Table 4.
Model settings
| Algorithms | Sentiment series | Market indicators | Historical USD/CNY rate | Parameters |
|---|---|---|---|---|
| ARIMA | X | X | O | Order = (2,1,2) |
| BPNN | O | O | O | Hidden units: 52; Activation function: ReLU; Learning rate: 0.05} |
| ELM | O | O | O | Hidden units: 77; Activation function: sigmoid} |
| SVR | O | O | O | Kernel function: sigmoid; C: 100.0; γ: 0.1} |
| CNN | O | O | O | Convolutional layer: Conv1D, Filter: 64, Kernal size: 2; Pooling layer: MaxPooling1D, Pool size: 2; Optimizer: Adam; Activate function: ReLU |
| LSTM-single | X | X | O | Hidden units: 32; Learning rate: 0.0001; Epoch: 200 |
| LSTM-market | O | O | O | Hidden units: 60; Learning rate: 0.0001; Epoch: 200 |
| LSTM-sentiment | O | X | O | Hidden units: 50; Learning rate: 0.0001; Epoch: 200 |
| LSTM-all | X | O | O | Hidden units: 64; Learning rate: 0.0001; Epoch:200 |
| MF-LSTM | O | O | O | Hidden units: 50(LSTMC1), 60(LSTMC2), 64(FC layer); Learning rate: 0.0001; Epoch: 200 |
O means that the factor is applied, while X means not