surge-fm-v2 — Chronos-2 fine-tuned for US grid day-ahead load forecasting

Full fine-tune of amazon/chronos-2 on 7 years (2018–2023) of hourly load data across 7 major US balancing authorities — PJM, CAISO, ERCOT, MISO, NYISO, ISO-NE, SPP — with hourly 2-m temperature (NOAA ASOS) and US-calendar features as covariates.

Results on 2025 hold-out (macro-averaged over 7 BAs)

Model	Test MASE	95% CI	vs seasonal-naive-24
seasonal-naive-24 (baseline)	1.044	[1.019, 1.071]	—
XGBoost hourly-binned (Roy '25)	0.901	[0.879, 0.924]	−14%
N-BEATS (Pelekis '23)	0.714	[0.692, 0.738]	−32%
Chronos-Bolt zero-shot	0.688	[0.668, 0.708]	−34%
Chronos-2 zero-shot + covariates	0.567	[0.550, 0.586]	−46%
surge-fm-v2 (this repo)	0.492	[0.477, 0.509]	−53%

PJM specifically: ~1.7 % MAPE day-ahead, matching published ISO-internal accuracy.

See github.com/tylergibbs1/surge for full methodology, benchmark code, and a Next.js playground.

Use

import torch
from chronos import BaseChronosPipeline

pipe = BaseChronosPipeline.from_pretrained(
    "Tylerbry1/surge-fm-v2",
    device_map="cuda" if torch.cuda.is_available() else "cpu",
    torch_dtype=torch.bfloat16,
)

# One-shot probabilistic forecast with covariates. See the Surge repo's
# `src/surge/api/forecaster.py` for the full feature-construction recipe.

Training setup

• Backbone: amazon/chronos-2 (119 M params)
• Mode: full fine-tune, LR 5e-6 (cosine), 3 000 steps, batch 16, bf16, seed 42
• Context: 2 048 hours · horizon: 24 hours
• Loss: quantile regression at the full Chronos-2 quantile grid (21 levels)
• Covariates: 2-m temperature (ASOS), hour-of-day sin/cos, day-of-week sin/cos, is_weekend, is_holiday (US federal)
• Trained in ~5 minutes on a single NVIDIA H100 80 GB

Data splits

Split	Period	Rows per BA
Train	2018-07 → 2023-12	~47 400
Val	2024	8 808
Test (reported above)	2025	8 808

Limitations

• Not bankable. For research and reference use only. Do not use for regulated bidding, financial settlement, or any decision where a legally-attested forecast is required.
• Univariate target. The model forecasts load only; other covariates are passed through but not predicted.
• Weather ideality. Evaluation used ground-truth ASOS temperature as "future covariate" — the upper bound. A real production pipeline with HRRR/GFS forecast temperature will degrade 10–20 % in MAPE terms.
• Not all BAs. Only the 7 largest US BAs in EIA-930. Small BAs and non-RTO regions aren't represented in training.

License

MIT. Base model (Chronos-2) is Apache 2.0.

Cite

@software{surge_fm_v2,
  title  = {surge-fm-v2 — Chronos-2 fine-tuned for US grid load forecasting},
  author = {Surge contributors},
  year   = {2026},
  url    = {https://github.com/tylergibbs1/surge}
}