NHITS: Neural Hierarchical Interpolation for Time Series Forecasting
Authors: Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, Artur Dubrawski
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Additionally, we conduct extensive large-scale dataset experiments from the long-horizon forecasting literature, demonstrating the advantages of our method over the state-of-the-art methods, where NHITS provides an average accuracy improvement of almost 20% over the latest Transformer architectures while reducing the computation time by an order of magnitude (50 times). |
| Researcher Affiliation | Collaboration | 1Auton Lab, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA 2Unity Technologies, Labs, Montreal, QC, Canada 3Nixtla, Pittsburgh, PA, USA |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/Nixtla/neuralforecast. |
| Open Datasets | Yes | All large-scale datasets used in our empirical studies are publicly available and have been used in neural forecasting literature, particularly in the context of long-horizon (Lai et al. 2017; Zhou et al. 2019; Li et al. 2019; Wu et al. 2021). |
| Dataset Splits | Yes | Datasets are partitioned into train, validation and test splits. Train split is used to train model parameters, validation split is used to tune hyperparameters, and test split is used to compute metrics reported in Table 1. Appendix C shows partitioning into train, validation and test splits: seventy, ten, and twenty percent of the available observations respectively, with the exception of ETTm2 that uses twenty percent as validation. |
| Hardware Specification | Yes | All our experiments are conducted on a Ge Force RTX 2080 GPU. |
| Software Dependencies | No | The paper mentions 'NHITS is implemented in Py Torch (Paszke et al. 2019)' and 'ADAM optimizer (Kingma and Ba 2014)', but does not provide specific version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | We tune the kernel size for multi-rate sampling from Equation (1) and the number of coefficients from Equation (2), some matching common seasonalities and others exponentially increasing. Additionally, we tune the random seed to escape underperforming local minima. Details are reported in Table A3 in Appendix D. During the hyperparameter optimization phase, we measure MAE on the validation set and use a Bayesian optimization library (HYPEROPT; Bergstra et al. 2011), with 20 iterations. We use the optimal configuration based on the validation loss to make predictions on the test set. We refer to the combination of hyperparameter optimization and test prediction as a run. NHITS is implemented in Py Torch (Paszke et al. 2019) and trained using ADAM optimizer (Kingma and Ba 2014), MAE loss, batch size 256 and initial learning rate of 1e-3, halved three times across the training procedure. |