PriceFM: An Open-Source Deep Neural Network for Day-Ahead Price Forecasting

Overview

Full model architecture and ablation figure — a, Price, lag and lead features enter separate branches, followed by looped cross-attention, temporal mixing and spatial gating to produce probabilistic price forecasts. b, Pinball loss for the full model and four ablations without multiple branches, looped cross-attention, temporal mixing or spatial gating.

Full European electricity data figure with temporal patterns, regional maps and an example week — European electricity data across time and regions. a, Daily and seasonal variation in price, generation, power flows, load and net load. b, Regional differences in those features. c, An example week showing the relationship between price and power-system features.

Features

PriceFM and PriceFM+ both produce day-ahead price trajectories with uncertainty. PriceFM+ adds cross-border flow information and adapts the fusion of input features through looped cross-attention.

Feature	PriceFM	PriceFM+
Input	Price, renewable generation and load	Price, renewable generation, load and cross-border power flows (imports and exports)
Output	Price trajectory with uncertainty	Price trajectory with uncertainty
Adaptiveness	Tune each graph cutoff per region	Self-adapted through looped cross-attention

Results

Testing performance across 40 European regions and the three feature cases. Scores average over observed price slots and 273 test days across three folds. All values are in EUR/MWh and lower is better. The lowest value in each column is bold, and the second-lowest is underlined. Statistical significance is assessed with the two-sided Diebold–Mariano test on paired daily losses.

Realistic

Model	Pinball loss	Interval score	MAE	RMSE

TSO

Model	Pinball loss	Interval score	MAE	RMSE

Oracle

Model	Pinball loss	Interval score	MAE	RMSE

Forecasts

Choose a test day and feature case. Each panel shows one European region over the delivery day. For regions recorded hourly, the archived value is held within its hour. The black line is the observed price. Colored lines are median forecasts, with shaded 10th–90th percentile ranges. Select any combination of the four generic foundation models, PriceFM+, MLP, LSTM, Transformer and three persistence baselines.

Delivery day (UTC) Feature case

Models

Loading forecast index…

Cite

If you find our work useful or use our forecasts, please cite us.

PriceFM+

@misc{YuPriceFMPlus2026,
  title  = {Training Paradigms and Foundation Models for European Electricity Price Forecasting},
  author = {Runyao Yu and Derek W. Bunn and Rafal Weron and Jalal Kazempour and Pierre Pinson and Takuji Matsumoto and Peter Palensky and Ricardo Jorge Bessa and Qingsong Wen and Jochen L. Cremer},
  year   = {2026},
  note   = {Manuscript in preparation}
}

PriceFM

@misc{YuPriceFM,
  title         = {PriceFM: Foundation Model for Probabilistic Electricity Price Forecasting},
  author        = {Runyao Yu and Chenhui Gu and Jochen Stiasny and Qingsong Wen and Wasim Sarwar Dilov and Lianlian Qi and Jochen L. Cremer},
  year          = {2026},
  eprint        = {2508.04875},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CE},
  url           = {https://arxiv.org/abs/2508.04875}
}

We will keep upgrading PriceFM by adding more features. If you have ideas to improve PriceFM and want to contribute, please contact ryu@london.edu. If you are interested in other market price forecasting and want to know more, please visit our website www.deepprior.com.