【作者】 刘涛；

【导师】 杨天奇；

【作者基本信息】 暨南大学 ， 计算机技术（专业学位）， 2017， 硕士

【摘要】 随着ATM的普及,关于ATM现金流的预测成了各大银行关注的热点和难点问题,为了确保银行在保持良好运营的基础上还能使得利益最大化,所以准确预测ATM中的现金流量对于银行意义重大。由于影响ATM现金流的不确定因素有很多,利用单一的预测算法往往很难满足我们的预测需求,所以本文提出利用组合预测算法来实现ATM现金流的预测。通过对目前国内外ATM现金流预测算法的研究,本文选取了BP神经网络,支持向量机(SVM),时间序列分析(ARIMA)和统计分析四种预测方法作为组合预测算法的预测单项。基于上述四种预测方法的研究发现,BP神经网络存在误差收敛速度慢,学习时间过长的主要原因在于反向传播过程中保持固定的学习率,所以对BP神经网络提出的改进是反向传播过程中对于不同的节点采取不同的学习率。时间序列分析作为统计学中常用的预测模型,在确定模型参数阶段存在人为的误差,所以对时间序列分析(ARIMA)提出误差最小模型最优的动态确定模型参数的改进方法。最后通过具体实验对四种预测方法的预测性能优势做出对比。组合预测算法的实质是结合各单项预测算法的优势,通过系统自动调整组合内各预测单项对应的权值,从而得到一个精确度高、适应性强的预测算法。基于对现有组合预测算法的研究,本文在确定各预测单项权值时提出了指数变权值组合预测法,并通过具体的实验验证了预测相同时间段的现金流数据时组合预测算法的预测精度要优于各单项预测算法。本文的最终目的是结合SpringMVC框架开发出一个以组合预测算法为核心算法的ATM现金流预测系统。通过对预测系统的需求分析以及功能页面的展示,最后总结全文,提出本文研究过程中存在的不足之处以及后续研究的方向,展望未来。更多还原

【Abstract】 With the popularity of ATM,the prediction of ATM cash flow has become a hot and difficult issue for major banks.To maximize the benefits of banks on the ground of maintaining good operation,the accurate prediction of ATM cash flow is of great importance.For ATM cash flow is influenced by many uncertain factors,the use of one single prediction could not satisfy our prediction needs.So against this background,a combined prediction algorithm is proposed in the prediction of ATM cash flowBased on the research of ATM cash flow prediction algorithm at home and abroad,four prediction methods are selected as the prediction individuals,including BP artificial neural network,Support Vector Machine(SVM),Autoregressive Integrated Moving Average Model(ARIMA)and statistical analysis.By studying the above four prediction methods,it can be found out that BP artificial neural network has the disadvantages of slow error convergence rate and long learning time due to fixed learning rates in the back propagation process,so the different learning rates are adopted in the back propagation process which can improve BP artificial neural network.ARIMA,as a frequently used prediction model in statistics,will lead to artificial errors in determining the parameters of models.Therefore,to improve ARIMA,the method of identifying model parameters dynamically is put forward,which is the optimal option with minimum errors.In the end,the comparisons of the prediction advantages of the four prediction methods through concrete experiments are given.The essence of combined prediction algorithm is obtaining a high-accuracy and strong-adaptability prediction algorithm by combining the advantages of each single prediction algorithm and through the automatic adjustment of the weights of each prediction model in the system.Based on the study of the existing combined prediction algorithm,the combined prediction method of changing indexes into weights is proposed when determining the weights of each single prediction algorithm.The result of concrete experiment proves that combined prediction algorithm has higher accuracy than each single prediction algorithm in predicting the cash flow data of the same period of time period.The ultimate goal of this thesis is to develop an ATM cash flow prediction system that takes the combined prediction algorithm as the core algorithm by combining the Spring MVC framework.Finally,through the analysis of the requirements of the prediction system and the display of the functional pages,the summary of this thesis is made by pointing out the shortcomings of the research process and the direction of the future search and by looking into the future.更多还原