Application Research Of Deep Learning And Transfer Learning In Predicting Foreign Exchange Rate—Based On LSTM Model

Correctly predicting exchange rates have always been of great significance in relevant economic and financial policies making and for companies to avoid foreign exchange risks.As a dynamic market with non-linear changes,the foreign exchange rate market has very significant nonlinear and historical dependence feature.The neural network model is widely used in foreign exchange forecasting because of its advantages in dealing with nonlinear feature systems and has been proved to have higher accuracy than previous time series prediction methods.However,the traditional neural network model ignores the sequential timing relationship within the sequence.Later,some scholars proposed a recurrent neural network(RNN)with memory persistence function to solve the problem of time series dependence.However,in practice,RNN often has problem with gradient disappearing and model training.In order to solve this problem,a long-term short-term memory unit neural network called LSTM neural network(Long Short-Term Memory)was proposed.This neural network has a unique "gate" control and memory unit that makes it popular in financial time series processing.It has been proven to have higher prediction accuracy than traditional RNN.This paper uses the LSTM model to predict foreign exchange rates for different time spans.On the one hand,based on the minute level exchange rate of six currency pairs data,the LSTM neural network and several commonly used traditional prediction models(BP neural network,support vector regression)is used to predict the short-term exchange rate in different minute level.Various error evaluation indicators shows that the prediction error of the exchange rate prediction of the deep LSTM model is better than the two traditional prediction models.On the other hand,in the prediction of the longer-term span of foreign exchange rate,the prediction error of neural network will increase when time interval increasing.Based on LSTM model,this paper explores using Transfer learning to improve the exchange rate prediction,timetransfer learning and currency pair transfer learning is tested on three daily exchange rate data.Comparing with the LSTM direct prediction model,the empirical results show that: The LSTM model of time granularity transfer learning and currency pair transfer learning has lower prediction error;2.In the prediction of the exchange rate between the Australian dollar and the US dollar against the US dollar,the currency pair migration has a greater improvement for LSTM model than the time granularity migration.In addition,taking the RMB exchange rate with a time span of 3 days as an example and With the EEMD(Ensemble Empirical Mode Decomposition)method,the sequence of the migration LSTM prediction result and the LSTM model prediction result are decomposed and compared with the component sequence of the real sequence decomposition.The empirical results shows the reason why migration learning can effectively improve the performance of LSTM model for medium and long-term exchange rate forecasting: transfer learning can more fully learn the long-term trend characteristics of exchange rate fluctuations during the pre-training of low-time-span exchange rate sequences.Therefore,it is more accurate for transfer learning LSTM mode to predict longer time span exchange rate sequences.