Transparency
We publish our error bars
A forecast is only as good as its record against reality. This page shows how our European power-market model is tested — calibrate on one year, blind-score the next — and what the errors actually are, including the markets where we are weakest.
Latest validation run: July 2026 · re-executed on every calibration update
The protocol: held-out, not in-sample
In-sample fit proves nothing — any model can reproduce the year it was tuned on. Our validation is held-out: the calibration is trained on one year of observed hourly market prices only, then the model blind-scores the following year it has never seen (2024 → 2025 full-year, plus a rolling 2025 → 2026 half-year check). Scoring runs on monthly averages per market — error, bias, correlation and quarterly drift — and the whole run is re-executed automatically on every calibration update, so the published record cannot silently go stale.
Where it stands, market by market
Monthly-average error on the blind 2024 → 2025 test, best-performing major markets:
| Market | Monthly MAPE (held-out) |
|---|---|
| Bulgaria | 10.3% |
| Greece | 10.4% |
| Croatia | 10.8% |
| Great Britain | 11.2% |
| Poland | 12.1% |
| Italy | 12.1% |
| Germany | 14.5% |
MAPE = mean absolute percentage error on monthly average prices. Hydro-dominated and coupled markets run higher — see below, where we publish them rather than average them away.
Case study: fixing the hydro-coupled markets — in public
Hydro-dominated markets that price off their neighbours — Switzerland, Austria, Norway, Sweden — were our weakest group: a domestic-cost model structurally misses import parity and the water value of stored hydro. In July 2026 we rebuilt their price formation around actual coupling partners and hydro opportunity cost. The held-out test improved on every metric in all four markets:
| Market | Monthly MAPE | Annual level error |
|---|---|---|
| Austria | 17.6%13.5% | −17.6%−10.7% |
| Switzerland | 27.4%22.5% | −25.3%−19.1% |
| Norway | 27.4%18.2% | −28.2%−18.6% |
| Sweden | 132.0%90.7% | +49.7%+27.4% |
Held-out 2024 → 2025, before → after the July 2026 coupled-market rebuild. Sweden remains our hardest market — hydro-driven inter-annual swings move monthly prices far more than any fundamental model tracks — so its outlook ships with deliberately wide uncertainty bands, and we say so rather than hide it.
Capture and curtailment: benchmarked against independent data
Merchant capture rates
Our per-technology capture rates are set from observed outturn and cross-checked against netztransparenz.de, Bruegel's renewables value tracker and Ricardo/WSP. 2024 reference points:
| Market | Wind | Solar |
|---|---|---|
| Germany | ≈84% | ≈59% |
| Spain | ≈88% | ≈68% |
| Great Britain | ≈89% | ≈89% |
Capture rate = volume-weighted achieved price ÷ baseload average. Future years decline with build-out (cannibalisation) and are floored conservatively.
Curtailment anchors
Curtailment starts from measured TSO and regulator outturn — never a modelling guess — and each stored figure carries its source and a quality grade (measured / derived / estimate). 2024 anchors include:
| Market | Tech | 2024 | Source |
|---|---|---|---|
| Great Britain | Wind | 13% | NESO |
| Ireland | Wind | 10.1% | EirGrid / SONI |
| Greece | Solar | ≈10% | IPTO / ADMIE |
| Germany | Solar | 1.9% | Bundesnetzagentur |
Weather risk, checked against the literature
The 30-year weather ensemble behind our P90 yields reproduces the inter-annual variability ranges reported in the academic record — solar varying a few percent year-to-year, wind two to three times more. It is why two projects with the same P50 revenue can carry very different downside risk, and why our P90s come from thirty real weather years rather than a rule of thumb.
What we haven't validated yet
An accuracy page that only lists wins is marketing. These are the open items, published until they are closed:
- Nodal (locational) price structure covers 20 of the 28 markets today; the eight markets added in mid-2026 carry zonal structure until their nodal solve lands.
- 2026 actuals are half-year only — full-year 2026 scoring happens when the year closes.
- Curtailment anchors are 2024 outturn carried into the 2025 base year; the 2025 re-sourcing round is scheduled, and each figure keeps its vintage label until replaced.
- Norway’s price regime shifted structurally between the calibration years; its factor is flagged non-stationary and its outlook ships with deliberately wide bands.
- Continental battery-revenue backtests are still building out market by market as imbalance and ancillary price history is ingested; Great Britain is the validated reference market.
What's behind these numbers
The architecture, data sources and scenario dials are documented on the transparency page; subscribers get the full methodology and the per-market validation record inside the platform.