⚠️ Experimental checkpoint — do not use as a general model
This is a research-only NanoForecast ablation, not a v0.6 release.
Result
It improves electricity and traffic substantially, but regresses ETT and exchange-rate accuracy. It therefore fails the release adoption gate.
| Dataset | v0.5 | v06-HC |
|---|---|---|
| ETTh1 | 0.676 | 1.141 |
| ETTh2 | 1.110 | 1.244 |
| ETTm1 | 0.287 | 1.642 |
| exchange_rate | 4.317 | 4.675 |
| electricity | 2.029 | 0.610 |
| traffic | 1.805 | 0.515 |
| Overall | 1.704 | 1.638 |
Official model
Use eulogik/nanoforecast-v05 for general forecasting.
Experiment
USE_DART_NORM = False- Electricity/traffic channels:
8 - Electricity/traffic upweight:
2x - Epochs:
100 - Best epoch:
12 - Loss:
MultiTaskLoss
Citation
@article{kishore2026nanoforecast,
title={NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization},
author={Kishore, Gautam},
journal={arXiv preprint arXiv:2609.31669},
year={2026}
}
Paper: arxiv.org/abs/2609.31669