| Model | Number of Units | MSE(validation) | MSE(test) |
| LSTM | 50 | 51.0 | 50.7 |
| GRU | 60 | 52.1 | 52.4 |
| EURNN(Jing et al., 2016) | 128 | 51.8 | 51.9 |
| oRNN | 128 | 46.2 | 46.9 |
| uRNN | 108 | --- | --- |
| GORU (EURNN FFT-style) | 64 | 45.5 | 45.7 |
| GORU (Householder) | 64 | 40.9 | 43.0 |
| GORU (with ReLU) | 64 | 45.8 | 47.4 |
| GORU (with tanh) | 64 | 59.7 | 59.6 |
| GORU (without reset gate) | 64 | 45.9 | 46.9 |
| GORU (without update gate) | 64 | 46.3 | 47.9 |
| Model | Number of Units | MSE(validation) | MSE(test) |
| LSTM | 50 | 51.0 | 50.7 |
| GRU | 60 | 52.1 | 52.4 |
| EURNN(Jing et al., 2016) | 128 | 51.8 | 51.9 |
| oRNN | 128 | 46.2 | 46.9 |
| uRNN | 108 | --- | --- |
| GORU (EURNN FFT-style) | 64 | 45.5 | 45.7 |
| GORU (Householder) | 64 | 40.9 | 43.0 |
| GORU (with ReLU) | 64 | 45.8 | 47.4 |
| GORU (with tanh) | 64 | 59.7 | 59.6 |
| GORU (without reset gate) | 64 | 45.9 | 46.9 |
| GORU (without update gate) | 64 | 46.3 | 47.9 |
Note: The hidden size of each model is set to match the total number of parameters. uRNN failed to converge in this task. GORU significantly outperforms all other RNN models.