Stealing profits: Spread-based temporal hierarchy forecasting for day-ahead electricity markets
Arkadiusz Lipiecki, Nikolaos Kourentzes, Rafal Weron
q-fin.ST · cs.LG · econ.EM
Quick verdict
Adds direct spread forecasts to THieF reconciliation for day-ahead power prices; gains look consistent, but the BESS test is stylized and no code is shared.
Summary
The paper extends temporal hierarchy forecasting (THieF) for day-ahead electricity prices by adding direct forecasts of all 276 intraday pairwise price spreads to a reconciliation system that already contains the 24 hourly prices, yielding a 300x24 summing matrix reconciled via minimum-trace (MinT) weighting with diagonal-target shrinkage covariance estimation. Base forecasts come from three architectures — linear ARX, a shallow NARX neural-network ensemble, and the pretrained TabPFN-2 foundation model — applied to German (EPEX-DE) and Spanish (OMIE) prices over a five-year out-of-sample period (2021-2025) with 3-year rolling windows. Spread THieF improves hourly RMSE by 2.9-8.8% and daily-average-price metrics by up to 19.7% over unreconciled forecasts, with multivariate conditional predictive ability tests rejecting equality at the 1% level in every model-market combination; unlike Block THieF, it also improves the already strong TabPFN-2 forecasts. Economically, a stylized 1 MWh battery arbitrage exercise (one daily cycle, eta in {1, 0.95, 0.9}, 25 EUR round-trip cost) shows profit gains of about 1.1-9.7% over Block THieF and up to 10.4% over unreconciled forecasts, with NARX plus Spread THieF most profitable despite TabPFN's statistical lead. Main limitations: the trading setup ignores price impact, intraday markets and multi-cycle scheduling; spread-forecast accuracy is never reported separately; and no code or data archive is provided.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● 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
- Jointly reconciling direct forecasts of all 276 intraday spreads with the 24 hourly prices (Spread THieF) improves forecast accuracy across ARX, NARX, and TabPFN-2 in both Germany and Spain, with skill scores up to 19.7% (dMAE, Germany ARX, Table 1).
- Unlike Block THieF, which slightly degrades the highly accurate TabPFN-2 forecasts in both markets, Spread THieF improves them (e.g., +5.9% RMSE in Germany).
- Multivariate CPA tests reject equal predictive ability in favor of Spread THieF at the 1% level against both Base and Block THieF under MAE and MSE loss for every model and market.
- In a stylized 1 MWh single-cycle BESS arbitrage, Spread THieF raises profits by roughly 1.1-9.7% over Block THieF and up to 10.4% over unreconciled forecasts, and even its worst specification beats the best non-Spread alternative in every market-efficiency combination.
- Statistical accuracy and economic value rank architectures differently: TabPFN is the most accurate model, but NARX with Spread THieF earns the highest profit in all six market-efficiency combinations (Table 2).
Strengths
Unusual and well-motivated hierarchy design
Section 2.2 defines a 300x24 summing matrix pairing 24 hourly prices with all 276 pairwise spreads (Eq. 8-9), introducing a genuinely new information source into MinT reconciliation rather than just another temporal aggregation as in Block THieF (Lipiecki et al., 2026).
Rigorous, long-horizon statistical evidence
Five years out-of-sample (T=1826 days, Table 1) across two markets and three architectures; multivariate CPA tests (Sec. 5.2) reject equal predictive ability in favor of Spread THieF at the 1% level in every model-market-loss combination, including against Block THieF.
Improves even the strongest base model where the predecessor method fails
Decision-centric evaluation with honest null results
Transparency about data revisions and method hygiene
Weaknesses
Stylized trading setup limits external validity of profit claims
Direct spread-forecast quality never evaluated separately
No code or data availability statement
Economic differences are not formally tested
Point forecasts only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes **Spread THieF**: a forecast reconciliation system in which direct forecasts of all 276 intraday pairwise price spreads join the 24 hourly day-ahead prices in a 300x24 linear structure (Sec. 2.2, Eqs. 8-9), combined via minimum-trace (MinT) reconciliation with Schaefer-Strimmer diagonal shrinkage of the error covariance (Sec. 2.3-2.4). Three claims are tested on EPEX-DE (Germany) and OMIE (Spain) over 2021-2025 (T=1826 out-of-sample days, 3-year rolling windows) with three base architectures (ARX, NARX, TabPFN-2): (i) accuracy gains over unreconciled forecasts up to 19.7% (SSdMAE, Germany ARX; hourly measures 1.8-16.4% in Germany, 2.9-10.7% in Spain, Table 1); (ii) gains persist for the strong TabPFN-2 (+5.9% RMSE in Germany) where predecessor Block THieF actually degrades it (-1.1%); (iii) economic value in a stylized 1 MWh single-cycle BESS arbitrage: profit gains up to 10.4% vs Base (Spain, NARX, eta=0.9: 43,261 -> 47,746 EUR, Table 2) and roughly 1.1-9.7% vs Block THieF. All three claims are supported within the stated setting, and the CPA evidence (Sec. 5.2) is unusually uniform: 1% rejections in every model-market-loss combination.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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
What is Spread THieF in electricity price forecasting?
Spread THieF is a forecast reconciliation framework that treats the 24 hourly day-ahead electricity prices and all 276 pairwise intraday price spreads (differences between two hours) as one coherent system. Each series is forecast independently, then all 300 forecasts are combined with minimum-trace (MinT) reconciliation weights so that the final hourly prices and spreads are mutually consistent and share information.
How much does Spread THieF improve forecast accuracy?
Over five years of out-of-sample data (2021-2025) on the German and Spanish day-ahead markets, it reduces hourly RMSE by 2.9-8.8% and improves daily-average-price error metrics by up to 19.7% relative to unreconciled hourly forecasts, depending on the base model (ARX gains most, TabPFN-2 least but still positive). Formal CPA tests favor it at the 1% significance level in every tested model-market combination.
Does better forecast accuracy mean more battery trading profit?
Not necessarily. In the paper's stylized 1 MWh battery arbitrage experiment, the statistically most accurate model (TabPFN-2) was not the most profitable: NARX combined with Spread THieF earned the most in all six market-efficiency combinations. Spread THieF raised profits by about 1.1-9.7% over Block THieF and up to 10.4% over unreconciled forecasts, but this is under a simplified setup (one charge-discharge cycle per day, fixed 25 EUR round-trip cost, no market impact).
Is code or data available for this paper?
The paper text contains no code or data repository statement. The inputs come from public sources (ENTSO-E Transparency Platform, OMIE, Investing.com), but the authors note that ENTSO-E revises historical data, so exact replication would require their specific July 2026 data snapshot and a re-implementation of the 336-series forecasting and reconciliation pipeline.
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.
- Figures 1-5 are images in the extraction; all quantitative statements here come from the embedded text and Tables 1-2, which are present in full.
- Exact CPA test statistics and p-values are not shown in the text, only significance claims at the 1% level.
- Absence of a code/data availability statement is verified only for this extracted version of the manuscript.
- Several cited works are 2026-dated (journal articles and arXiv preprints) that could not be verified externally.
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.