Exchange Rate Prediction Based On SRU Deep Learning

On January 1,1994,China merged the official exchange rate with the foreign exchange adjustment price,which opened the process of RMB market reform.On July 21,2005,the People's Bank of China announced the reform plan for the RMB exchange rate formation mechanism and implemented the “floating exchange rate system based on market supply and demand,with reference to a basket of currency adjustments and management”.The fluctuation of the non-US dollar against RMB spot exchange rate has been expanded from 1.5% to 3%.In 2007,the floating range of the US dollar against the RMB spot exchange rate was expanded from 3‰ to 5‰.With the deepening of market-oriented reforms,the 8.11 exchange reform in 2015 further liberalized the control of the exchange rate middle price parity,publicized the RMB exchange rate intermediate price quotation mechanism,and increased the reference intensity for a basket of currencies.With the gradual increase of RMB exchange rate fluctuations,how to accurately predict the exchange rate and then formulate effective investment strategies and risk management decisions has become an important issue of extensive concern in the academic and practical circles.Current exchange rate forecasts face challenges.Meese and Rogoff(1983)have pointed out that even the simplest random walk model can defeat complex structural econometric models,and a large number of studies support this conclusion.Various time series models are widely used in exchange rate forecasting,but these models are usually in linear form.The estimation method basically adopts the maximum likelihood method,and assumes that the variables obey the normal distribution.But in reality,the exchange rate does not have linear characteristics and cannot be linear predicted.Kuan and Liu(1995)pointed out that linear models do not have an advantage in exchange rate prediction.A natural question is whether the nonlinear model has more advantages in exchange rate forecasting?In this paper,I first use SRU deep learning network which is newly proposed by ASAPP and MIT in order to predict the exchange rate.In addition to the complete ability to predict nonlinear features,the model also has parallel computing,simple structure,prevention of over-fitting and gradient disappearance.All of these are significantly improving the learning speed of deep networks.The data in this paper is selected from August 11,2015 to August 9,2018,which uses mid-price of USD/CNY,EUR/CNY,JPY/CNY,GBP/CNY and other currency pairs as forecast object.The fixed prediction method is used to compare the SRU model with the same type of LSTM and GRU neural networks,commonly used BP networks,probabilistic neural networks,and traditional econometric models.The results show that the prediction error of SRU is 49% lower than traditional econometric model.The prediction error is 36.6% to 44.7% lower than normal neural network.The prediction error is 29.7% to 35% lower than the same type of deep neural network.While the prediction accuracy is improved,the learning speed is also increased by 18% to 23%.This paper also compares the random walk model.The prediction accuracy of SRU for the other three currency pairs is better than random walk model except USD/CNY,and the prediction advantage of SRU increases with the increase of prediction length.This result is exciting,which shows that deep learning can more effectively discover the distributed features hidden in data by recombining the low-level features to form more abstract high-level features.At present,the trend of cross-disciplinary between finance and artificial intelligence is becoming more and more obvious in the world.The research in this paper shows that there are many factors affecting exchange rate behavior in a developing country with an open economy,which makes exchange rate change more complicated.In order to improve the accuracy of exchange rate prediction under this environment,we must strengthen the non-linear prediction including deep learning.The research and application of this model is a great significance to the improvement of China's economic forecasting ability and forward-looking decision.The follow-up study will further expand the research on the currency and sample size,strengthen the research on optimal network structure of deep learning,and develop an automated trading system based on deep learning to improve the level of financial artificial intelligence in China.