序列预测已成为各个领域关注的热点。然而，变幻莫测的经济环境使得时间序列的特征和影响因素异常复杂，进而致使先前取得的一些成果可能无法满足新环境下人们对时间序列预测的需求。因此，在前人研究的基础上进一步探索和研究时间序列预测方法具有重要的理论价值和应用价值。

  先前很多研究表明，分解集成策略是针对复杂时间序列预测的一个有效预测框架，但大多数研究忽视了在分解集成框架下开展系统的分解方法、深度学习预测技术及多模态预测因子的研究。因此，本文从分解方法、深度混合学习预测技术及多模态预测因子出发，并结合混合智能优化算法开展了一系列时间序列预测方法的研究。具体研究内容和主要结论包括以下四个方面：

  （1）借助一系列统计检验方法分析了研究数据包含的复杂特性，以为后文预测模型中方法的选择提供参考。实验结果表明，研究数据具有非线性、混沌性、长记忆性和递归性特征。基于上述检验结论，本文在后续预测模型中采用了降低数据复杂性的分解集成策略，并在预测技术方面侧重采用了能捕捉序列复杂特征的深度学习和混合预测技术，以及从导致数据复杂特征的源头选择了多模态预测因子。

  （2）先前很多研究采用了分解集成的思想进行了时间序列的预测，但鲜有研究对分解方法进行系统的对比研究。为此，本文以系统研究分解方法为目的，构建了基于混沌理论和不同分解方法的混合点预测模型。实证研究结果表明，基于多变量分解方法的模型比基于单变量分解方法的模型有更好的预测性能，且在单变量分解方法中基于改进自适应噪声完备集合经验模态分解方法模型的预测性能略优于基于变分模态分解方法的模型，但明显优于基于其他分解方法的模型。同时，在多变量分解方法中基于多元变分模态分解模型的预测性能优于基于多元经验模态分解方法的模型。

  （3）针对单变量的时间序列预测及前期分解集成预测研究中多数关注单一人工智能预测技术的问题，本文选择第二部分中的最优单变量分解方法为分解技术，并构建基于深度混合预测技术及多目标混合智能优化算法的混合点预测模型。研究结果表明，构建模型较基准模型是最优的。同时，在所有设置的预测技术中，基于本文构建的深度混合学习预测技术模型的预测性能是最优的，其次是基于单一深度学习的，紧接着是基于机器学习的，最后是基于计量经济学方法的。此外，为了研究不同集成方法对预测结果的影响，本文也对单个非线性集成和基于多目标混合灰狼优化算法的线性集成进行了简单的对比研究。结果表明多目标混合灰狼优化算法的线性集成略胜于单一技术的非线性集成。

  （4）针对多模态影响因素的时间序列预测及大多数现有研究关注点预测，忽视区间预测的问题，本文选择第二部分研究内容中的最优多变量分解方法作为分解技术，并构建基于多模态预测因子与单目标混合智能优化算法结合的混合区间预测模型。研究结果表明，构建模型较基准模型是最优的。同时，基于“历史+新闻+金融”预测因子模型的点预测和区间预测性能优于基于其他预测因子的模型。

  本文的主要贡献和创新为：从整体而言，本文融合预测模型的基本要素并结合不同方法的优点搭建了时间序列预测的理论框架，且从方法论的视角而言，每个模型可独立于其他模型。从具体构建的系列模型来看，首先，本文构建了基于混沌理论和不同分解方法的预测模型，并侧重研究了不同分解方法的分解效果。该研究结论对非平稳、强波动时间序列的预测具有重要意义，特别是对研究中分解方法的选择具有重要的参考价值。同时，这也可能是对单变量和多变量分解方法的首次系统研究。此外，该模型中的另一个创新是在特征选择阶段构造了基于数据混沌特性的多阶段特征选择方法，该方法的优点是对不同混沌特性的序列采用了不同滞后阶数确定的方法。其次，以研究预测技术为侧重点，本文构建了基于深度混合预测技术和多目标混合智能优化算法的预测模型。该模型中的创新为：一是基于卷积神经网络和不同特征循环神经网络模型构建了深度混合预测技术；二是通过改进灰狼算法中的线性递减收敛因子优化了原始多目标灰狼优化算法的性能。再次，以研究预测因子为重点，本文构建了基于不同数据模态及单目标混合智能优化的混合区间预测模型。该模型中的创新为：一是在多变量分解集成框架下研究多模态预测因子组合下模型的区间预测性能；二是借助蜻蜓优化算法的权重因子思想和改进灰狼优化算法中的线性递减收敛因子优化了原始单目标灰狼优化算法的性能。

    In recent years, with the rise of artificial intelligence and text mining technology, time series forecasting has become a hot topic in various fields. However, the unpredictable economic environment makes the characteristics and influencing factors of time series more complicated, which makes some achievements previously obtained may not meet the needs of people for time series forecasting in the new environment. Therefore, it is of great theoretical and application value to further explore and study time series forecasting methods on the basis of previous studies.

    Although many previous studies have shown that the decomposition and ensemble strategy is an effective forecasting framework for the forecasting of complex time series, most studies have ignored the research for decomposition methods, deep learning forecasting techniques, and multimodal predictors under the decomposition-ensemble framework. Therefore, to further improve the accuracy of time series forecasting, this paper carries out a series of research for decomposition method, deep mixed learning forecasting technology, multimodal predictor, and hybrid intelligent optimization algorithm. The specific research content and main conclusions include the following four aspects:

    (1) A series of statistical test methods are used to analyze the complex characteristics contained in the research data, so as to provide reference for the selection of methods in the following forecasting model. The experiment results show that the study data has the characteristics of nonlinear, chaotic, long memory, and recursion. Based on the test conclusions above, this paper adopts the decomposition-ensemble strategy to reduce data complexity in the subsequent forecasting model, and in terms of forecasting technology, this paper focuses on deep learning and hybrid forecasting technology that can capture complex sequence features. In addition, the multimodal predictors are selected from sources that lead to complex features.

    (2) Although previous many studies have used the idea of decomposition-ensemble to forecast time series, few studies have conducted systematic comparative studies for decomposition methods. In order to study the decomposition methods systematically, this paper constructs a mixed point forecasting model with chaos theory and different decomposition methods. The empirical study shows that the model with multivariate decomposition method has better forecasting performance than that of the model with univariate decomposition method. In univariate decomposition method, the prediction performance of the model with improved complete ensemble empirical mode decomposition with adaptive noise method is slightly better than that of the model with variational mode decomposition method, but is obviously better than that of the model with the other decomposition methods. At the same time, the forecasting performance of the model with multivariate variational mode decomposition is better than that of the model with multivariate empirical mode decomposition.

    (3) For univariate time series forecasting, and the problem that most researches on decomposition-ensemble forecasting focused on the single artificial intelligence forecasting technology, this paper chooses the optimal univariate decomposition method in the second part as the decomposition technology, and builds a mixed point forecasting model with deep hybrid forecasting technology and multi-objective hybrid intelligent optimization algorithm. The results show that the constructed model is optimal compared with the benchmark models. At the same time, among all the predictive technologies set up, the forecasting performance of the model with the deep mixed learning predictive technology constructed in this paper is the best, followed by the model with single deep learning, next the model with the machine learning, and finally the model with the econometric methods. In addition, to study the influence of different ensemble methods on the forecasting results, the single nonlinear integration and linear integration with multi-objective hybrid gray wolf optimization algorithm are also studied. The results show that the linear integration of multi-objective hybrid gray wolf optimization algorithm is slightly better than the nonlinear integration of single technique.

    (4) For the time series forecasting of multimodal influencing factors and the problem that most existing studies focus on value forecasting while ignoring interval forecasting, this paper chooses the optimal multivariable decomposition method in the second part as the decomposition technology, and a hybrid interval forecasting model with multimodal predictor and single objective hybrid intelligent optimization algorithm is constructed. The experimental results show that the constructed model is optimal compared with the benchmark models. Meantime, the performance of point forecasting and interval forecasting models with the “history + news + finance” is better than that of the model with other predictors.

    The main contributions and innovations of this paper are as follows: As a whole, this paper combines the basic elements of forecasting models and the advantages of different methods to build a theoretical framework for time series forecasting, and from the perspective of methodology, each model can be independent of other models. From the perspective of specific constructed models, firstly, this paper constructs a forecasting model with chaos theory and different decomposition methods, and focuses on the decomposition effects of different decomposition methods. The research conclusions in this model are of great significance for the forecasting of non-stationary and highly fluctuating time series; in particular, it has important reference value for the choice of decomposition method in the research. This may be the first systematic study for univariate and multivariate decomposition methods. In addition, another innovation in the model is to construct a feature selection method with data chaos characteristics in the feature selection stage, and its advantage is that sequences with different chaos characteristics adopt different hysteresis order determination methods. Secondly, focusing on the research of forecasting technology, this paper constructs a forecasting model with deep hybrid forecasting technology and multi-objective hybrid intelligent optimization algorithm. The first innovation in the model is to build a deep hybrid forecasting technique with convolutional neural network and recurrent neural network of different feature. The second innovation is to optimize the performance of the original multi-objective grey wolf optimization algorithm by improving the linear decreasing convergence factor. Thirdly, focusing on the study of predictors, this paper constructs an interval forecasting model with different data modes and single objective hybrid intelligent optimization. The first innovation of the model is to study the interval forecasting performance of the model under the combination of multi-source predictors. The second innovation is to optimize the performance of the original single-objective gray wolf optimization algorithm by using the weight factor idea of dragonfly optimization algorithm and improving the linear decreasing convergence factor of original gray wolf optimization algorithm.