Renewable penetration makes short-term electricity prices increasingly weather-driven, and a growing literature studies how traders convert forecasts into profit — almost entirely on the continuous intraday market and almost entirely in single markets. We complement that work at the foundation of the value stack: the day-ahead auction. We define a resolution-disciplined, ceiling-relative forecast-capture metric — the ratio of a forecast-driven, price-taking battery's day-ahead arbitrage P&L to the perfect-foresight ceiling — and measure it across 43 markets spanning Europe, GB, the Nordics and the US, with hourly and native 15-minute markets each handled on their own resolution. Measured with a standard, reproducible reference forecaster — internally constructed from open methods on public data — a price-taking battery captures a money-weighted 92% of the ceiling on average, with a wide spread we decompose into market archetypes (saturated, forecast-powered, flat-hydro). We then test whether raw archived numerical weather adds capture beyond the transmission operator's own published day-ahead RES forecast, and find a structured, not blanket residual signal: weather is orthogonally informative — adding +0.6 to +2.4 percentage points — only in wind-dominated zones with capture headroom (Latvia, Great Britain, Estonia, Lithuania, southern Sweden), while it is neutral-to-harmful in solar-heavy, high-capture markets where the operator's RES forecast already prices it. A near-saturated wind market (Poland, +0.16pp) isolates headroom, not RES type, as the primary driver. We cross-validate against independent third-party benchmarks on the day-ahead slice (Electricity Maps and Chronos-2 for Germany/Austria; Suena for Poland). Day-ahead is the layer every battery trades first — the measurable foundation beneath the intraday value ladder, and the plain complement to the intraday weather "arms race." In short: across 43 markets we measure how much day-ahead trading value a standard price forecast captures, and we identify the markets where adding weather forecasts still improves on the grid operator's own wind and solar forecasts.
Rising variable-renewable share makes short-term electricity prices increasingly weather-driven: as wind and solar set the residual load, forecast error becomes the structural trading signal (Bunn 2000; Kiesel & Paraschiv 2017; Koch & Hirth 2019). A growing literature studies how a battery converts forecasts into profit — but almost entirely on the continuous intraday market, and almost entirely in a single market.
That literature has mapped the value stack with precision. Semmelmann, Yu, Cary & Bunn (2026) establish a clean ladder on the German market: perfect-foresight day-ahead €1.37m < rolling-intrinsic intraday €1.77m < a forecast-driven intraday backstop €2.06m (40 MWh / 40 MW, 2024). The intraday layers carry the headline money — but day-ahead is the floor of the stack, the layer every asset touches first, and the one this paper measures.
The field has, moreover, raced on intraday weather, not day-ahead capture. Kuppelwieser & Wozabal propose the first weather-based intraday strategy and frame better forecasts → more profit as an "arms race in weather forecasting"; Semmelmann et al. (2026) trade the German intraday market on XGBoost weather-forecast updates. All of it is intraday, order-book, single-market, and keyed to the revision between forecast vintages. Nobody has asked the complementary day-ahead question: once the transmission operator has already published a day-ahead RES forecast that the auction prices in, does raw weather still carry residual signal — and does the answer generalise across markets?
Nor does a cross-market day-ahead forecast-capture benchmark exist. Commercial references (Modo, enspired, Aurora Energy Research, Suena) report single or few markets and rarely separate forecast skill from the ceiling it is measured against; academic day-ahead work focuses on price-forecast accuracy, not on the dispatch value a forecast actually realises, at scale, resolution-matched.
We make five contributions. (i) A resolution-disciplined, ceiling-relative forecast-capture metric. (ii) The first 43-market measurement of it. (iii) A market-archetype decomposition of the capture spread. (iv) The weather-orthogonality test — raw NWP against the operator's own published RES forecast — as a day-ahead residual signal, with a structural map of where it lives and where it does not. (v) Independent cross-validation against three third-party references (Electricity Maps, Chronos-2 and enspired for Germany/Austria; Suena for Poland). We make no model-novelty claim: the measuring instrument is a standard, reproducible forecaster, and the contribution is the benchmark, the findings, and the independent revenue-assurance methodology they rest on — not a proprietary model.
Our work sits at the meeting point of three literatures, and is distinct from each along a clear axis.
Battery revenue benchmarking is the closest neighbour, and almost all of it is intraday and single-market. Semmelmann, Yu, Cary & Bunn (2026) establish the value ladder on DE-LU 2024 from EPEX transaction data with order-book (OrderFusion) forecasts — deep in one market, on the intraday strategy; Miskiw, Ludwig, Semmelmann & Weinhardt (2025) give the open-source rolling-intrinsic base that underpins it. Commercial references (Modo, enspired, Aurora, Suena) report single or few markets and typically full-stack revenue (day-ahead + intraday + ancillary) rather than day-ahead forecast-capture isolated. We are complementary, not competing: we measure the day-ahead floor across 43 markets, fence strictly to the day-ahead slice, and separate forecast skill from ceiling — then cross-validate against enspired (Germany/Austria) and Suena (Poland) on that day-ahead slice only.
Weather in power trading is where our empirical finding is adjacent, and here we are candid about kinship. Kuppelwieser & Wozabal propose the seminal first weather-based intraday strategy and frame the "arms race in weather forecasting" on German order-book data, keyed to forecast error; Semmelmann, Miskiw, Thederan & Weinhardt (2026) trade the German intraday market on XGBoost weather-forecast updates. Our distinctness is a different axis rather than greater depth: day-ahead (not intraday), 43 markets (not one), and a specific mechanism — raw numerical weather tested for orthogonality to the operator's own published day-ahead RES forecast. The overlap is thematic. Stated honestly, our weather result is novel in lane and breadth, not in kind: we cite this cluster as the adjacent prior art and ask the complementary day-ahead / residual question it leaves open.
Electricity price forecasting is the methodological backdrop for our measuring instrument; for the open-benchmark and review literature see Lago et al. (2021) and Weron (2014). We use an internally constructed, standard forecaster in that tradition — built from open methods on public data and deliberately left unspecified in its details. The paper asks what capture a competent, reproducible, standard forecast realises across markets and what adds to it, not whether a new model wins: the novelty is the benchmark and the findings, not the forecaster. Closest to our day-ahead question is the emerging foundation-model strand: Lipiecki & Weron (2026) benchmark general-purpose time-series foundation models — Chronos-2 among them — for electricity-price forecasting and battery arbitrage across Germany, Poland and Spain, asking whether they can replace market-specific models. We share their battery-arbitrage lens and cross-market framing but ask the orthogonal question — not which model captures most, but how much a standard forecast captures across 43 markets and where raw weather still adds to it — and we use one of their models (Chronos-2) as an independent cross-validation anchor (§4.4), not a competitor.
3.1 Forecast-capture metric. For a delivery day, a price-taking battery commits its charge/discharge schedule on the forecast and settles on the realised price. Capture = realised forecast-driven P&L ÷ perfect-foresight-ceiling P&L (the same battery dispatched with hindsight). A naive baseline uses yesterday's price curve (lag-1-day persistence). Capture ∈ [naive, 100%]; it isolates forecast skill at converting price structure into dispatch value, independent of absolute price level or battery size.
3.2 Resolution discipline. Markets that cleared hourly vs natively at 15-minute MTU are measured on their own resolution; the perfect-foresight ceiling is computed at the native resolution so the 15-min vs hourly ceiling difference (≈1.28× in Germany) is never a confound. Markets that switched (Germany Oct 2025; Poland 30 Sep 2025) are handled within their pure-resolution eras.
3.3 Battery & dispatch. The canonical battery is a 4-hour, one-cycle-per-day price-taker (usable energy E = 4 MWh per MW of power, round-trip efficiency η = 0.86), dispatched by the daily top-bottom rule: charge in the forecast-cheapest hours, discharge in the forecast-dearest, and settle the resulting schedule at realised prices. Price-taker, no market impact — a deliberate scope choice: we measure forecast capture, not full-stack optimisation (that is the commercial-optimiser domain, and conflating the two is the scope trap we avoid). The cross-validation case studies (§4.4) are each run at their own benchmark's spec — e.g. the Electricity Maps harness at 1.6 MWh / 1.0 MW / η 0.9025 / 1.5 cycles — so every head-to-head is matched rather than compared across specs.
3.4 The reference forecaster (the measuring instrument, not the product). The benchmark is measured with a deliberately standard, reproducible specification — an internally constructed reference forecaster, built from standard open methods on public data and averaged over a small set of rolling calibration windows (months to years); fundamentals = TSO day-ahead load / wind / solar forecasts (gate-safe, published D-1); no bespoke per-zone tuning. This choice is principled: the paper's question is how much capture a competent, reproducible forecast realises across markets, and what adds to it — so the instrument must be reproducible, and must not leak the production layer. CausalGrids' commercial identity is the independent revenue-performance-assurance reference, and its uniqueness is the measurement framework, the cross-market breadth, and the findings — not a proprietary forecaster. The production assurance service applies proprietary per-market calibration and model selection on top of this reference baseline; that production layer is deliberately not disclosed here. We make, and need, no model-novelty claim: a vendor "our model wins" result would forfeit the independence that makes an assurance benchmark credible.
3.5 Weather arm (the orthogonality test). Gate-safe archived short-lead NWP (open-meteo historical- forecast archive, single consistent model) → derived features (wind-power via cube-law, solar-power, CDD/HDD). A/B: BASE (TSO fundamentals only) vs +FCW (BASE + raw NWP features). Crucially the NWP is added on top of the TSO's own published RES forecast, so a positive Δ measures residual weather signal the auction has not already priced from the TSO forecast — orthogonality, not duplication. Faithfulness gate: the BASE arm must reproduce the board's base capture for the zone (guards against a mis-specified panel). An ERA5-realized oracle arm bounds the ceiling (look-ahead; upper bound only, never claimed as gate-safe).
3.6 Data. 43 zones; sources = ENTSO-E (DA prices + load / wind / solar forecasts), EXAA / APG, Elexon (GB, FX-converted to EUR), US ISOs (CAISO, PJM; USD-basis). Each zone walks a 180-day out-of-sample test on a 365-day rolling train, drawn from up to 1,095 days of history (2023–2026). Because capture is a ratio of committed-to-ceiling P&L, it stays comparable across currency bases: US ceilings are computed on a USD basis and GB is FX-converted, but the ratio is unit-free. Honest-degradation discipline: fundamentals completeness is preflight-checked at the board's exact cache window; silent price-only fallback is banned (see §3.7, pipeline hardening). Full per-feed provenance and vintage — source, access, coverage window, resolution, snapshot date and any transform — is tabulated in Appendix B (audit trail).
3.7 Pipeline hardening (honest-degradation discipline). Fundamentals arrive over a throttled public API, and a naïve pipeline silently imputes missing load / wind / solar and caches the degraded panel — producing a forecast that looks healthy but is price-only. We ban this failure mode by construction. Before any capture run, a preflight asserts that each fundamentals series clears a completeness floor (≥ 98% finite over the board's exact cache window) and aborts loud otherwise; a degraded panel is re-pulled serially and re-validated, never median-imputed in place. The discipline is not cosmetic — §4.1 reports the audit event in which it moved the honest headline by half a percentage point, and it is a transferable methodological point: a cross-market capture benchmark is only as trustworthy as its weakest silently-degraded panel.
4.1 The 43-market capture board. The internally constructed reference forecaster, on a 4h price-taking battery (180-day walk-forward, η = 0.86), captures a money-weighted 92.1% of the perfect-foresight day-ahead ceiling across 43 markets (continental Europe, GB, the Nordics, US ISOs), resolution-matched throughout. The median market captures 92.3% and the equal-weighted mean 88.3% — the latter dragged below the headline only by the five deep-Nordic flat-hydro markets (NO_3–NO_5, SE_1–SE_2), where compressed hydro price structure caps the achievable; we headline the money-weighted figure as the economically meaningful aggregate, each market weighted by the arbitrage € actually at stake (a 99% capture in a large, volatile market counts for more than a 35% capture in a tiny flat one). That the median (92.3%) sits essentially on the money-weighted mean confirms the headline is a property of the typical market, not an artefact of the weighting. The per-market table (capture / naïve / headroom-recovered / ceiling / n_days) is reported in full in the companion data appendix. Integrity: the headline is post-audit — a first completeness pass run at the wrong cache window missed that 18 of the 39 zones then on the board had silently degraded to price-only panels under ENTSO-E throttling (a bare-except fallback that median-imputed the fundamentals and cached the result without complaint). Re-pulled serially and re-captured, the honest headline rose 91.6% → 92.1% (surgical restores — FI +14.1, EE +4.6, DE-LU +2.3 — with the persistence baseline pinned unchanged on every zone as a clean sanity check). A durable preflight now auto-heals any degraded panel before capture, and a completeness audit at the board's exact cache window is a required close-out — a reproducibility discipline we recommend for any cross-market capture study.
Figure 1. Day-ahead capture by market, grouped by archetype. The flat-hydro-floor zones sit far below; the forecast-powered zones (cayenne) straddle the money-weighted mean, where skill converts to euros; the saturated core (steel) rides the ceiling. The spread, not the ~92% mean, is the structural result.
4.2 Archetype structure. The capture spread is not noise; it decomposes into market archetypes governed by headroom (the ceiling-to-naïve gap) × RES share × price volatility. Saturated markets sit near the ceiling with little left for skill (e.g. DE-LU ~99% on the restored panel — persistence already grabs most of the structured diurnal shape, so the forecast's differentiation narrows even as absolute capture is high). Forecast-powered markets carry real headroom where forecast skill pays (the Baltics, GB). Flat-hydro Nordic-north zones have compressed price structure that caps the achievable — and are the archetype most easily mis-diagnosed: Finland read as a 66% "flat-hydro floor" purely because its panel was throttle-degraded, correcting to ~80% once repaired. A structural floor can be a data artefact, so fundamentals must be verified before a zone is written off.
4.3 Weather orthogonality — the headline result. The central test — does raw archived NWP weather, added on top of the TSO's own published day-ahead RES forecast, still buy capture? — returns a structured, not blanket, answer, which is the paper's novel contribution. Across the tested zones (gate-safe short-lead NWP arm; eligibility = Δ ≥ +0.5pp and a faithful BASE arm that reproduces the board's base capture):
| Eligible (Δ ≥ +0.5pp, faithful) | Δpp | Not eligible | Δpp | |
|---|---|---|---|---|
| Latvia (LV) | +2.41 | SE_3 | +0.06 | |
| Great Britain (GB) | +1.99 | Italy — CNOR | −0.09 | |
| Estonia (EE) | +1.60 | Italy — NORD | −0.31 | |
| Lithuania (LT) | +0.68 | Italy — CALA | −0.35 | |
| S. Sweden (SE_4) | +0.58 | Italy — CSUD | −0.82 | |
| Italy — SUD | −0.91 (faithful = False) | |||
| Italy — SARD | −0.94 | |||
| Italy — SICI | −2.67 | |||
| Bulgaria (BG) | −3.24 | |||
| Poland (PL) | +0.16 |
Figure 2. Day-ahead capture uplift from raw NWP weather vs base-case capture, by market. The five eligible wind-dominated zones (cayenne) sit inside the +0.5pp eligible band; the solar-heavy and saturated zones (steel) fall below the zero baseline; Poland — wind-heavy but near-saturated — sits near zero, isolating headroom as the primary driver.
Weather is orthogonally informative only in wind-dominated zones with capture headroom (all five eligibles), and is neutral-to-harmful in solar-heavy, high-capture markets (all seven Italian zones + Bulgaria), where the operator's own RES forecast already prices the weather so raw NWP adds noise, not residual signal. The decisive disambiguation is Poland: wind-heavy but near-saturated (base capture ≈ 96%), it returns just +0.16pp — demonstrating that headroom is the primary driver and RES-type / TSO-forecast-quality the secondary one. On the eligible zones the gate-safe forecast-weather arm reaches or approaches the realised-weather (ERA5) oracle, consistent with the residual-signal reading: the market clears on the forecast weather, not the realisation. The mechanism is the finding — a day-ahead residual-weather signal exists, structurally, where wind dominates and capture has room; it is distinct from, and complementary to, the intraday forecast-error signal the arms-race literature exploits.
4.4 Case studies — independent cross-validation (DA slice only). Germany / Austria vs two frontier references: measured on identical resolution-matched day-ahead, the standard forecast is at statistical parity (point nudging ahead) with both Electricity Maps (on its own harness — 1.6 MWh / 1.0 MW / η 0.9025 / 1.5 cycles, 15-min, low-volatility Q1 2026: 87.59% vs 87.4%) and Chronos-2, a market-leading foundation model whose published day-ahead capture on the neutral public fr4borsa/bess-arbitrage judge (DE-LU, hourly, 1 cycle) is 94.2%; our standard forecast, run through the same judge, sits in the same top-of-field band at 94.6% — we cite Chronos-2's published figure and do not re-run the model; we run our own forecast through the same open judge (independently re-confirmed on a freshly-cloned judge, 10 Oct 2026, repo @ dd186c2). The perfect-foresight ceiling is independently validated by enspired's KPMG-assured day-ahead benchmark (reconciled at 5.0% MAE). These three German figures are not comparable to one another — nor to the board's DE-LU 99.2% (Appendix A; 4h / 1 cycle·day / η 0.86, blended-resolution ceiling): capture % is spec-dependent — battery duration, cycles per day, round-trip efficiency, market resolution and the test window all move it — so each is matched to its own counterpart harness and never read across specs. A longer-duration, single-cycle battery on a blended-resolution ceiling (our board) sits higher than a 1.5-cycle 15-minute harness scored on a low-volatility winter quarter (Electricity Maps); that spread is spec, not skill.
Figure 3. Day-ahead capture vs the two published frontier references, resolution-matched. The cayenne (CausalGrids, independent) and steel (frontier) markers nearly coincide on both rows — statistical parity, the point estimate nudging ahead — with CausalGrids and Chronos-2 sitting at the top of the public field.
Poland vs a full-stack optimiser: on the day-ahead-auction slice (native 15-min, measured at the matched 1.5-cycle spec — distinct from the 4h board, which puts Poland at 96.9%, Appendix A), the forecast-driven dispatch is at or above Suena's published standalone day-ahead figure in both months (Aug €16.2k vs €14.0k /MW·mo; Sep €15.8k vs €15.4k), at ~93% capture — consistent with the German ~92–98%, i.e. the skill travels to a second market. This Poland comparison is a two-month point comparison against a published monthly figure, not the deflation-tested daily win-rate we apply to the Electricity Maps and Chronos-2 head-to-heads, and we report it as such. Every comparison is fenced to the day-ahead slice; no claim is made against any full-stack (DA + intraday + ancillary) revenue figure.
What drives capture variation. The spread in day-ahead capture is not forecaster noise but market structure. Three factors govern it, and they compound: headroom — the gap between the perfect-foresight ceiling and the naïve-persistence floor — sets how much a forecast can add; RES share sets how much of the residual-load (and hence price) structure is weather-driven and therefore forecastable; and price volatility / spread depth sets the euro value of getting the dispatch right. Saturated markets (high base capture, thin ceiling-to-naïve gap) leave little for skill to claim — persistence already captures most of a stable diurnal shape — so even a frontier forecast reads as only marginally ahead. Forecast-powered markets (real headroom, wind-heavy) are where skill converts to euros. This also reframes a naïve "higher capture = better market" reading: a 98% saturated market and an 82% forecast-powered one can reward forecast investment in the opposite order.
Independent corroboration. An independent, public, MIT-licensed battery-arbitrage benchmark (fr4borsa/bess-arbitrage, the day-ahead judge we also use for the Chronos-2 cross-validation in §4.4) reaches the same structural conclusions on its own codebase, without coordination. It reports that fundamentals-based features buy +6–7 percentage points in solar-driven Germany and lose 5–7 points in nuclear-dominated France, where the merit-order curve does not transfer — our headroom-and-archetype split in another lab's words, and a clean replication of the finding that forecast skill converts to euros only where market structure permits. It further documents a measured capture-saturation law — a free persistence floor near 85%, sharply diminishing euro returns from each added layer of forecast skill, and a few points of ceiling permanently out of reach — and notes that several independent 2026 analyses converge on it. That a third party, on a different implementation, recovers both our archetype asymmetry and our diminishing-returns framing is external, reproducible support for the central claim.
The weather-orthogonality reading. The headline result (§4.3) says raw NWP weather carries information the auction has not already priced from the TSO's published day-ahead RES forecast — but only where the structure permits it. The signal is a day-ahead residual, mechanistically distinct from the intraday forecast-error signal the arms-race literature exploits (Kuppelwieser–Wozabal; Semmelmann et al.): there, the edge is the revision between day-ahead and intraday vintages; here, it is the orthogonal content of a raw NWP field relative to a single published RES forecast, measured as dispatch capture. It lives where wind dominates — where hub-height wind carries structure the operator's own wind forecast has not fully captured — and where capture still has room; it vanishes or reverses in solar-heavy, high-base markets where the operator's forecast is already good and raw weather only adds noise. Poland is the clean separator — wind-heavy yet near-saturated, and essentially null — isolating headroom as primary. A desk reading this knows where a better day-ahead weather model is worth buying and where it is wasted motion.
A practical implication — per-market model selection. The archetype structure implies that no single global forecaster is optimal everywhere: the forecast characteristics that create value in a gas-on-margin thermal market are not those that create value in a wind-dominated one, and the weather result (§4.3) is itself an instance of a market-structure-specific enhancement. The operational response is therefore per-market model selection rather than one model applied uniformly. Which approach is deployed in which market — and how each is calibrated — is the production layer and is out of scope here; this paper reports a single, uniform reference forecaster throughout, precisely so the benchmark stays reproducible and separable from that production layer.
What the capture spread is worth, in euros. Because capture is a ratio, its money value scales with the arbitrage at stake. Anchored to an independently assured figure — enspired's KPMG-assured German day-ahead top-bottom value of roughly €210–220 per MW·day in 2026, i.e. on the order of €75–100k per MW·year for a 1 MW / 4-hour battery — one percentage point of day-ahead capture is worth roughly €0.8–1.0k per MW·year. On that basis the residual-weather edge of §4.3 (up to +2.4pp where wind dominates and capture has headroom) is worth about €1.8–2.4k per MW·year, and the archetype spread between a saturated and a forecast-powered market — comparable to the gap between a mid-field and a top-of-field forecaster, ~6pp — about €4.5–6k per MW·year. The euros concentrate in the markets with headroom, not in parity-level fractions of a point: a 0.2pp difference between two already-strong forecasts is only about €150–200 per MW·year. This is the practical reading of the archetype structure — it is where forecast investment, a better day-ahead weather model included, actually pays.
Figure 4. Annual value of a 40 MWh battery on the German 2024 market, by strategy layer (after Semmelmann et al. 2026). Day-ahead perfect foresight (cayenne) is the floor; rolling-intrinsic and forecast-driven intraday (steel) add on top. This paper characterises the floor — broadly, across 43 markets.
Position in the value ladder. Semmelmann, Yu, Cary & Bunn place the day-ahead layer at the floor of the BESS value stack, below rolling-intrinsic and forecast-driven intraday. That is exactly the layer this benchmark characterises — and the point is complementary, not competitive: day-ahead is the measurable, cross-market foundation every asset touches first, and the intraday prize above it is powered by the same price-quantile forecasting capability we measure here. Mapping the floor broadly (43 markets) and mapping the prize deeply (one market, intraday) are two halves of the same picture; the residual-weather structure we document at day-ahead is the unasked complement to the intraday weather arms race.
A standard forecast captures a money-weighted ~92% of the day-ahead arbitrage ceiling on average across 43 markets, with a structured, archetype-driven spread: saturated markets leave little for skill, forecast-powered markets reward it. Raw weather still buys residual edge at the day-ahead stage — but structurally: +0.6 to +2.4pp in wind-dominated markets with headroom, neutral-to-negative in solar-heavy saturated ones, with headroom the primary driver. Day-ahead forecast-capture is thus a measurable, cross-market foundation beneath the intraday value ladder, and the residual-weather map tells a trading desk precisely where a better day-ahead weather read still pays — the complementary, unasked question to the intraday weather arms race.
Formatting note: items marked ✓ carry a DOI / venue verified against Crossref, arXiv or the publisher; items marked ⚠ remain to be confirmed before submission.
fr4borsa/bess-arbitrage day-ahead judge — an MIT-licensed open GitHub benchmark (github.com/fr4borsa/bess-arbitrage) that scores any hourly day-ahead price forecast in euros captured by a perfect-foresight-relative battery LP (HiGHS) on energy-charts.info (Fraunhofer ISE) prices, with a stated ex-ante contract and a bring-your-own-forecast scorer ✓. Foundation-model vs market-specific EPF + battery-arbitrage benchmark context (Germany / Poland / Spain, incl. Chronos-2): Lipiecki, P., Weron, R. (2026), Foundation models for electricity price forecasting and battery arbitrage, arXiv:2609.00089.Scope. This appendix reports only the reproducible reference-forecaster outputs — per-zone capture %, the naïve-persistence floor, the headroom recovered (the forecast's skill share), the perfect-foresight ceiling basis, the test-window length, and the archetype label. Production calibrations and model selection are not part of this benchmark and are not reported here. Every row is the standard reference forecaster on free public data, so any reader can reproduce it.
Source. The standard reference-forecaster 43-market board, resolution-matched, post the fundamentals-completeness audit (§4.1). The headline money-weighted mean across the 43 markets is 92.1% (§4.1).
| Market | Capture % | Naïve % | Headroom rec. % | n_days | Ceiling basis | Archetype |
|---|---|---|---|---|---|---|
| DE_LU | 99.2 | 95.6 | 82 | 179 | 15-min + hourly | saturated |
| ES | 99.2 | 98.2 | 53 | 179 | hourly | saturated |
| NL | 98.3 | 95.2 | 64 | 179 | hourly | saturated |
| PT | 98.0 | 96.9 | 34 | 179 | hourly | saturated |
| CH | 97.3 | 93.2 | 60 | 179 | hourly | saturated |
| FR | 97.3 | 94.7 | 48 | 179 | hourly | saturated |
| BE | 97.0 | 92.6 | 60 | 179 | hourly | saturated |
| PL | 96.9 | 90.3 | 68 | 179 | 15-min + hourly | saturated |
| AT | 96.7 | 93.4 | 50 | 179 | hourly | saturated |
| DK_1 | 96.6 | 93.8 | 46 | 179 | hourly | saturated |
| HU | 96.4 | 94.8 | 32 | 179 | hourly | saturated |
| CZ | 96.4 | 93.4 | 45 | 179 | hourly | saturated |
| RO | 96.3 | 94.1 | 38 | 179 | hourly | saturated |
| HR | 96.2 | 94.7 | 28 | 179 | hourly | saturated |
| DK_2 | 96.1 | 92.7 | 46 | 179 | hourly | saturated |
| US_PJM_WESTHUB | 95.9 | 93.4 | 39 | 179 | hourly (USD) | saturated |
| SK | 95.7 | 93.6 | 33 | 179 | hourly | saturated |
| SI | 95.5 | 93.2 | 34 | 179 | hourly | saturated |
| GR | 94.2 | 91.0 | 35 | 179 | hourly | saturated |
| IT_SUD | 93.9 | 86.4 | 55 | 179 | hourly | forecast-powered |
| IT_CSUD | 93.3 | 85.6 | 53 | 179 | hourly | forecast-powered |
| IT_CALA | 92.3 | 84.8 | 49 | 179 | hourly | forecast-powered |
| IT_SICI | 91.9 | 87.1 | 38 | 179 | hourly | forecast-powered |
| IT_CNOR | 91.9 | 83.0 | 52 | 179 | hourly | forecast-powered |
| SE_4 † | 91.8 | 82.5 | 53 | 179 | hourly | forecast-powered |
| IT_NORD | 91.4 | 81.6 | 53 | 179 | hourly | forecast-powered |
| LT † | 91.2 | 80.5 | 55 | 180 | hourly | forecast-powered |
| IT_SARD | 91.1 | 83.7 | 45 | 179 | hourly | forecast-powered |
| NO_2 | 90.0 | 81.9 | 45 | 179 | hourly | forecast-powered |
| BG | 90.0 | 87.1 | 22 | 179 | hourly | forecast-powered |
| LV † | 89.9 | 81.0 | 47 | 180 | hourly | forecast-powered |
| SE_3 | 89.6 | 76.7 | 55 | 179 | hourly | forecast-powered |
| US_CAISO_SP15 | 85.3 | 74.6 | 42 | 179 | hourly (USD) | forecast-powered |
| NO_1 | 85.1 | 75.4 | 40 | 179 | hourly | forecast-powered |
| GB † | 83.3 | 61.8 | 56 | 177 | hourly (Elexon, FX→EUR) | forecast-powered |
| FI | 81.0 | 56.6 | 56 | 180 | hourly | forecast-powered |
| IE_SEM | 77.9 | 58.8 | 46 | 179 | hourly | forecast-powered |
| EE † | 75.1 | 47.4 | 53 | 180 | hourly | forecast-powered |
| SE_2 | 73.4 | 47.8 | 49 | 179 | hourly | flat-hydro |
| SE_1 | 71.3 | 54.5 | 37 | 179 | hourly | flat-hydro |
| NO_5 | 58.1 | 45.9 | 23 | 179 | hourly | flat-hydro |
| NO_4 | 42.3 | 19.1 | 29 | 179 | hourly | flat-hydro |
| NO_3 | 35.4 | 21.0 | 18 | 179 | hourly | flat-hydro |
Archetypes follow §4.2, assigned from the capture / naïve / headroom-recovered triple (headroom recovered = the share of the ceiling-above-naïve gap the forecast captured): the top 19 markets (capture ≥ 94%) are saturated, the middle 19 forecast-powered, the lowest 5 the flat-hydro floor. † marks the five weather-orthogonality-eligible markets (§4.3) — all in the forecast-powered band, as expected. Money-weighted mean across the 43 = 92.1% (§4.1). US (PJM, CAISO) ceilings are USD-basis; GB via Elexon, FX-converted to EUR; DE-LU and PL span a 15-min/hourly resolution era-split, each measured within its pure-resolution era (§3.2).
Note. Per-zone values above are the standard reference-forecaster harness. Production-layer figures may differ by ±0.3–1.6 pp per the two-tier architecture disclosed in §3.4; see §5 for the by-construction reconciliation vs the Board revenue-assurance memo.
Every number in this paper traces to one of the sources below. The discipline is deliberate: CausalGrids is an independent revenue-performance-assurance reference, and an assurance benchmark is only as credible as its provenance. All feeds are free and public; exact snapshot dates are given at finalisation. Here "vintage" means the data-coverage window and the snapshot as-of date — not the publication year (that is the references).
B.1 Primary data feeds (capture-board inputs). All gate-safe (day-ahead, published D-1); fundamentals preflight-checked to ≥ 98% completeness (§3.7).
| Feed | Series | Access | Coverage window | Native resolution | Snapshot vintage | Transforms |
|---|---|---|---|---|---|---|
| ENTSO-E Transparency | DA auction prices (EU zones) | REST API (private token; never in text) | 2023-01 → 2026-10 | hourly / 15-min (per-zone era) | board cache, Oct 2026 | none |
| ENTSO-E Transparency | DA load / wind / solar forecasts (TSO) | REST API | 2023-01 → 2026-10 | hourly | board cache, Oct 2026 | completeness preflight |
| EXAA / APG | AT DA prices + EXAA early auction | API | 2019 → 2026-10 | hourly → 15-min | Oct 2026 | AT zone + shadow |
| Elexon BMRS | GB DA prices | API | 2023 → 2026-10 | 30-min | Oct 2026 | GBP → EUR (FX) |
| CAISO, PJM | US DA LMP | ISO public portals | 2023 → 2026-10 | hourly | Oct 2026 | USD basis (ratio is unit-free) |
Per zone: a 180-day out-of-sample test on a 365-day rolling train, drawn from up to 1,095 days of the windows above.
B.3 External comparison provenance (all fenced to the day-ahead slice). The figures appear in §4.4; this table is their source and period.
| Benchmark | Market / resolution | Period | Spec matched | Their figure | Ours | Assurance / judge |
|---|---|---|---|---|---|---|
| Electricity Maps (Re-Twin harness) | DE, 15-min | Q1 2026 | 1.6 MWh / 1.0 MW / η 0.9025 / 1.5 cyc | 87.4% | 87.59% | re-twin.energy |
| enspired | DE/AT, DA ceiling | Q2–Q3 2026 | DA top-bottom | — (ceiling anchor) | — | KPMG-Austria-assured; ceiling reconciled at 5.0% MAE |
| Chronos-2 (foundation model) | hourly | 2026 | public fr4borsa/bess-arbitrage judge | 94.2% (their published figure) | 94.6% | neutral public benchmark |
| Suena | PL, DA slice (15-min) | Aug + Sep 2026 | standalone DA (slice only) | Aug €14.0k / Sep €15.4k /MW·mo | Aug €16.2k / Sep €15.8k | suena.energy configurator |
No comparison is made against any full-stack (DA + intraday + ancillary) figure — the fence holds on every row.
B.5 Reproducibility. The reference forecaster plus the public feeds above make the headline board reproducible by any reader; the production calibration and model-selection layer are not disclosed. The references are verified (DOI / venue against Crossref, arXiv or the publisher), including the Chronos-2 cross-validation judge — the public, MIT-licensed fr4borsa/bess-arbitrage GitHub benchmark (github.com/fr4borsa/bess-arbitrage), a perfect-foresight battery-LP scorer on energy-charts.info prices with a bring-your-own-forecast front door. Before submission, stamp the exact snapshot dates into B.1–B.4. Vintage: data October 2026.
An independent, cross-market day-ahead forecast-capture benchmark. Walk forward, don't look ahead.
First complete draft · October 2026 · Ikenga / CausalGrids Lab · DA-slice fenced.