基于人工蜂群优化循环神经网络的财务危机预测

Financial crisis prediction based on artificial bee colony optimized recurrent neural network

为了提高财务危机预测的性能,采用循环神经网络(Recurrent neural network,RNN)对关键指标进行分析训练,以解决因为时间变化带来的深度学习网络预测准确率性能下降的问题。选取关键指标特征生成预测样本,并充分利用RNN在时间序列的循环计算优势,采用差异化时间序列的赋权策略,记忆不同历史时间序列对RNN预测分析的影响。经过RNN训练,并采用隐藏层输出不断循环的方式,将历史时间段输入不断作用于当前训练输出。引入人工蜂群(Artificial bee colony,ABC)算法在RNN反向传播过程中对时间序列权重进行更新。将RNN网络输出值与预测值的均方误差作为ABC的适用度函数,获得全局最优的ABC-RNN预测模型。试验证明,合理优化历史时间序列权重,能够获得较高的危机预测准确率。和常用预测算法对比,所提ABC-RNN算法的预测准确率更高,且曲线面积(Area under the curve,AUC)值更高。

In order to improve the performance of financial crisis prediction,the recurrent neural network(RNN)is used to analyze and train key indicators,so as to solve the problem that the prediction accuracy of deep learning network decreases due to time change.Firstly,select key index features to generate prediction samples,and make full use of the advantages of RNN in the cyclic calculation of time series,and adopt the weighting strategy of differentiated time series to remember the influence of different historical time series on RNN prediction analysis.Then,after RNN training,the historical time period input is continuously applied to the current training output by using the hidden layer output to continuously circulate.Secondly,artificial bee colony(ABC)algorithm is introduced to update the weight of time series in the process of RNN back propagation.The mean square error between the output value of RNN network and the predicted value is taken as the fitness function of ABC so as to obtain the globally optimal ABC-RNN prediction model.Experiments show that reasonably optimizing the weight of historical time series can obtain higher accuracy of crisis prediction.Compared with common prediction algorithms,the proposed ABC-RNN algorithm has higher prediction accuracy and higher value of area under curve(AUC).