计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (14): 269-281.DOI: 10.3778/j.issn.1002-8331.2010-0247

• 工程与应用 • 上一篇    下一篇

指数平滑与自回归融合预测模型及实证

包研科,陈然,郑宏杰,冯永安,王江   

  1. 1.辽宁工程技术大学 理学院,辽宁 阜新 123000
    2.辽宁工程技术大学 智能工程与数学研究院,辽宁 阜新 123000
    3.北京泛鹏天地科技股份有限公司,北京 100000
    4.辽宁工程技术大学 软件学院,辽宁 葫芦岛 125000
  • 出版日期:2022-07-15 发布日期:2022-07-15

Index Smoothing and Autoregressive Fusion Forecasting Model and Its Empirical Study

BAO Yanke, CHEN Ran, ZHENG Hongjie, FENG Yong’an, WANG Jiang   

  1. 1.School of Science, Liaoning Technical University, Fuxin, Liaoning 123000, China
    2.Institute of Intelligent Engineering and Mathematics, Liaoning Technical University, Fuxin, Liaoning 123000, China
    3.Beijing Vantage Point Technology Co., Ltd., Beijing 100000, China
    4.School of Software, Liaoning Technical University, Huludao, Liaoning 125000, China
  • Online:2022-07-15 Published:2022-07-15

摘要: 针对银行业务中的隔夜头寸预测问题,融合指数平滑与自回归预测的思想方法,提出了一个时间序列一步预测的微分动力学方程,证明了方程的离散化结构同无隐层BP神经网络的等价性。讨论了模型预测有效性问题,并进行了实证分析。对12个样本预报偏差的拟合优度检验表明,绝对预报偏差近似服从指数分布;实证分析了环比波动特征同模型预报误差以及预报同态度的关系,表明在一定条件下模型的预报是有效的,同NLP框架下的LSTM和GRU进行了对比实验表明该模型有更好的表现;定义了稳态指数和转折指数来描述时间序列的环比波动特征,分析表明稳态指数和转折指数可以预估模型预报的误差水平和同态度水平;研究了模型前端降噪对预报结果的影响,结果表明前端降噪可以抑制模型预报的“过敏”行为,有利于对时间序列变化趋势的动态跟踪。

关键词: 时间序列预测, 指数平滑, 自回归, BP神经网络, 实证分析

Abstract: To solve the problem of overnight position prediction in banking business, a differential dynamic equation for one-step prediction of time series is proposed by integrating exponential smoothing and autoregressive prediction. It is proved that the discretization structure of the equation is equivalent to that of BP neural network without hidden layer. Then, the prediction validity of the model is discussed and an empirical analysis is made. Firstly, the goodness of fit test of 12 samples shows that the absolute prediction deviation approximately follows the exponential distribution. Secondly, the empirical analysis of the relationship between the ring ratio fluctuation characteristics and the prediction error, as well as the relationship between ring ratio fluctuation characteristics and the prediction homomorphism rate, shows that the prediction of the model is effective under certain conditions. In addition, the model is compared with LSTM and GRU in NLP framework. It is found that the model in this paper has better performance. Thirdly, the steady-state index and turning index are defined to describe the month on month fluctuation characteristics of time series. The results reflect that the steady-state index and turning index can predict the error level and homomorphism level of the model prediction. Finally, the influence of the front-end noise reduction on the prediction results is discussed. The results are indicative that the front-end noise reduction can suppress the “allergic” behavior of the model prediction, which is conducive to the dynamic tracking of the change trend of the time series.

Key words: time series prediction, exponential smoothing, autoregression, BP neural network, empirical analysis