Research On The Application Of Foreign Exchange Rate Prediction Based On Deep Learning And Migration Learning

With the development of the global economy,exchanges between countries around the world have become closer and foreign exchange has become an important link in international trade.The analysis of exchange rate movements between international currencies can,to a certain extent,assist in financial policy and can also help multinational companies to avoid financial risks.However,accurate forecasting of foreign exchange rates is difficult because of the non-linear and dynamic nature of the foreign exchange rate market and the historically dependent nature of exchange rate fluctuations.Traditional time series forecasting models often adopt a linear form through the maximum likelihood method,assuming that the time series variables obey a normal distribution,ignoring the backward and forward time series relationships within the series,and the actual data situation does not all have linear characteristics,so forecasting cannot be done through linear characteristics alone.The model proposed in this paper combines deep learning with empirical modal decomposition methods and migration learning methods.In addition to its ability to fit non-linear features to a high degree,the model also has a simple structure that prevents gradient disappearance and overfitting,which can improve the learning ability of deep learning from the root and achieve a higher degree of prediction accuracy.In this paper,we first use a deep learning neural network model to replace the traditional serial prediction model and perform a serial error comparison.The data were selected from AUD/JPY from 2 January 2001 to 5 February 2001,USD/JPY from 2 January 2001 to 31 January 2001,CHF/JPY from 3 January 2000 to 24 February 2000 and GBP/JPY from 2January 2001 to 2 February 2001,and the exchange rates of several currency pairs between opening prices as forecasters.It is shown through experiments that the use of deep learning-based LSTM models in exchange rate series forecasting can effectively improve the accuracy of exchange rate forecasting.Secondly,an Ada-LSTM series forecasting model incorporating migration learning is proposed.By decomposing the model and fitting the parameters to the exchange rate data series before formal training,the model learns the deeper features of the time series and retains the optimal parameters for subsequent forecasting training,so that the model can already know the series before training and further reduce the error of exchange rate forecasting.Experiments show that the model is able to further improve the forecasting performance compared to the neural network LSTM-based model.Again,the combined EMD-Ada-LSTM model was used to forecast short-term exchange rate series in different time contexts and compared with some previous forecasting methods.The final results show that the combined model with the inclusion of an empirical modal decomposition of the time series has lower forecasting errors and better forecasting performance than some previous series forecasting models,with the final four sets of exchange rate data forecast results based on RMSE error analysis The final results for the four sets of exchange rate data were based on RMSE error analysis,with errors of 0.003,0.022,0.096 and 0.026 respectively;and based on MAE error analysis,with errors of 0.002,0.017,0.074 and 0.029 respectively,which can basically be considered as the best forecasting results.