Research And Application Of Long And Short Term Forecasting Model Of Financial Time Series Based On ARIMA And Recurrent Neural Network

Financial time series is a special high-noise dynamic time series,showing complex features of discrete and non-stationary.Analyzing,predicting and controlling such data is a requirement in the context of the era of artificial intelligence,and it is also the key research object of technologies such as robo advisors in financial technology.The time series data in this paper is selected by taking the CSI 300 index on the Chinese securities market as an example.,and the validity of the model prediction is evaluated through quantitative indicators RMSE,MAE,SMAPE,MAPE and R~2.The main research contents of this paper are divided into the following three points:(1)A test for non-stationary and discrete characteristics of financial time series.In this paper,statistics such as skewness and kurtosis are used to describe the discrete degree of time series,and frequency statistics are used to describe the distribution state of data.By decomposing the constituent elements of the time series,the hidden internal characteristics of the time series are excavated,and the non-stationarity of the time series is tested by ACF,PACF,ADF and KPSS.(2)Aiming at the characteristics of thick tails,peaks and long-term dependence of financial time series,this paper selects the ARIMA model to predict the medium and long-term trend of time series.This paper improves the empirical flow of the model from parameter tuning to diagnostic testing,and determines ARIMA(3,2,3)as the best model choice according to AIC,BIC and HQIC criteria.The research shows that the model has strong linear trend forecasting ability,and the long-term forecasting performance on the CSI 300 index is better.(3)In view of the nonlinear,high volatility and random characteristics of financial time series,this paper selects Simple RNN,LSTM and GRU three kinds of recurrent neural networks to build a short-term forecasting model of time series.This paper improves the structure and training methods of three recurrent networks,and analyzes the experimental process from algorithm optimization to performance evaluation.The research shows that the three kinds of recurrent neural networks can achieve high-precision short-term T+1 prediction,and the improved GRU network is better than the other two networks in terms of training time and prediction accuracy.When solving practical problems,the model can be applied to the prediction of other non-stationary discrete time series according to the training method and parameter settings in this paper,or the data in this paper can be replaced with tick level,so as to realize the modeling and prediction of high-frequency financial data.This paper provides a practical basis and reference for the modeling of time series problems in the domestic financial market,and also provides a more scientific alternative for the decision-making of financial institutional investors.In addition,it also has certain reference value for the practical application of artificial intelligence technology in the financial field.