20世纪末，伴随着全球化产业链的发展和进出口贸易的持续增长，国家之间的金融联系日趋紧密，各国货币市场相互交融。而外汇作为连结各国经济金融市场的核心枢纽，更是能够深度影响国家经济发展。在这一背景下，对于汇率变动情况的研究就显得尤为重要。汇率数据作为一种金融时间序列数据，具有较强的非线性特征和往期依赖性。基于这样的数据特征，神经网络作为一种擅长于提取非线性关系的算法被引入到汇率预测领域。然而传统的神经网络算法缺乏对时序数据前后影响关系的分析能力，因此并不能很好地应用于此类数据。1990年，循环神经网络（RNN）被正式提出，随后便有学者通过门控机制对RNN的梯度消失和梯度爆炸问题进行了改善，开发出了长短期记忆神经网络（LSTM）。这一网络由于其出色的结构设计，被广泛运用于序列数据领域，如语言文本分析，股票预测等。但是目前较少有学者利用这一方法进行人民币汇率预测。基于此，本文将这一方法拓展至人民币汇率预测领域，并分析了其性能表现。

 本文整理了2005年至今的美元、欧元、日元与人民币汇率进行建模预测，对比了LSTM网络与移动自回归平均-广义异方差模型（ARMA-GARCH）和BP神经网络的性能表现，结果显示LSTM网络在预测精度上优于上述两种算法。但是作为一种时间序列预测模型，其远期预测能力会出现一定下降，因此在使用中需要不断滚动更新，以保证模型可以学习到较新的规律。同时，本文发现LSTM网络在涨跌预测准确性上同样好于其他两个模型。可以对汇率的未来涨跌情况进行较好估计。最后，结合一个简化后的外汇交易模型，本文初步论证了LSTM模型具备一定的套利能力，可以被应用在外汇交易市场上。

At the end of the 20th century, with the development of global industrial chain and the continuous growth of import and export trade, financial connections between countries became increasingly close, and the currency markets between countries were intertwined. As the pivot connecting the economies and finances of various countries, the foreign exchange market can deeply influence the economic development of a country. In this context, the study of exchange rate fluctuations is particularly important. Exchange rate data, as a financial time series data, has strong nonlinear characteristics and past period dependency. Based on such data features, neural networks have been introduced into the field of exchange rate prediction as an algorithm that excels in extracting nonlinear relationships. However, traditional neural network algorithms lack the ability to analyze the relationship between past and current time series data, so they cannot be well applied to such data. In 1990, the Recurrent Neural Network (RNN) was officially proposed. Later, scholars improved the gradient vanishing and exploding problem of RNN through gating mechanisms and developed a Long-short Term Memory neural network (LSTM). Due to its excellent structural design, this network is widely used in the field of sequence data, such as language text analysis, stock prediction and so on. However, there are only few scholars using this method to predict the RMB exchange rate. This article extends this method to the field of RMB exchange rate prediction and analyzes its performance.

This article summarizes the modeling and prediction of the exchange rates of the US dollar, Euro, Japanese yen and RMB from 2005 to present, and compares the performance of LSTM network with the Autoregressive Moving Average-Generalized Autoregressive Conditional Heteroskedasticity model（ARMA-GARCH） and BP neural network. The results show that LSTM network outperforms the above two algorithms in prediction accuracy. However, as a time series prediction model, its long-term prediction ability will significantly decrease, so continuous rolling updates are required in practice to ensure that the model can learn new regulations. At the same time, this article found that the LSTM network is also better than the other two models in terms of rise and fall prediction accuracy. It is possible to make a good estimate of the future fluctuations in exchange rates. Finally, combined with a simplified foreign currency trading model, this article preliminarily demonstrates that the LSTM model has certain arbitrage ability and can be applied in the foreign exchange trading market.