Single-step forecasting on GiftEval (held-out evaluation set)
10.89CRPSChronos-2
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Chronos-2Forecast horizon (h)=12026.05 | 10.89 | 2,886 | 13.38 | 0.83 | 13.38 | 0.192 | 0.192 | 0.192 | 13.3 | 0.077 | 72.4 | |
| ARIMAForecast horizon (h)=12026.05 | 11.94 | 3,327 | 14.86 | 0.805 | 14.86 | 0.201 | 0.201 | 0.201 | 13.8 | 0.083 | 70.8 | |
| Chronos-BoltForecast horizon (h)=12026.05 | 14.94 | 5,409 | 18.91 | 0.934 | 18.91 | 0.24 | 0.24 | 0.24 | 14.6 | 0.096 | 67.8 | |
| NanoTS LoRA AttnInitialization=Pretrained, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 20.1 | 8,818 | 24.68 | 1.607 | 24.68 | 0.243 | 0.243 | 0.243 | 16.8 | 0.107 | 47.8 | |
| NanoTS LoRA Attn+IOInitialization=Pretrained, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 22.54 | 11,220 | 26.11 | 1.702 | 26.11 | 0.27 | 0.27 | 0.27 | 17.4 | 0.124 | 45 | |
| Chronos-T5Forecast horizon (h)=12026.05 | 29.88 | 12,764 | 32.99 | 1.212 | 32.99 | 0.173 | 0.173 | 0.173 | 12.4 | 0.073 | 64.6 | |
| NanoTS Full FinetuneInitialization=Pretrained, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 49.51 | 10,461 | 27.6 | 1.561 | 27.6 | 0.286 | 0.286 | 0.286 | 17.5 | 0.199 | 45.3 | |
| NanoTS LoRA Attn+IOInitialization=Random Init, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 50.68 | 59,864 | 62.9 | 3.438 | 62.9 | 0.599 | 0.599 | 0.599 | 28.1 | 0.22 | 47.6 | |
| NanoTS IO OnlyInitialization=Pretrained, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 59.56 | 84,846 | 66.53 | 3.43 | 66.53 | 0.614 | 0.614 | 0.614 | 30 | 0.25 | 46.9 | |
| NanoTS IO OnlyInitialization=Random Init, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 111.6 | 280,932 | 142.3 | 6.565 | 142.3 | 0.706 | 0.706 | 0.706 | 46.2 | 0.313 | 54 | |
| NanoTS Full FinetuneInitialization=Random Init, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 151.8 | 415,339 | 180.4 | 7.819 | 180.4 | 0.914 | 0.914 | 0.914 | 53.2 | 0.463 | 54.2 | |
| NanoTS LoRA AttnInitialization=Random Init, Base model=Qwen3-0.6B, Training step=128, Forecast horizon (h)=12026.05 | 154.5 | 430,387 | 182.2 | 7.872 | 182.2 | 0.917 | 0.917 | 0.917 | 53.2 | 0.481 | 53.5 | |
| Qwen3 TextForecast horizon (h)=12026.05 | 213.5 | 2.6 | 213.5 | 376,667 | 213.5 | 39,508 | 39,508 | 39,508 | 56.3 | 19,754 | 50.3 |