🤖 machine learning

Foundation models for electricity price forecasting and battery arbitrage: Can they replace market-specific forecasting models?

While foundation models like TabPFN demonstrate superior statistical accuracy in electricity price forecasting across multiple markets, they do not universally replace market-specific models because their economic value in battery arbitrage depends heavily on the specific decision strategy and risk tolerance.

Original authors: Arkadiusz Lipiecki, Rafał Weron

Published 2026-09-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arkadiusz Lipiecki, Rafał Weron

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Electricity is the invisible lifeblood of modern society, flowing through wires to power everything from streetlights to hospitals. But the price of that electricity is not fixed; it fluctuates wildly, sometimes rising to dizzying heights and occasionally plunging into negative territory. These price swings are driven by a complex mix of factors: how much sun and wind are available to generate power, how much fuel costs, and how much energy people are using at any given moment. Because these prices change so unpredictably, companies that own batteries or manage power grids need to guess the future price with remarkable accuracy. If they can predict a price spike, they can buy cheap electricity, store it in a battery, and sell it later for a profit. If they guess wrong, they lose money. For decades, experts have built specialized computer models designed just for these specific markets, learning the unique rhythms of each country's power grid.

Recently, a new type of artificial intelligence has emerged, promising to do the same job without needing to be taught the specifics of any single market. These are called foundation models. Imagine a student who has read millions of books on every possible subject; when you ask them a question about a specific topic, they can answer immediately without needing a new textbook. Similarly, these foundation models are trained on vast amounts of data from many different fields, hoping they can apply what they have learned to electricity prices instantly, without any extra training. The big question for the energy industry is whether these general-purpose AI tools are good enough to replace the specialized, market-tuned experts that have been the standard for years.

A team of researchers from Poland and Denmark set out to answer this question by putting these new AI models to a rigorous test. They gathered five years of electricity price data from three very different European markets: Germany, Poland, and Spain. These countries were chosen because they have different mixes of power sources; Germany and Spain rely heavily on wind and solar, while Poland still uses a significant amount of coal. The researchers pitted nine different versions of these new foundation models against two of the best existing, specialized models used by professionals. They did not just ask the models to guess the price; they tested how well the models could predict the entire range of possible prices, including the rare but dangerous spikes and drops. Finally, they took the most important step: they simulated a real-world trading strategy. They used the predictions from each model to run a virtual battery storage system, buying and selling electricity over five years to see which model actually made the most money.

The results revealed a surprising split between statistical accuracy and real-world profit. In terms of raw prediction skill, one family of foundation models, known as TabPFN, consistently outperformed the specialized experts across all three countries. These models were able to predict the price with greater precision than the traditional methods, even without being trained on the specific history of those markets. However, the story changed when the researchers looked at the bank account. While the TabPFN models were statistically superior, they did not automatically translate into the highest profits. The outcome depended entirely on how much risk the trader was willing to take. When the trading strategy was aggressive and willing to accept higher risks, the TabPFN models generated the most money. But when the strategy was cautious and risk-averse, the older, specialized model called DDNN-JSU actually made more profit, despite being less accurate in its raw predictions.

This finding suggests that the new foundation models are powerful tools, but they are not a universal replacement for the specialized models that have been refined for decades. The best choice depends on the specific goal. If the goal is to have the most accurate statistical picture of the future, the new foundation models are excellent. But if the goal is to maximize profit in a cautious trading environment, the old, specialized models still hold the advantage. The study concludes that while these new AI tools are a significant step forward and can improve forecasting, they should be viewed as powerful components in a larger system rather than a magic bullet that solves every problem. The future of electricity trading likely involves a blend of these new general-purpose tools and the careful, market-specific expertise that has been built up over years of experience.

Technical Summary: Foundation Models for Electricity Price Forecasting and Battery Arbitrage

Problem Statement
Electricity price forecasting (EPF) is critical for power market decision-making, yet it presents unique challenges: prices exhibit strong seasonality, abrupt spikes, heavy tails, and negative values, while being driven by complex interactions between weather, renewable generation, fuel costs, and market design. While foundation models (FMs) promise accurate, zero-shot forecasting across diverse time series without task-specific retraining, it remains unclear whether they can outperform models specifically designed for the idiosyncrasies of electricity markets. Existing literature often lacks comprehensive comparisons across multiple markets, extended test periods, and, crucially, the translation of statistical accuracy into economic value for trading strategies.

Methodology
The authors conduct a large-scale empirical evaluation comparing nine variants from five foundation model families against two state-of-the-art EPF benchmarks across three structurally distinct European markets: Germany, Poland, and Spain. The evaluation spans a five-year out-of-sample period (2021–2025).

Foundation Models Evaluated:
- Time Series FMs: Chronos-2 (and its synthetic/small variants), Moirai-2, and TimesFM-2.5. These are Transformer-based models (encoder-only or decoder-only) designed for general time series forecasting.
- Tabular FMs: TabPFN-2, TabPFN-3, TabPFN-TS-3, and Mitra (with conformal prediction). These leverage in-context learning on synthetic data to solve regression tasks, with TabPFN-TS-3 specifically adapting time series data into a tabular format.
- Input Handling: Time series FMs utilize multivariate or univariate contexts including historical prices, load forecasts, renewable generation forecasts, and fundamental variables (EUA, gas, oil, coal prices). Tabular models treat the forecasting task as a regression problem using lagged features and exogenous variables.
Benchmark Models:
- LEAR+CP: A Lasso-Estimated AutoRegressive model with conformal prediction, representing a robust, interpretable statistical baseline.
- DDNN-JSU: A Distributional Deep Neural Network outputting parameters for a Johnson's SU distribution, representing a sophisticated, market-specific deep learning baseline.
Evaluation Metrics:
- Statistical: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Continuous Ranked Probability Score (CRPS) for probabilistic accuracy.
- Economic: Profits derived from a 1 MWh Battery Energy Storage System (BESS) arbitrage strategy. Two strategies were tested:
  1. Unlimited-Bid (UB): Market orders executed whenever predicted profit is positive.
  2. Quantile-Based (QB): Limit orders based on specific quantiles, controlled by a risk parameter $\alpha$ (ranging from 90% to 50%), where higher $\alpha$ implies higher risk tolerance.
- Statistical Testing: Multivariate Conditional Predictive Ability (CPA) tests were used to determine if performance differences were statistically significant.

Key Contributions

Comprehensive Zero-Shot Comparison: Unlike prior studies limited to single markets or short periods, this work evaluates FMs across three diverse markets over five years, testing both time series and tabular architectures.
Statistical vs. Economic Decoupling: The study explicitly investigates whether gains in statistical forecast accuracy (MAE, CRPS) translate directly into economic gains in a trading context, a gap often overlooked in FM literature.
Risk-Dependent Performance: The analysis reveals that the "best" model depends heavily on the risk profile of the trading strategy, challenging the notion of a single universal winner.

Results

Statistical Accuracy:
- TabPFN Dominance: The TabPFN family (TabPFN-2, TabPFN-3, TabPFN-TS-3) consistently and significantly outperformed both benchmarks across all three markets and all statistical metrics (MAE, RMSE, CRPS). For instance, in Germany, TabPFN-3 reduced the CRPS from 10.70 (DDNN-JSU) to 9.15.
- Time Series FMs: General-purpose time series FMs (Chronos-2, Moirai-2, TimesFM-2.5) showed mixed results. While Chronos-2 variants performed well in Poland, they generally failed to significantly outperform the specialized DDNN-JSU benchmark in Germany or Spain. Moirai-2 and TimesFM-2.5 were among the weakest performers.
- Conclusion on Replacement: The statistical evidence does not support the claim that FMs can universally replace market-specific models; success is highly architecture-dependent, with tabular FMs (specifically TabPFN) showing superior transferability.
Economic Value (BESS Arbitrage):
- No Direct Translation: Statistical dominance did not guarantee economic dominance.
- Risk Sensitivity:
  - High Risk (Unlimited-Bid & High $\alpha$ ): TabPFN variants (particularly TabPFN-TS-3) generated the highest profits in Germany and Spain, and TabPFN-3 in Poland.
  - Low Risk (Low $\alpha$ ): The DDNN-JSU benchmark became the most profitable model when risk tolerance was lower (e.g., $\alpha = 50\%$ or $60\%$ ). This suggests that while TabPFN models may offer better distributional accuracy, the DDNN's specific calibration or point forecast characteristics were more conducive to conservative trading strategies.
- Market Variability: The relative ranking of models for trading profits varied more strongly across markets than for statistical errors.

Significance and Claims
The paper concludes that while foundation models offer a promising avenue for electricity price forecasting, they cannot universally replace models designed specifically for the EPF domain.

Architecture Matters: The success of FMs is contingent on architecture. Tabular foundation models (TabPFN) demonstrated robust zero-shot performance, whereas general time series FMs struggled to consistently beat specialized benchmarks without fine-tuning or specific adaptations.
Context Matters: The value of a forecast is not intrinsic but depends on the decision problem. A model with superior statistical accuracy (lowest CRPS) may not yield the highest profit if the trading strategy requires different risk characteristics (e.g., conservative limit orders vs. aggressive market orders).
Future Outlook: The authors suggest that FMs should be viewed as promising components within larger forecasting pipelines (e.g., for generating covariates or as part of ensembles) rather than as standalone, universal substitutes for carefully engineered market-specific models. The study highlights the necessity of evaluating forecasts through both statistical and economic lenses to fully understand their utility.

Technical Summary: Foundation Models for Electricity Price Forecasting and Battery Arbitrage

More like this