The aims are to analyze the fluctuation forecast of exchange rate markets, including the Chinese Yuan (CNY), and discuss applying the neural network model in the Value at Risk (VaR) measure. Therefore, six exchange rate markets are selected as the research objects, with the CNY exchange rate market as the main body, to analyze and explain the exchange rate fluctuation risks. Second, based on the overview of the neural network model, the Deep Belief Network (DBN), Multilayer Perceptron (MLP), and Long Short-Term Memory Network (LSTM) are introduced, and a VaR method based on risk measurement is proposed. Finally, based on the number of excess days (Exc) and the Kupiec test, the VaR measure results under different models are analyzed. Results demonstrate that the CNY exchange rate market’s historical data are relatively concentrated, with minor fluctuations, and the overall change is a sharp right shift. Compared with the benchmark model Generalized AutoRegressive Conditional Heteroskedasticity, the three neural network models show excellent risk measurement performance for different exchange rate markets. Based on Exc, the DBN model has the optimal risk forecast performance. In the CNY exchange rate market, the Exc values corresponding to the DBN and LSTM models are small, and the forecast performance is fair. Based on the Kupiec test, in addition to the Great Britain Pound exchange rate market, the three neural network models perform well in measuring the risks of the other five exchange rate markets. Besides, the MLP model has the optimal performance in measuring the CNY risks. Hence, the neural network models have excellent applicability in measuring the risks of exchange rate markets.

目的是分析包括人民币 (CNY) 在内的汇率市场的波动预测，并讨论在风险价值 (VaR) 度量中应用神经网络模型。因此，选取6个汇率市场为研究对象，以人民币汇率市场为主体，对汇率波动风险进行分析解释。其次，在神经网络模型概述的基础上，介绍了深度信念网络（DBN）、多层感知器（MLP）和长短期记忆网络（LSTM），提出了一种基于风险度量的VaR方法。最后，基于超额天数（Exc）和Kupiec检验，分析了不同模型下的VaR测度结果。结果表明，人民币汇率市场的历史数据相对集中，小幅波动，整体变化为急剧右移。与基准模型 Generalized AutoRegressive Conditional Heteroskedasticity 相比，三种神经网络模型对不同汇率市场表现出优异的风险度量性能。基于 Exc 的 DBN 模型具有最优的风险预测性能。在人民币汇率市场，DBN和LSTM模型对应的Exc值较小，预测表现公平。基于Kupiec检验，除了英镑汇率市场外，三个神经网络模型在衡量其他五个汇率市场的风险方面表现良好。此外，MLP 模型在衡量人民币风险方面具有最佳性能。因此，