The model against realised GB prices — and the forward wedge no hindcast can show.
The two questions a buyer should ask any revenue model: does it reproduce the prices that actually happened, and does it tell you anything a simple model would not? This page answers both from the committed evidence — every number below renders live on the methodology tab from the same evidence files, and the misses are published alongside the passes. Nothing is recomputed for this page.
The model is re-run for each historical year with the fleet that actually existed and compared against realised GB half-hourly market-index prices (Elexon). The published credibility target is an annual mean within 5%. On the realised-profiles basis — the model given the demand, wind and solar that actually happened — the record is:
| Year | Model mean | Realised mean | Error | Verdict (5% target) |
|---|---|---|---|---|
| 2022 (energy crisis) | £203.48/MWh | £195.48/MWh | 4.1% | PASS |
| 2023 | £89.12/MWh | £93.67/MWh | 4.8% | PASS |
| 2024 | £65.68/MWh | £70.16/MWh | 6.4% | FAIL — published |
| 2025 (held-out) | £79.15/MWh | £80.40/MWh | 1.6% | PASS |
Three of four years land inside 5%, across a regime range few models are even tested on — from a £195/MWh crisis year to a £70/MWh calm one. The 2024 miss is 1.4 points outside the target and is published as a FAIL rather than excused; each error also ships with a bootstrap confidence interval in the evidence, because a point estimate without one is marketing.
2025 is the row to interrogate. It is the pre-registered held-out year: every rule was frozen in dated commits before any 2025 data was fetched, and the result — the best annual mean in the record at 1.6% — was published unmodified, together with the P50/P95 distribution-shape failures the same test exposed (the full treatment is in the held-out note and the pre-registration record). A hindcast you can tune is a demo; a held-out year you froze the rules for is evidence.
On the default synthetic-profiles basis (the calibrated stylised weather the forward product runs on), the same years grade 2.8% / 16.8% / 10.4% / 12.0% — larger errors, published with the same discipline; why synthetic remains the forward default is disclosed on the methodology tab.
At 2022–2025 fleet sizes (roughly 1–6 GW of grid-scale batteries), cannibalisation was small: batteries were not yet large enough to compete away their own spreads. That is precisely why a well-calibrated model can reproduce those years — and it means a naive price-taker model can match this history almost as well. History alone cannot tell the two apart.
The difference appears at the fleet sizes your financing horizon actually lives in. In the committed model runs, prices and the whole fleet's dispatch are solved jointly, and the naive price-taker figure is reported alongside at every step. On the central scenario the price-taker forecast overstates battery revenue by +120% in 2026 and +192% in 2030 (£124,207 vs £56,401/MW·yr and £150,303 vs £51,488/MW·yr — see it yourself, no account needed, in the live demo); across the published scenarios the overstatement at 2030 fleet sizes runs 100–300%+ (the one-pager carries the figure).
Read together, the two halves are the product's claim:
For how to underwrite on these numbers — including exactly how to treat the published distribution-tail failures — see the underwriting note.
Forecasts are projections, not advice. Sources: Elexon BMRS realised market-index prices under the Elexon open-data licence; all model values from the committed, fingerprinted evidence record. Diligence walk-through: hello@compoundingenergy.com.