See the miss. Before the market does.
SkyrisQ studies the day-ahead miss of European solar and wind generation: what a wrong schedule costs, and what, if anything, predicts it. The work started with Hungarian solar, where the market and settlement data are best understood, and now also covers Spain, France, Germany and Austria for the same question, with wind to come as the data reach the same standard. Weather-model disagreement, market prices, settlement data and machine-learning models, analysed in the open on public sources, with the evidence and its limits written down.
plantsby risk
HU-PV-03risk 78 · cloud front ES-PV-07risk 61 · sea-breeze front DE-PV-11risk 44 · North Sea cloud band FR-PV-04risk 28 · front stalling offshore AT-PV-02risk 14 · clear skyECMWF and ICON disagree on cloud timing by ~1h. Cloud cover forecast rising in the west.
Illustrative data. A design sketch of one way results could be presented, with one invented plant per country (Hungary, Spain, Germany, France, Austria) to show how a cross-border view might look. No live scoring exists, and no plant data is processed.
€702 million net cost in nineteen months, from day-ahead solar schedules that missed by about 11.5% of scheduled energy, across five European markets.
Every quarter-hour, each country's solar producers settle the gap between what was expected the day before and what they actually produced, under their own market's settlement rules. I priced that gap for each market using its own ENTSO-E day-ahead solar forecast as the benchmark against its imbalance settlement price (MAVIR for Hungary; REE, Spain; RTE, France; Netztransparenz.de's reBAP, Germany; APCS, Austria) — each mechanism (single-price, dual-price, or conditional between the two) verified against the regulator's or TSO's own rulebook before any number below was computed. This started as a Hungary-only measurement; the chart below sums the same measurement across all five day-ahead markets tested so far — not all of Europe. See the five-country breakdown, including each country on its own, in the research notes.
- €1,271.5 million of adverse cost in the intervals where the miss lost money, summed across all five countries; €701.7 million net after the intervals where it happened to pay.
- A few extreme-price quarter-hours carry a large share: the top 1% of intervals (pooled across all five countries) hold about 28% of the adverse cost. That is why timing the risk matters.
- National totals per country from each market's own day-ahead forecast and actual generation, not any single plant. The five countries are the day-ahead markets whose settlement data has passed this project's own usability check so far — not a representative sample of Europe. Latest months use preliminary prices where final ones are not yet published.
Sources: ENTSO-E Transparency Platform; MAVIR (Hungary); REE (Spain); RTE (France); Netztransparenz.de (Germany); APCS (Austria). Own calculation, Jan 2025 to early Aug 2026, every country's window capped at Hungary's own cutoff.
Do the weather models disagree in a useful way? →
Where ECMWF, GFS and ICON split, the worst misses tend to be worse. The effect is small and sits in the tail: it shrinks from 0.30 to about 0.08 once sky state is held fixed, and depends on the controls used.
Testing whether observed clouds help. →
Weather models predict clouds; satellites see them. I am testing whether European satellite cloud observations improve short-range accuracy. It is kept only if a backtest against measured sunshine shows a gain.
Price goes with the miss. →
In hours with a day-ahead price at or below zero, actual output fell below the schedule in 85% of hours. The association survives sky-state controls but shrinks by 20 to 40% under stricter ones. I have not identified the cause.
Checked against real settlements, and audited. →
Every result is tested against actual generation and imbalance prices from the Hungarian market. An internal statistical audit found no code-level flaws and led me to soften several claims; the notes say where the evidence stops.
A first machine-learning model of the miss. →
A gradient-boosted quantile model, tested on months it never saw, beats a fair baseline at the median and at the high (surplus) tail of the miss. It does not beat it at the low (shortfall) tail, and its confidence intervals are not yet well calibrated. Mixed results are reported as mixed.
Written to be reproduced. →
Definitions, datasets, timeframes and the LaTeX methodology are written down in the research notes, with the tests behind each number. Questions and corrections are welcome.
Does Hungary's pattern hold elsewhere? →
I ran the same low-price-shortfall test on Spain, France, Germany and Austria. Four of five countries share Hungary's sign; only Austria reverses it, and that reversal survives a placebo check. The five countries are the day-ahead markets whose settlement data has passed this project's own usability check so far — not a representative sample of Europe.
Where it stands.
Done: reproduction of a published verification for Hungary, a leak-free day-ahead weather dataset, spread and price-regime tests, a first machine-learning model of the miss, an internal audit, and the same cost model and low-price-shortfall test repeated in Spain, France, Germany and Austria (four of five countries share Hungary's sign; Austria is the outlier). In progress: the machine-learning result for the other four countries, wind, independent weather sources, ensemble collection, a satellite cloud test. Not offered: no product, API or service exists, and no customer or plant data is processed.
Questions, corrections, comparisons.
The research is public and preliminary. If you work with solar scheduling, weather or market data for Hungary, Spain, France, Germany or Austria and can check a number or want to compare methods, send a note.
ben@skyrisq.com →Independent research, done in my own time with public data only; it uses no employer or client data and does not represent the views or activities of any employer. Informational only; not investment, financial or trading advice, and no product or service is offered. Data: ENTSO-E Transparency Platform; MAVIR (Hungary); REE (Spain); RTE (France); Netztransparenz.de (Germany); APCS (Austria); HUPX. Weather forecasts: Open-Meteo.com (CC BY 4.0). Database: Meteorological Database, HungaroMet Nonprofit Zrt. See the terms.