PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Arkadiusz Lipiecki, Rafał Weron

arXiv:2609.00089v1cs.LGecon.EM

TL;DR

The paper asks whether zero-shot foundation models can replace market-specific electricity-price forecasting models, given the importance of uncertainty for risk-sensitive decisions. It compares nine variants from five foundation-model families with two benchmarks across three European markets and evaluates statistical accuracy plus battery-arbitrage value. TabPFN consistently leads statistically, but economic performance depends on trading risk and favors DDNN under lower risk tolerance.

  • Problem

    The paper addresses whether zero-shot foundation models consistently outperform market-specific EPF models and whether statistical gains translate into greater economic value.

  • Method

    The study compares nine foundation-model variants with two EPF benchmarks across Germany, Poland, and Spain over 2021-2025 using point, probabilistic, and battery-arbitrage evaluations.

  • Results

    Only the TabPFN family achieves superior statistical accuracy across all markets; TabPFN leads riskier trading strategies, while DDNN is most profitable at lower risk tolerance.

  • Takeaways & Limitations

    Foundation models cannot universally replace market-specific EPF models, and their usefulness depends on architecture, exogenous-variable treatment, temporal structure, and the decision problem.

  • Takeaways & Limitations

    The analysis is limited to three markets, one battery specification, and one trading-strategy family, and cannot fully rule out test-period contamination for some pretrained models.

Abstract

from arXiv · show

Foundation models promise accurate forecasts with little or no task-specific training, but whether they can replace models designed specifically for electricity price forecasting remains unclear. We compare nine variants from five foundation model families, evaluated in zero-shot mode, with two state-of-the-art electricity price forecasting benchmarks in Germany, Poland, and Spain over 2021-2025. Their performance is assessed in terms of point and probabilistic forecasting accuracy, as well as economic value in battery energy storage arbitrage. Only the TabPFN models consistently and significantly outperform the benchmarks across all three markets and all statistical measures. However, this statistical dominance does not translate directly into economic dominance: TabPFN performs best under unlimited bids and riskier quantile-based strategies, whereas the Distributional Deep Neural Network benchmark is more profitable when risk tolerance is lower. Thus, foundation models cannot universally replace market-specific models, and their value depends on both model architecture and the decision problem.

1. Introduction

Electricity price forecasting is economically important but difficult because prices are seasonal, spiky, heavy-tailed, and shaped by interacting market fundamentals. The study tests whether zero-shot foundation models can outperform market-specific EPF models across statistical and economic tasks.

  • Forecast accuracy matters because electricity price forecasts guide power-company decisions and even modest improvements can have substantial economic value.
  • Daily and weekly seasonality, abrupt spikes, heavy tails, negative values, and multiple exogenous drivers make electricity prices difficult to forecast.
  • Probabilistic forecasts describe uncertainty that point forecasts omit, which is important when errors are asymmetric, extreme prices are costly, or decisions face risk constraints.
  • Zero-shot foundation models are pretrained broadly and applied without retraining, potentially transferring information across series and reducing manual feature engineering.
  • The study compares nine foundation-model variants with two state-of-the-art EPF benchmarks across Germany, Poland, and Spain during 2021-2025.
  • Only the TabPFN family achieves superior statistical accuracy across all markets, while economic rankings favor TabPFN under riskier strategies and DDNN under lower risk tolerance.

2. Literature review

Prior studies examine foundation models for point or probabilistic forecasting, exogenous information, and hybrid systems, but mostly use short test periods or weaker comparisons. This literature motivates a unified long-horizon evaluation against strong EPF-specific benchmarks.

  • Hornek et al. found zero-shot forecasting feasible, but none of six foundation models significantly outperformed the MSTL benchmark in five European markets during 2024.
  • Ponyuenyong et al. reported up to 37.4% lower MAPE than ARIMA, LSTM, and CNN-LSTM benchmarks in Singapore, although MAPE is discouraged for EPF.
  • Moirai zero-shot forecasts were generally less accurate than DDNN by CRPS, while fine-tuning improved calibration and exogenous-variable gains were limited.
  • PriceFM incorporates load, wind, solar, and transmission topology, but its training through 2024 prevents an extensive five-year test period without contamination.
  • Hybrid systems and temporal reconciliation show that foundation models can complement downstream models and that market structure may matter as much as model complexity.
  • Existing foundation-model studies separately examine point forecasting, probabilistic forecasting, exogenous variables, and hybrid systems, predominantly over short test periods.

3. Datasets

The dataset combines day-ahead prices, load and renewable-generation forecasts, and lagged commodity fundamentals for Germany, Poland, and Spain. It spans 2017-2025, with 2021-2025 reserved for out-of-sample evaluation.

  • The datasets come from ENTSO-E and Investing.com and cover 6 January 2017 to 31 December 2025, including a five-year test period beginning 1 January 2021.
  • Figure 2 includes day-ahead prices, load forecasts, renewable-generation forecasts, and EUA, TTF, Brent, and API2 commodity prices.
  • The study uses Germany, Poland, and Spain to test robustness across markets with different generation mixes, renewable penetration, and price dynamics.
  • Renewable generation aggregates onshore wind, offshore wind for Germany, and solar forecasts, while subhourly ENTSO-E series are aggregated to hourly values.
  • Fundamental variables use the most recent available nearest-to-delivery futures closing prices, corresponding to a two-day lag relative to the target day.
  • Preprocessing handles daylight-saving transitions, duplicates, and missing price, load, and forecast observations using preceding-hour or preceding-day values.

4. Foundation models

The study evaluates Transformer-based time-series and tabular foundation models in zero-shot forecasting. It compares their input structures and outputs with established benchmarks, including probabilistic forecasting capabilities.

  • The foundation models considered are Transformer-based and span both time-series and tabular prediction tasks.
  • Table 1 lists the foundation models and benchmarks, their parameter sizes, pretraining-data categories, output types, and selected licensing restrictions.
  • Zero-shot performance measures forecasts for unseen series without task-specific weight updates, using in-context learning instead.

4.1. Time series models

The time-series foundation models differ in architecture, scale, covariate support, and probabilistic forecasting design. The section covers Chronos-2, Moirai-2, and TimesFM-2.5 variants.

  • Chronos-2: Chronos-2 uses an encoder-only Transformer with time- and group-attention layers, followed by a quantile head producing forecasts at 21 probability levels.The levels include 19 values from 0.05 to 0.95 plus tail levels 0.01 and 0.99.
  • Chronos-2: Chronos-2 is the only foundation model in the study that natively supports multivariate forecasting.
  • Chronos-2: The study evaluates default Chronos-2, Chronos-2-synth pretrained only on synthetic data, and Chronos-2-small with 28 million parameters.The default model has 120 million parameters, while Chronos-2-small is approximately one-quarter its size.
  • Moirai-2: Moirai-2 is an 11.4-million-parameter decoder-only Transformer whose quantile head provides nine deciles and whose autoregressive decoding increases distributional resolution.It does not support covariates, and the study includes it after earlier Moirai variants performed considerably worse in preliminary tests.
  • TimesFM-2.5: TimesFM-2.5 is a roughly 200-million-parameter decoder-only model that directly produces nine deciles and can incorporate covariates through the external XReg residual-regression pipeline.

4.2. Tabular models

The tabular foundation models use synthetic-data pretraining and in-context learning, with TabPFN variants, Mitra, and a time-series-to-tabular pipeline represented in the comparison.

  • TabPFN: TabPFN is a Transformer pretrained solely on synthetic datasets under the Prior-Fitted Network paradigm for tabular classification and regression.Its pretraining samples datasets from a prior distribution over generative processes.
  • TabPFN: PFNs use in-context learning to infer predictor-response relationships from supplied data and generate predictions for unseen observations without task-specific training.The model receives training data together with predictor values for the unknown test target and produces a predictive distribution in one forward pass.
  • TabPFN variants: The study evaluates TabPFN-2, TabPFN-3, and TabPFN-TS-3, which transforms time-series forecasting into a tabular regression problem through feature generation and augmentation.TabPFN-2 has 11 million parameters, while TabPFN-3 has 58 million and processes substantially more observations or features.
  • TabPFN variants: TabPFN regression outputs represent predictive distributions over fixed target-space bins, from which quantile forecasts are derived.
  • Mitra: Mitra extends the PFN framework with structural-causal and tree-based priors, while Mitra+CP postprocesses its point forecasts to obtain probabilistic forecasts.

4.3. Context data

Context construction depends on model type: covariate-capable time-series models receive complete series, whereas tabular models receive explicitly engineered lagged predictors.

  • Time-series context: Chronos-2, TimesFM-2.5, and TabPFN-TS-3 receive relevant exogenous inputs as complete time series without a predefined lag structure.
  • Time-series context: Chronos-2 uses 24-hour price, load, and renewable-generation vectors, with a past context length of D = 1449 and four market variables lagged by two days.The market variables are European carbon allowances, natural gas, crude oil, and coal.
  • Time-series context: TimesFM-2.5 and TabPFN-TS-3 decompose forecasting into 24 univariate tasks, one for each hour of the day.
  • Time-series context: The univariate setup additionally uses one-day-lagged prices for the remaining hours as exogenous variables.
  • Time-series context: Moirai-2 treats electricity prices as one hourly univariate series and generates the following day’s 24 prices as a multi-horizon forecast.This strategy preserves cross-hour information through the common hourly series.
  • Tabular context: TabPFN-2, TabPFN-3, and Mitra+CP use explicit lags of prices, load, renewable generation, market variables, and day-of-week indicators as tabular regressors.TabPFN-2 and TabPFN-3 use 1,449 observations, while Mitra uses 1,085 because 364 observations are reserved for postprocessing.

5. Benchmark models

The benchmarks comprise LEAR, a regularized regression model with averaged forecasts and conformal postprocessing, and DDNN-JSU, which directly models hourly predictive distributions.

  • LEAR with conformal prediction: LEAR averages four Lasso regression forecasts trained with rolling windows of 56, 84, 728, and 1,092 days.The implementation uses two short and two long training windows, with three years available for estimation and the fourth year reserved for postprocessing.
  • LEAR with conformal prediction: LEAR’s regularization parameter is selected by seven-fold cross-validation at each retraining step, and an area hyperbolic sine transformation stabilizes variance.
  • LEAR with conformal prediction: LEAR produces conditional-mean point forecasts and uses a conformal-prediction pipeline to obtain probabilistic forecasts.It is trained on a context set of the 1,085 most recent observations and one-hot encodes the day-of-week variable.
  • DDNN-JSU: DDNN-JSU is a two-hidden-layer feedforward network whose output layer returns Johnson’s SU distribution parameters for each of the following day’s 24 hours.Its hyperparameters are selected annually using the preceding year as the evaluation set and optimized with the Tree-Structured Parzen Estimator.
  • DDNN-JSU: DDNN-JSU uses four independent 500-trial hyperparameter studies, with weights reestimated every 28 days using an expanding window.The initial training window contains 1,085 data points.
  • DDNN-JSU: DDNN-JSU jointly estimates distribution parameters for all 24 hourly prices using the same context data as TabPFN-2 and TabPFN-3.

6. Empirical evidence

Across Germany, Poland, and Spain, TabPFN models deliver the strongest and statistically significant forecast accuracy, but economic rankings depend on trading risk, market, and the decision strategy. Statistical superiority therefore does not translate directly into higher battery-arbitrage profits.

  • Forecast accuracy: 14.5%, 12.3%, and 12.4% are TabPFN-3’s relative CRPS reductions versus DDNN-JSU in Germany, Poland, and Spain, respectively.The corresponding CRPS reductions are from 10.70 to 9.15, 10.96 to 9.61, and 9.30 to 8.15.
  • Forecast accuracy: TabPFN models achieve the strongest statistical performance across all three markets and significantly outperform DDNN-JSU in daily MAE, RMSE, and CRPS at the 1% level.The same conclusion holds when LEAR+CP is used as the benchmark.
  • Forecast accuracy: Chronos-2 variants are competitive mainly in Poland, while Moirai-2, TimesFM-2.5, and other foundation models do not consistently outperform EPF-specific benchmarks.Chronos-2 and Chronos-2-synth significantly outperform DDNN-JSU across all statistical measures in Poland, but not in Germany or Spain.
  • BESS arbitrage: Risk appetite changes the economic ranking: TabPFN variants tend to lead under riskier quantile-based strategies, whereas DDNN-JSU becomes more competitive as strategies become more risk-averse.In Germany and Poland, TabPFN-TS-3 leads for α = 90–70%, while DDNN-JSU leads at α = 60% and 50%; in Spain, DDNN-JSU leads except at α = 90%.
  • BESS arbitrage: TabPFN-TS-3 obtains the highest unlimited-bid profits in Germany and Spain, while TabPFN-3 leads in Poland, although differences between leading models are small.In Poland, profits are 106,685 EUR for TabPFN-3 and 106,060 EUR for TabPFN-TS-3.
  • BESS arbitrage: Trading profits reflect both settlement frequency and conditional profit distributions, so the model with the best statistical errors does not always maximize arbitrage returns.Quantile-based strategies are more selective and generally produce higher mean and median profits per settled transaction, while model-based strategies remain below Oracle profits.

7. Conclusions

Zero-shot foundation models can outperform strong market-specific benchmarks, but gains are concentrated in TabPFN and do not extend across architectures. Economic value varies with market and trading configuration, so foundation models cannot universally replace carefully designed EPF models.

  • Evaluation: Nine foundation-model variants were evaluated against two state-of-the-art benchmarks from 2021-2025 in Germany, Poland, and Spain.Evaluation covered MAE, RMSE, CRPS, and profits from quantile-based and unlimited-bid battery trading strategies.
  • Economic evaluation: TabPFN variants tended to perform best under unlimited bids and riskier quantile strategies, whereas DDNN-JSU performed best under risk-averse strategies.Quantile-based strategies generally traded less frequently but earned higher profits per settled transaction, so lower statistical errors did not necessarily imply higher cumulative profits.
  • Implications and limitations: Foundation-model usefulness depends on architecture, exogenous-variable treatment, temporal representation, and the decision problem.The study is limited to three markets, one battery specification, and one family of trading strategies, with possible test-period contamination for some empirically pretrained models.
Loading 2609.00089v1…