A CNN-STLSTM-AM model for forecasting USD/RMB exchange rate

Abstract

USD/RMB exchange rate plays a significant role in China's financial market. Effective forecasting of the USD/RMB exchange rate can help people analyze the economic situation and avoid financial risks. This paper puts forward a hybrid model CNN-STLSTM-AM to forecast the closing price of the USD/RMB exchange rate. Convolution neural network (CNN) extracts local characters of the input data. Special tanh long short-term memory (STLSTM), an improved model proposed in the paper, forecasts the closing price of the USD/RMB exchange rate through the local characters. Attention mechanism (AM) distributes feature weight based on the impact of the local characters on the closing price. By introducing tanh (x) + 0.2 function to the input gate, the output range is changed from (0, 1) to (0.2, 0.96) to prevent complete discarding or retention. The closing price forecasted by CNN, support vector regression (SVR), long short-term memory (LSTM), CNN-LSTM, gated recurrent unit - LSTM (GRU-LSTM), CNN-LSTM-AM, and CNN-STLSTM-AM are compared and analyzed. The results of experiments reflect the CNN-STLSTM-AM has the highest forecasting accuracy.