chronos-bolt-base, fine-tuned on hourly electricity demand
amazon/chronos-bolt-base fine-tuned on Monash electricity_hourly (321 hourly series, 2012-2014, aggregated from UCI ElectricityLoadDiagrams20112014).
Fine-tuning: randomly sampled (context 2048 -> target 64) windows,
AdamW lr=1e-05, batch 24, early stopping on a held-out tail
(stopped at step 400; ceiling 3000). Seed 42.
Benchmark (5 rolling-origin windows x 168-step horizon, this training run)
| model | MASE | WQL | coverage@80 |
|---|---|---|---|
seasonal_naive_168 | 1.1130 | 0.1279 | 0.095 |
chronos_bolt_base | 0.9378 | 0.0900 | 0.712 |
chronos_bolt_base_ft | 0.9358 | 0.0896 | 0.701 |
Lower MASE/WQL is better; the full leaderboard (classical baselines, LightGBM, PatchTST) and the evaluation protocol are in the companion notebook.
Fine-tuning matches the zero-shot base on typical weeks; its gain concentrates in the benchmark's hardest, anomalous week (Christmas), where it improves on zero-shot for 58% of the 321 series (sign test p≈0.002) and cuts that week's MASE from 1.317 to 1.282 — useful exactly when forecasts matter most. Note this rests on the single holiday episode present in the test period.
Companion Resources
- GitHub project repository: https://github.com/mskayacioglu/energy-demand-forecasting-foundation-models
- Kaggle EDA notebook: https://www.kaggle.com/code/mskayacioglu/energy-demand-forecasting-eda
- Kaggle benchmark notebook: https://www.kaggle.com/code/mskayacioglu/energy-demand-forecasting-with-foundation-models
Intended Use
This model is intended for probabilistic forecasting of hourly electricity-demand-like univariate time series with strong daily and weekly seasonality. It uses only historical target values, with no calendar, weather, price, or other exogenous covariates.
It is best suited as a research benchmark model or as a starting point for demand forecasting systems that can tolerate validation and calibration on their own data.
Usage
import torch
from chronos import BaseChronosPipeline
pipe = BaseChronosPipeline.from_pretrained("mskayacioglu/chronos-bolt-base-monash-electricity-hourly",
device_map="cuda", dtype=torch.bfloat16)
quantiles, mean = pipe.predict_quantiles(
context=torch.tensor(history), # (n_series, <=2048), most recent last
prediction_length=64,
quantile_levels=[0.1, 0.5, 0.9],
)
Horizons beyond 64 steps: predict in blocks of 64, feeding the median back as context.
Out-of-Scope Use
This model should not be used as-is for safety-critical grid operations, financial decisions, or non-hourly/non-electricity domains without separate validation. For applications requiring guaranteed prediction interval coverage, apply post-hoc calibration such as conformal adjustment.
Limitations
Fine-tuned for hourly demand with strong daily/weekly cycles; expect degraded transfer to other
frequencies or domains (start from amazon/chronos-bolt-base instead). Quantiles are the model's
own; coverage was ~70-71% for nominal 80% on this benchmark — recalibrate (e.g. conformal) if you
need guaranteed coverage.
Data Source
This model was fine-tuned on the electricity_hourly dataset from the Monash Time Series
Forecasting Repository, derived from the UCI ElectricityLoadDiagrams20112014 dataset.
- Dataset archive: https://zenodo.org/records/4656140
- Monash forecasting archive paper: https://arxiv.org/abs/2105.06643
Citation
If you use this model, please cite the Chronos paper and the Monash Time Series Forecasting Archive:
@article{ansari2024chronos,
title={Chronos: Learning the Language of Time Series},
author={Ansari, Abdul Fatir and Stella, Lorenzo and Turkmen, Caner and Zhang, Xiyuan and Mercado, Pedro and Shen, Huibin and Shchur, Oleksandr and Rangapuram, Syama Sundar and Arango, Sebastian Pineda and Kapoor, Shubham and Zschiegner, Jasper and Maddix, Danielle C. and Mahoney, Michael W. and Torkkola, Kari and Wilson, Andrew Gordon and Bohlke-Schneider, Michael and Wang, Yuyang},
journal={Transactions on Machine Learning Research},
year={2024},
url={https://openreview.net/forum?id=gerNCVqqtR}
}
@article{godahewa2021monash,
title={Monash Time Series Forecasting Archive},
author={Godahewa, Rakshitha and Bergmeir, Christoph and Webb, Geoffrey I. and Hyndman, Rob J. and Montero-Manso, Pablo},
journal={arXiv preprint arXiv:2105.06643},
year={2021},
url={https://arxiv.org/abs/2105.06643}
}
References
- Chronos / Chronos-Bolt base model: https://arxiv.org/abs/2403.07815
- Monash Time Series Forecasting Archive: https://arxiv.org/abs/2105.06643
- Dataset archive: https://zenodo.org/records/4656140