Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
SKOLR: Structured Koopman Operator Linear RNN for Time-Series Forecasting
Authors: Yitian Zhang, Liheng Ma, Antonios Valkanas, Boris N. Oreshkin, Mark Coates
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments on various forecasting benchmarks and dynamical systems show that this streamlined, Koopman-theorybased design delivers exceptional performance. Our code is available at: https://github. com/networkslab/SKOLR. 4. Experiments |
| Researcher Affiliation | Collaboration | 1Department of Electrical and Computer Engineering, Mc Gill University, Montreal, Canada 2Mila Quebec Artificial Intelligence Institute, Montreal, Canada 3ILLS International Laboratory on Learning Systems, Montreal, Canada 4Amazon Science. |
| Pseudocode | No | The paper describes the methodology using mathematical equations and descriptive text, but it does not include a distinct block or figure explicitly labeled as "Pseudocode" or "Algorithm". |
| Open Source Code | Yes | Our code is available at: https://github. com/networkslab/SKOLR. |
| Open Datasets | Yes | We evaluate SKOLR on widely-used public benchmark datasets. For long-term forecasting, we use Weather, Traffic, Electricity, ILI and four ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2). We assess short-term performance on M4 dataset (Makridakis et al., 2020). |
| Dataset Splits | Yes | Following the standard pipelines, the dataset is split into training, validation, and test sets with the ratio of 6:2:2 for four ETT datasets and 7:1:2 for the remaining datasets. |
| Hardware Specification | Yes | Figure 5: Model comparison on error and training epoch time on P100 GPU. Table 12: Model Efficiency and Performance Comparison for Different Datasets with T = 96. Parameters (Params) are measured in millions (M), GPU memory (GPU) in Mi B, computation time per epoch in seconds (s) on NVIDIA V100 GPU with batch size 32. |
| Software Dependencies | No | We implement SKOLR in Py Torch, applying instance-normalization and denormalization (Kim et al., 2022) to inputs and predictions respectively. The LRU is trained using the Adam W optimizer with no weight decay applied to the recurrent parameters. |
| Experiment Setup | Yes | Following Koopa (Liu et al., 2023), we set the lookback window length L = 2T for prediction horizon T {48, 96, 144, 192} for all datasets, except ILI, for which we use T {24, 36, 48, 60}. We train using Adam W optimizer (Loshchilov, 2017) with learning rate 1 e 4 and weight decay 5 e 4, using batch size 32 across all datasets. Complete hyperparameter configurations are detailed in Table 7. |