Exchange Rate Forecasting with Fused Financial Exogenous Variables by CAE-SCINet

Exchange rate forecasting is essential to decision- making and optimization in various aspects, including international trade, investment decisions, and risk management within the financial domain. The most of existing methods focus on using historical data to predict future exchange rate, that is just using the temporal relation of time series without considering other exogenous variables such as some market sentiment and economic indicator. Recently, some researchers begin to introduce exogenous variables to exchange rate forecasting, while they neglect extracting the spatial relationships of them, resulting in suboptimal forecasting accuracy. This paper proposes an exchange rate forecasting model CAE-SCINet to capture the spatial nonlinear characteristics of multifaceted financial factors, and predict exchange rate by the captured features. Firstly, the spatial characteristics of multifaceted financial factors are extracted by convolutional autoencoder (CAE), and concatenated with exchange rate data. Then, the dynamic features of the joint time series are captured by sample convolution and interaction network (SCINet) at multiple temporal scales. Finally, the multi-resolution dynamic time features are sequenced, added to the original time series by residual connections, decoded by a fully connected network and utilized for predictions. We collected the daily exchange rate data of RMB against USD and related financial factors from August 11, 2015 to September 26, 2023 as the experimental data. The experimental results demonstrate that the proposed model by fusing multifaceted financial factors is effective for the exchange rate forecasting.