Foundation models for electricity price forecasting and battery arbitrage: Can they replace market-specific forecasting models?
Arkadiusz Lipiecki, Rafał Weron
cs.LG · econ.EM
Quick verdict
Rigorous 5-year, 3-market zero-shot FM test: only TabPFN beats EPF benchmarks statistically, but BESS arbitrage value hinges on risk appetite.
Summary
This study tests whether time series and tabular foundation models (FMs), applied zero-shot, can replace dedicated electricity price forecasting (EPF) models. Nine FM variants from five families — three Chronos-2 versions, Moirai-2, TimesFM-2.5, TabPFN-2, TabPFN-3, TabPFN-TS-3, and Mitra+CP — are benchmarked against two state-of-the-art EPF models (LEAR with conformal prediction and the distributional DDNN-JSU network) on the German, Polish, and Spanish day-ahead markets over a five-year test period (2021-2025). Evaluation spans point accuracy (MAE, RMSE), probabilistic accuracy (CRPS), and economic value in 1 MWh battery arbitrage under quantile-based strategies (alpha from 90% to 50%) and unlimited bids, with one-sided multivariate conditional predictive ability tests. Only the TabPFN family significantly outperforms the DDNN-JSU benchmark on every statistical measure in all three markets — for instance, TabPFN-3 cuts German CRPS from 10.70 to 9.15 (about 14.5%). Chronos-2 variants help only in Poland, while Moirai-2 and TimesFM-2.5 trail. Economically, TabPFN leads under unlimited bids and riskier quantile strategies, but DDNN-JSU is most profitable at low risk tolerance (alpha = 60% and 50%), so statistical accuracy does not translate into economic dominance. Acknowledged limitations: three markets, one battery setup and strategy family, possible test-period contamination for FMs pretrained on empirical data, and no fine-tuning experiments.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Only the TabPFN family (TabPFN-2, TabPFN-3, TabPFN-TS-3), used zero-shot, significantly outperforms both state-of-the-art EPF benchmarks on MAE, RMSE, and CRPS in all three markets over 2021-2025.
- Statistical dominance does not translate into economic dominance: TabPFN variants lead under unlimited-bid and riskier quantile-based BESS strategies, while DDNN-JSU is most profitable at lower risk tolerance (alpha = 60% and 50%).
- General-purpose time series FMs (Chronos-2 variants, Moirai-2, TimesFM-2.5, Mitra+CP) do not systematically outperform market-specific EPF models; Chronos-2 helps only in Poland.
- Foundation models therefore cannot universally replace market-specific EPF models; their value depends on model architecture and the decision problem.
Strengths
Longest test horizon in the FM-EPF literature
Five-year out-of-sample window (2021-2025, 1,826 trading days) across three structurally different markets (Sec. 1.1, Sec. 3), versus the single-year tests that dominate prior foundation-model studies reviewed in Sec. 2.
Statistical plus economic evaluation
Beyond MAE/RMSE/CRPS, forecasts are priced via 1 MWh BESS arbitrage with risk parameter alpha (90%-50%), unlimited bids, and Oracle profit bounds (143,474/123,439/77,558 EUR in DE/PL/ES; Sec. 6.2, Tables 3-5), exposing that the CRPS-best model is not the most profitable at low risk tolerance.
Genuinely strong EPF-specific benchmarks
Formal significance testing
Candid limitation handling
Weaknesses
Unquantified contamination risk for mixed-pretraining FMs
Tuning asymmetry between benchmarks and FMs
Thin uncertainty quantification for economic rankings
Second benchmark's test results not shown
Zero-shot scope excludes the practical middle ground
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a sharply framed question: can pretrained foundation models (FMs), applied zero-shot, replace purpose-built electricity price forecasting (EPF) models? The authors (Lipiecki and Weron; Wroclaw University of Science and Technology / Aarhus University) benchmark nine FM variants from five families — Chronos-2, Chronos-2-synth, Chronos-2-small, Moirai-2, TimesFM-2.5, TabPFN-2, TabPFN-3, TabPFN-TS-3, and Mitra+CP — against two state-of-the-art EPF benchmarks (LEAR with conformal prediction; DDNN-JSU, a distributional neural network with Johnson's SU output) in three European day-ahead markets (Germany/DE-LU, Poland, Spain) over a five-year out-of-sample window (2021-2025; 1,826 test days). Claims are evaluated on three axes: point accuracy (MAE/RMSE), probabilistic accuracy (CRPS), and economic value in 1 MWh battery (BESS) arbitrage. The headline claims — only TabPFN wins statistically everywhere; statistical dominance does not imply economic dominance; FMs cannot universally replace market-specific models — are all directly supported by the reported evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can foundation models replace specialized electricity price forecasting models?
According to this study, not universally. Only the TabPFN family of tabular foundation models, used zero-shot, consistently and significantly beat two state-of-the-art EPF benchmarks on statistical accuracy in Germany, Poland, and Spain over 2021-2025. Other foundation models (Chronos-2, Moirai-2, TimesFM-2.5, Mitra) performed inconsistently, and even TabPFN's statistical edge did not always translate into higher battery-trading profits.
Which foundation model was best for electricity price forecasting?
TabPFN-3 achieved the lowest CRPS in all three markets (for example 9.15 vs 10.70 for the DDNN-JSU benchmark in Germany, a roughly 14.5% improvement) and the lowest MAE/RMSE in Germany and Poland; TabPFN-2 was best on MAE/RMSE in Spain. Among time series foundation models, only Chronos-2 significantly beat the benchmark, and only in Poland.
Does better forecast accuracy mean higher battery trading profits?
Not necessarily. In the paper's 1 MWh battery arbitrage setup, TabPFN models earned the most under unlimited bids and riskier quantile strategies (alpha of 90-70%), but the DDNN-JSU benchmark was more profitable at low risk tolerance (alpha of 60% and 50%) in all markets, because conservative limit orders avoid large losses on settled trades.
What data and markets did the study use, and is it reproducible?
It used public day-ahead prices, load, and renewable generation forecasts from ENTSO-E plus fuel and carbon futures (EUA, TTF, Brent, API2) for the German, Polish, and Spanish day-ahead markets, testing 2021-2025. Data sources are public and model pipelines are described in detail, but no code repository is mentioned, and two models (Moirai-2, TabPFN-3/TS-3) carry commercial-license restrictions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Model and author names appear with inconsistent glyph variants throughout the extraction (e.g., 'TabPFN'/'TabPFN', 'DDNN'/'DDNN', 'LEAR'/'LEAR', 'Chronos'/'Chronos'); these are PDF font artifacts of identical names, not distinct entities, and were normalized when citing.
- The extraction ends naturally with reference [62] and no mid-sentence cut; Table 2's categorical set is fully enumerated ({elu, tanh, relu, softplus, sigmoid}), so no truncation-related content loss was detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.