Planning Of Exchange Rate Prediction Schemes Based On News And Neural Networks

Investors can build technical indicators based on technical analysis to predict asset price trends.The process of building technical indicators in technical analysis is the process of extracting features from historical price information.Neural networks can automatically extract features;investors can also analyze the market reaction to macro news to predict asset price trends,at the same time,the neural network language model can learn the probability distribution of text and get the vector representation of news text.This paper uses neural network to connect two kinds of information to build a model to predict the market.This article takes the foreign exchange market as the research object,and selects seven mainstream trading currency pairs with the US dollar as the core in the market.Through analysis,it shows the correlation between the news and currency pairs selected in this article,which proves the feasibility of the scheme.By introducing the framework of machine learning,the prediction problem is transformed into a supervised binary classification problem.This paper studies the features and deep neural network models constructed in this paper to predict the correct rate of the seven currency pairs at different periods.Specifically,this article collects seven major currency pairs of EUR/USD,GBP/USD,AUD/USD,NZD/USD,USD/CAD,USD/CNY,and USD/JPY from January 1,2013 The historical K-line daily data to December 20,2019,crawled the FX168 Finance website's forex news from November 25,2013 to December 20,2019,using past 120 trading day's K-line data as feature,using the document embedding model to vectorize daily news data as feature,and both,to build neural network model to predict the exchange rate of these seven currency pairs in 3,5,10,20 trading days.The research in this paper shows that the neural network that combines the two features has shown significant prediction power in the forecast of fluctuations in all cycles of all currency pairs,and its average out-of-sample prediction accuracy rate is significantly higher than that constructed using one of the features alone.The forecasting scheme,which is characterized by historical price K-line or news alone,only shows a relatively obvious forecasting effect in certain cycles of certain currency pairs.The two schemes are less effective than their combination.At the same time,this article also finds that the price-based forecasting scheme is higher in its forecast of ups and downs after 10 or 20 trading days than its forecast of ups and downs after 3 or 5 trading days,and the news program performs more evenly.This paper also finds that all models have high out-of-sample prediction accuracy rates on the four currency pairs of NZD/USD,EUR / USD,GBP / USD,and USD / JPY,while the performance on the other three currency pairs is poor,indicating the predictability of currency pairs is related to itself,and it is not clear why.Finally,this article analyzes the reasons for the differences in the performance of various models.The neural network of the price K-line scheme is not overfit due to technical conditions.This is the reason for the poor performance of the price model.The inefficient representation of the news text caused by the low dimension of news features may be the reason for the poor performance of the news model.This article proposes some improvements and prospects.