PatchTST Baselines (Standard Protocol)
Channel-independent PatchTST baselines trained on a free Colab T4 with the exact hyperparameters from the PatchTST paper. Used as reference baselines for NanoForecast standard-protocol benchmarking.
What is This?
These are reference baselines for comparing against NanoForecast. Each model was trained independently on a single dataset (channel-independent: each channel is a separate sample).
Training Configuration
| Parameter | Value |
|---|---|
| d_model | 512 |
| e_layers | 3 |
| patch_len | 16 |
| stride | 8 |
| dropout | 0.3 |
| learning rate | 1e-4 |
| patience | 3 |
| lookback | 512 |
| horizon | 48 |
Files
{ds}.pt— trained model weights (PyTorch){ds}.json— per-dataset benchmark resultsstandard_benchmark.json— aggregated MASE/MAE/MSE/sMAPE under the protocol inbenchmark_standard.py
Benchmark Results
| Dataset | MASE | MAE | MSE | sMAPE |
|---|---|---|---|---|
| ETTh1 | 0.781 | 0.574 | 0.342 | 5.74 |
| ETTh2 | 1.467 | 0.907 | 0.684 | 4.39 |
| ETTm1 | 0.488 | 0.299 | 0.210 | 3.59 |
| exchange_rate | 3.861 | 0.020 | 0.001 | 1.19 |
| electricity | 1.347 | 48.718 | 42.173 | 1.41 |
| traffic | 1.379 | 0.011 | 0.000 | 43.23 |
Usage
import torch
model = torch.load("ETTh1.pt")
Citation
If you use these baselines, please cite the original PatchTST paper:
@inproceedings{nie2023time,
title={A Time Series is Worth 64 Words: Long-term Forecasting with Transformers},
author={Nie, Yuqi and Nguyen, Nam H and Sinthong, Phanwadee and Kalagnanam, Jayant},
booktitle={International Conference on Learning Representations},
year={2023}
}
Baselines generated by Eulogik for NanoForecast benchmarking.