Model Identification For ARCH Time Series Through Convolutional Neural Networks

In the field of modern finance,in order to be able to scientifically analyze the uncertainty of financial markets,the most common method is to model the financial time series.However,in the traditional financial econometric model,there is an assumption of independent homoskedasticity.Now,a large number of studies on financial markets have shown that the variance is always changing,resulting in the traditional financial econometric model cannot accurately describe the Economic activity.It wasn't until Professor Engle(1982)first proposed the concept of autoregressive conditional heteroscedasticity(ARCH for short),which made a breakthrough in the modeling of time-varying volatility.Professor Engle thus obtained Nobel Economics in 2003 prize.After the ARCH model was proposed,it quickly became a non-linear financial time series model that is very popular with researchers.Convolutional Neural Network(CNN)is currently a kind of neural network with very high learning efficiency.One of the characteristics of CNN is to combine the process of feature extraction and classification to train the neural network.In image classification and large-scale continuous Great success in speech.Applying CNN to other fields is a very important work at present.We experimentally compare the three methods of the traditional convolutional neural network model,the newly proposed"Double addition" network model,AIC?BIC and the likelihood function criterion.We find that the "Double addition" network model performs best on the ordering problem of the ARCH model.When we use the ARCH model,a very important step is to determine the order of the ARCH model.In this paper,we first use a traditional convolutional neural network to predict the order of the autoregressive conditional heteroscedasticity model(ARCH model).Secondly,we propose an improved double addition residual neural network structure based on the traditional residual convolutional neural network.