Design of Experiments vs. TOPSIS to Select Hyperparameters of Neural Attention Models in Time Series Prediction

Attention models are used in neural machine translation to overcome the challenges of classical encoderdecoder models. In the present research, design of experiments and TOPSIS methods are used to select hyperparameters of a neural attention model for time series prediction. The configurations selected by both methods are compared with outof-sample data in time interval between January 2020 and April 2020 when global economies were significantly impacted due to Covid-19 pandemic. Results demonstrated that both selection methods outperformed each other in terms of different output features. On the other hand, our results with more than 95 % coefficient of determination and less than 0.23 % MAPE verified that neural attention models had strong capabilities in exchange rate prediction even in extraordinary situations in global economies.