NanoForecast v0.6-HC — DART-Norm + High-Cardinality Ablation
Experimental checkpoint — DART-Norm (causal mean/std normalization) + HC channel expansion (8 channels for electricity/traffic) + 2x upweighting. Single-session training (100 epochs, ~10h on Colab T4).
Ablation Results
| Dataset | v0.5 MASE | v0.6-HC MASE | Change |
|---|---|---|---|
| ETTh1 | 0.681 | 0.834 | +23% worse |
| ETTh2 | 1.110 | 0.760 | -31% better |
| ETTm1 | 0.287 | 1.192 | +315% worse |
| exchange_rate | 4.317 | 14.058 | +226% worse |
| electricity | 2.029 | 0.421 | -79% better |
| traffic | 1.805 | 2.779 | +54% worse |
| Overall | 1.704 | 3.341 | +96% worse |
Key Findings
- Electricity win is real: 0.421 MASE (79% improvement) — beats TimesFM's 0.923
- ETTh2 improved: 0.760 MASE (31% improvement)
- But exchange_rate/ETTm1 collapsed: DART-Norm too aggressive for some datasets
- Released v0.5 remains official — adoption gate failed (traffic worsened)
Training Details
| Parameter | Value |
|---|---|
| USE_DART_NORM | True |
| HC_MAX_CHANNELS | 8 |
| HC_UPWEIGHT | 2x |
| Epochs | 100 |
| Best epoch | 6 |
| Wall time | ~10h (Colab T4) |
| Loss | MultiTaskLoss |
Next Steps
Ablation notebooks available:
colab_training_v06_dart_only.ipynb— DART-Norm only (no HC)colab_training_v06_hc_only.ipynb— HC upweighting only (no DART)
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
Built by Eulogik — deployable AI for the real world