Long- and Short-Term Memory Networks in Financial Market Predictions: A Study on USD Fluctuations

Abstract

The financial market is very sensitive to changes in the USD exchange rate. People may better assess the economic condition and steer clear of financial hazards with accurate predictions of the USD rate of exchange. This study proposes a CNN-STLSTM-AM hybrid method for closing USD price forecasting. A convolutional neural network (CNN) uses the input data to extract local characters. One better model suggested in the article, Specialized Tanh Long and Short-term Memory (STLSTM), uses local characters to predict the closing value of the USD exchange rate. The feature weights are distributed through the Attention Mechanism (AM) according to the effect of the localized character on the closing price. To prohibit total discarding or retention, the output range is altered from (0, to 0.2, 0.96). This is achieved by adding the tanh (x) plus 0.2 functions to the input gate. The following neural networks are examined and compared: SVR, CNN, CNN-LSTM, CNN-LSTM-AM, LSTM, GRU-LSTM, and CNN-STLSTM-AM, all of which predict the closing price. Experiments show that CNN-STLSTM to AM is the most accurate prediction model.