Short-term electricity prices can change quickly, and many parts of the energy system rely on accurate price forecasts to plan operations, trade power and manage risk. Foundation model–based probabilistic price forecasting uses a pre-trained time-series model to generate short-term wholesale electricity price distributions. Instead of training a bespoke forecasting model for each market, the method adapts a large time-series model that has already learned general temporal patterns from diverse datasets.  

The forecasting workflow involves two distinct steps. First, the model is applied to reconstruct missing values in historical price series, producing a consistent input dataset. Second, the same model generates probabilistic forecasts that describe a range of plausible future prices rather than a single point estimate. Because the model relies on representations learned during pre-training, it can remain robust during periods of elevated volatility, when price behavior deviates from historical norms. This reduces the need for extensive feature engineering and hyperparameter tuning while supporting faster deployment across markets.[1,2] 

Main Outcomes of the Solution 

• Reliability & Flexibility: Supports consistent short-term price forecasting during volatile market conditions by leveraging generalized temporal patterns learned during large-scale pre-training, reducing sensitivity to regime shifts. 
• Business Value: Shortens model development and deployment timelines for price forecasting, enabling trading, hedging and operational planning without requiring large, highly specialized modelling teams. 

Key Deployment and Scaling Factors 

1. Ensure clean and complete wholesale price datasets aligned with official market index definitions. 
2. Apply robust missing-data reconstruction methods to produce a consistent historical price series. 
3. Use systematic back-testing approaches, such as daily walk-forward evaluation, to assess performance under changing market conditions. 
4. Design deployment architectures that allow extension to additional markets through configuration changes and limited fine-tuning. 
5. Provision appropriate compute resources to support efficient inference, with stable numerical precision suitable for probabilistic outputs. 

Case Study

Accenture (2025), United Kingdom

For anonymous client. The forecasting system was evaluated using Great Britain’s Market Index Data (MID), a half-hourly wholesale electricity price series published under the Market Index Definition Statement. The dataset covered the period from late 2016 through October 2025 and included price data only, without supplementary inputs such as demand, weather or fuel prices. 

The foundation model was adapted to this dataset and assessed using a daily walk-forward evaluation framework. Performance remained stable across the full sample, including the highly volatile market conditions observed during 2021–2022. The implementation was operationally lightweight: training, fine tuning and full-day inference were executed on a single consumer-grade GPU, enabling integration into trading, hedging and operational planning workflows without significant infrastructure requirements. 

While further development is underway to incorporate additional explanatory variables and deeper fine-tuning, the price-only configuration provides a transparent and reproducible baseline for probabilistic short-term electricity price forecasting.[3]

References:

1. API Documentation I Exelon
2. Probabilistic Price Forecasting for Day-Ahead and Intraday Markets: Beyond the Statistical Model I Sustainability Vol. 9 Issue 11 (October 2017)
3. Information provided to World Economic Forum (November 2025)