CausalGrids Lab CausalGrids Lab Walk forward, don't look ahead.

CausalGrids Lab · Power-Trading Intelligence · Benchmark Analysis · v2 · Two-Frontier

Day-Ahead BESS Forecast-Capture — Benchmarked Against Both Frontier Models

CausalGrids measured on the identical German day-ahead market, period and battery against the two published frontier references anyone can check — Electricity Maps (a forecast competitor, 15-minute) and Chronos-2 (the market-leading foundation model, on a neutral open judge, hourly) — plus enspired, a KPMG-assured optimiser whose benchmark independently validates our ceiling. On both frontiers the honest answer is statistical parity — which we read as a sign that our independent revenue-assurance benchmark is frontier-grade, scored to an auditor's deflation standard and independent of every vendor it grades.

German day-ahead (plus Polish DA, §6b) · January–September 2026 · Prepared by CausalGrids · v2 · validated · 6 October 2026 · Confidential

Status. This is the validated (v2) release. The exact forecasters, windows and judges described here were re-run at scale on the Universidad Politécnica de Madrid (UPM) high-performance cluster; the figures below are the HPC-validated set. Every number is gate-safe walk-forward — the forecaster at time t sees only information available at t, with no look-ahead anywhere — so the HPC run confirmed, rather than revised, the ex-ante results.

1Executive summary

On the German day-ahead market a battery that dispatches on the CausalGrids forecast captures ~92–98% of the perfect-foresight ceiling in every regime we tested. Measured the same way against the two frontier references anyone can reproduce, the honest answer on both is statistical parity — reached with an independent, frontier-grade benchmark whose numbers are published deflated for overfitting.

The two frontiers · German day-ahead · matched resolution on each
Frontier referenceResolution / judgeReferenceCausalGridsRead
Electricity Maps (forecast competitor)15-min · Re-Twin harness87.4%87.59%Level · nudging ahead
Chronos-2 (market-leading foundation model)hourly · neutral fr4borsa judge94.4%94.6%Level · nudging ahead

What we lead on. Both frontier references are forecasting tools a desk uses to trade: Electricity Maps is a subscription feed; Chronos-2 is Amazon's open-weight (Apache-2.0) foundation model. CausalGrids sits on the other side of the table — the independent revenue-assurance reference that measures how much of the available revenue a battery's strategy actually captured, scored to an auditor's standard against the persistence floor and the perfect-foresight ceiling. Reaching parity with the frontier on its own ground is what makes that yardstick credible: a frontier-grade benchmark that answers to none of the vendors it grades. And we publish our capture figures deflated (Deflated Sharpe ≈ 1.000, Probability of Backtest Overfitting 20.6%) — a discipline neither frontier publishes. On capture we are level with the frontier; on independence, transparency and statistical rigor — the things a revenue-assurance read must have — the advantage is ours.

The three capture claims we stand behind, in order of strength

  1. Capture is regime-robust — and rises across the year. Native 15-minute capture rises 87.2% (Q1) → 91.5% (Q2) → 91.8% (Q3) — a 9-month average of 90.66% (Jan–Sep 2026, on an enspired-validated ceiling); hourly capture rises 90.7% → 96.4% across the regimes (neutral judge). The EM-benchmark quarter (Q1) is our weakest — the full-year picture is materially stronger.
  2. We are level with both frontier models on their matched benchmarks — point nudging ahead on each. EM 15-minute: 87.59% vs 87.4% (level, point nudging ahead). Chronos-2 hourly, neutral judge: 94.6% vs 94.4% (level, point nudging ahead) — the only vendor-independent benchmark in the top tier.
  3. Hourly forecast skill is regime-robust. Gap-closure over persistence holds at 64.5% → 65.4% across regimes (neutral judge), while absolute hourly capture rises 90.7% → 96.4%. Skill is not borrowed from an easy quarter.
The honest caveats we volunteer, up front (§5, §8). (i) At 15-minute resolution the relative edge over persistence narrows in a highly persistent summer — gap-closure 51% → 34–36% across Q1→Q3 — a genuine market property we measure rather than average away (even as absolute capture rises to 91.8%). (ii) Capture in low-spread winter (January–February) is structurally lower (~78–85%) than in high-spread months (March ~89%), because thin intra-day spread compresses the capture ratio — and these are the lowest-revenue months, so the € impact is modest. §8 shows exactly where this sits and how v2 closes it.

For an asset owner — your three questions, in euros. Beneath the benchmark language, this report answers what a BESS owner, investment committee or lender actually asks.

  1. How much day-ahead value could this asset capture? On the German day-ahead auction a well-forecast 1 MW / 4-hour battery captures ~92–98% of the perfect-foresight ceiling across the regimes tested. In money, that ceiling — anchored to enspired's KPMG-assured day-ahead top-bottom value of €210–220 per MW·day in 2026, i.e. ~€75–100k per MW·year — translates to roughly €70–98k per MW·year of realised day-ahead arbitrage.
  2. How does your contracted BESS optimiser perform? That is precisely what CausalGrids measures — independently, against the same persistence floor and perfect-foresight ceiling, and deflated for overfitting. The gap between your optimiser's capture and the frontier is the leakage: an optimiser running ~6 points below the top of the field is leaving ~€4.5–6k per MW·year on the table, every year, compounding over the asset's life.
  3. What improvement is credible, in euros? Each percentage point of day-ahead capture is worth ~€0.8–1.0k per MW·year. So a better day-ahead weather read in a wind-headroom market (up to +2.4pp — see the companion 43-market paper) is worth ~€1.8–2.4k per MW·year; closing a 6-point optimiser gap is ~€4.5–6k per MW·year. Parity-level differences are small — a 0.2-point edge between two already-strong forecasts is only ~€150–200 per MW·year — which is the honest point: the euros are in closing real leakage and in the right markets, not in chasing fractions of a point.

2The standard of proof (discipline is a differentiator)

Before any number, the discipline that produced it — the evidence hierarchy an auditor would apply to someone else's results. Neither frontier vendor publishes capture to it.

3Frontier 1 — Electricity Maps (forecast competitor · 15-minute)

The setup. Electricity Maps publishes a day-ahead price forecast whose battery-capture is measured on the Re-Twin harness — the cleanest available third-party point measured the way we measure ourselves. We put the CausalGrids forecast through the identical market, period and battery (E = 1.6 MWh, P = 1.0 MW, η_RT = 0.9025, 1.5 cycles/day), Jan–Mar 2026.

German day-ahead · January–March 2026 · 15-minute, matched resolution
MetricCausalGridsElectricity MapsRe-Twin
Forecast-driven capture (15-min, matched)87.59%¹87.4%83.6%
Forecast MAE vs realized (€/MWh) — direct quality< 17.217.221.4
Naive-persistence MAE (€/MWh) — same-market check27.8527.9—

The read. On the resolution-matched 15-minute benchmark — the only comparison our resolution rule admits — CausalGrids is statistically level with Electricity Maps, within residual cross-harness uncertainty, and both clear Re-Twin. The two harnesses price the same market: our naive-persistence MAE of 27.85 €/MWh matches the published 27.9 almost exactly. We do not quote a cross-provider "edge over persistence" — the naive floor is harness-specific. Forecast quality is carried on the direct forecast-MAE axis and the neutral public judge (§4).

Banked and deflation-clean. The production ensemble captures 87.0%, rising to 87.59% with the gate-safe winter-shape overlay — both statistically indistinguishable from Electricity Maps' 87.4% over the Jan–March 2026 window (the ~0.2-point margin sits inside the paired-bootstrap interval; a single low-volatility quarter cannot resolve it); both clear Re-Twin (83.6%), deflation-clean (DSR / PBO pass).

¹ Money lens. Our 15-minute capture rises with the daily arbitrage pie — equal-weighted ~83%, €-weighted 87% (71% on the thinnest-spread quintile to 91% on the fattest). The headline reflects performance on the high-value days, and because EM's 87.4% is likewise a pooled €-weighted figure, the head-to-head is money-vs-money on both sides.

Tier 0 Electricity Maps × Re-Twin case study (re-twin.energy/blogs/impact-of-day-ahead-forecast, 07/09/2026; re-verified 25/09/2026): perfect foresight €13,914 (100%), previous-day naive €10,904 (78.4%, MAE 27.9), Re-Twin €11,628 (83.6%, MAE 21.4), Electricity Maps €12,163 (87.4%, MAE 17.0–17.2) — same 10 MW / 0.9025 round-trip / 2h / 1.5-cycle battery. Our harness prices a larger pie (ceiling ≈1.28× theirs), so absolute € is never compared cross-harness — only within-harness capture fractions and direct forecast MAE.

4Frontier 2 — Chronos-2 (market-leading foundation model · hourly · neutral judge)

Why this is the harder, more meaningful frontier. Chronos-2 is the strongest publicly-benchmarked price-forecasting foundation model — open-weight (Apache-2.0), pretrained. Matching it is the ambition; doing so independently, as the benchmark rather than a rival feed, is what matters. We measure on fr4borsa/bess-arbitrage — an independent, reproducible public scorer that grades any forecast CSV on a fixed HiGHS LP against a published leaderboard. DE-LU, hourly, H1 2026.

Neutral fr4borsa judge · DE-LU hourly · H1 2026 · capture of perfect foresight
ForecasterCaptureIndependent?
CausalGrids94.6%¹✓ independent
Chronos-2 (open-weight foundation model)94.4%¹✗ forecast model
isotonic ex-ante fundamentals91.7%—
Chronos-Bolt (open)90.7%—
PriceFM90.4%—
learned-linear86.2%—

¹ Head-to-head on the common apples-to-apples window (2026-01-08..06-29, 4,127 settled hours, identical perfect-foresight pie both sides — Chronos-2's series starts 08 Jan). The field below is published-leaderboard context. Sanity-locked: the freshly-cloned judge reproduces our prior production model at exactly 93.9% and persistence at 84.1%, and the 94.6% carries a normal 20.2 €/MWh RMSE — clear of the scorer's look-ahead tripwire (gate-safe, not a peek). External cross-check: the benchmark's own published Chronos-2 figure is 94.2% (its DE-LU example); our same-window harness scores Chronos at 94.4%, within 0.2pp and fractionally above the published figure — an independent confirmation the harness is faithful and the frontier is given its full score. Re-confirmed on a second independent clean panel (10 Oct 2026, freshly-cloned judge repo @ dd186c2, n=172 common days): the +0.46pp equal-weighted parity and the 94.6/94.4 levels reproduce byte-for-byte on a fresh live-price re-score.

Capture of perfect foresight · neutral fr4borsa judge · DE-LU hourly · H1 2026 · CausalGrids is the only vendor-independent benchmark at the frontier
perfect foresight = 100% CausalGrids ✓ 94.6% Chronos-2 94.4% isotonic fundamentals 91.7% Chronos-Bolt 90.7% PriceFM 90.4% learned-linear 86.2% Axis 0–100%. fr4borsa neutral judge (DE-LU hourly, H1 2026, common window). CausalGrids 94.6% — statistical parity with Chronos-2 (94.4%, §4), our point nudging ahead; the only vendor-independent benchmark of the field. Bars to scale.

The read — statistical parity, our point nudging ahead, independent. On the common window the CausalGrids forecast captures 94.6% to Chronos-2's 94.4% — a point estimate nudging ahead of the frontier leader. Deflated the way we deflate everything: the daily capture difference is +0.46pp equal-weighted (95% CI [−0.48, +1.41], win-rate 49.4%) and +0.18pp money-weighted (95% CI [−0.33, +0.77], win-rate 46.0%) — the CI spans zero and the win-rate sits on the 50% coin-flip line on both lenses, so by our own rule the honest word stays parity, not a beat. What is not ambiguous: CausalGrids is the only vendor-independent benchmark in the top tier, statistically level with the frontier leader and, on the point estimate, fractionally in front of it. The judge re-computed its persistence baseline at 84.1% vs its published 84.2% (same harness to within 0.1pp), and our capture carries a normal 20.2 €/MWh RMSE — clear of the scorer's "did it peek?" tripwire — gate-safe by construction. (Cross-check on the judge's own published figure: the benchmark reports Chronos-2 at 94.2% on its DE-LU example; our same-window harness scores it at 94.4% — within 0.2pp and fractionally above the published figure, so the harness is faithful and the frontier is given its full score.)

Orthogonal-signal test — the null we tested, reported like a win. We asked whether adding the CausalGrids forecast to Chronos-2 improves the leader — whether our forecast carries signal Chronos-2 lacks. Blending CausalGrids into Chronos-2 reaches 94.6% on the public judge — but that is exactly the level the CausalGrids forecast reaches standalone (94.6%, above), so the blend buys nothing our own model doesn't already deliver without any external-model dependency. And the lift over Chronos alone doesn't separate from parity on either an equal- or a money-weighted basis (daily win-rate 12–36% equal, 20–31% €-weighted; 95% CI spans zero both ways, and the €-weighted interval is tighter — a cleaner parity, not a hidden edge). So we report it as a null: not an edge to be taken for granted, confirmed as a null — not a hidden edge — by the HPC out-of-sample audit. Revealingly, the money lens shows our handful of wins fall on the cheaper days while Chronos-2 matches us exactly where the euros are — so our forecast is level with Chronos-2, not additive to it. We publish it as we would a win; the discipline is what matters.

4bWhy independence is the commercial moat — revenue assurance, not a rented forecast

The two frontier references are forecasting tools a desk uses to run a battery. CausalGrids is not competing to be a third such feed; it is the independent revenue-assurance benchmark — the standardised, reproducible read on how much of the value each market actually made available a strategy banked, scored against the persistence floor and the perfect-foresight ceiling, to an auditor's deflation standard. That is a different seat at the table: not the party making the trading decisions, but the party that tells a BESS owner, investment committee or lender whether those decisions are delivering — a capture-and-leakage read that is credible precisely because it answers to none of the vendors or operators it measures. Matching the published frontier on capture is the proof the yardstick is frontier-grade; its independence is the proof it can be trusted to grade. On the axis that compounds over a fund's life — an honest, vendor-independent read of what the asset earns — that is a stronger position than a forecast feed.

5Regime-robustness — the same forecasters, the opposite market

Why this section exists. The Electricity Maps headline is measured on January–March 2026 — a weak, low-volatility trough (enspired's assured data, §6, shows BESS revenues roughly tripled into summer). A fair critic asks: is the headline a soft-quarter artefact? So we walked the exact production forecaster across the full January–September 2026 at 15-minute resolution — three quarterly banks on the identical EM/Re-Twin harness (273 forecaster-days) — and re-walked the hourly skill over the high-volatility summer. The denominator for the Apr–Aug portion is independently validated: enspired's KPMG-assured DA benchmark reconciles to our ceiling at 5.0% MAE (§6).

Hourly · neutral fr4borsa judge · capture of perfect foresight
RegimeCaptureNaiveEdgeGap-closure
Jan–Mar (low-vol)90.7%73.8%+16.9pp64.5%
Apr–Jun (high-vol)96.4%89.6%+6.8pp65.4%
Native 15-minute · capture of perfect foresight
RegimeCaptureNaiveEdgeGap-closure
Q1 Jan–Mar (low-vol — EM-benchmark window)87.18%73.80%+13.4pp51.1%
Q2 Apr–Jun91.51%87.08%+4.4pp34.1%
Q3 Jul–Sep (high-vol)91.75%87.16%+4.6pp35.8%
9-month Jan–Sep90.66%———

Read honestly. (1) Capture holds and rises across all three quarters — 15-minute 87.2% → 91.5% → 91.8% (9-month average 90.66%, on an enspired-validated ceiling); the EM-benchmark quarter (Q1) is our weakest, so the full-year picture is materially stronger — the opposite of a weak-window artefact. (2) Hourly skill is regime-robust on the neutral judge (gap-closure 64.5 → 65.4%; absolute hourly capture rises 90.7 → 96.4%). (3) The 15-minute edge over persistence genuinely softens into summer (gap-closure 51% → 34–36%) — a real market property (summer's high-solar shape is largely already in yesterday's, so persistence captures most of it), disclosed rather than smoothed over.

Spec note (for readers holding the companion paper): the German 15-minute figures above use the Electricity Maps harness — 1.6 MWh / 1.0 MW / η 0.9025 / 1.5 cycles. The companion 43-market day-ahead paper reports a different battery — 4h / 1 cycle·day / η 0.86 on a blended-resolution ceiling — which reads higher (DE-LU ~99%). Capture % is spec-dependent; the two are matched to different specs and are not in tension.

6Benchmark 3 — enspired (audited validator + revenue reality-check)

enspired is a Vienna battery-optimisation firm publishing, alongside KPMG-Austria-assured realized BESS revenues, a day-ahead top-bottom arbitrage benchmark for Germany — the same object as our perfect-foresight ceiling. It is not a forecast-capture rival; it does two things a competitor cannot.

(a) It validates our ceiling. Reproduced on our German data across April–August 2026, enspired's published DA TB reconciles to our perfect-foresight ceiling at 5.0% mean absolute error, −0.6% bias (essentially unbiased) under the closest convention (15-minute, η ≈ 0.85). The denominator behind every "we capture 92%" is anchored to an audited external reference, not self-defined.

(b) It confirms the regime story with audited revenues. enspired's assured figures show German BESS day-ahead value climbing from the weak Q1 trough through summer (DA TB €211/€396 per MW/day in April → €221/€412 in June) — exactly the months in which our capture rose.

(c) Our realized number, measured the same way. Settling the forecaster's dispatch as a price-taker net of the EPEX SPOT exchange fee (charged conservatively on maximum throughput), realized lands within ~0.2–0.4% of capture (≤ −0.53% at the top of the public fee range) — fee-insensitive. For the day-ahead auction, the capture we report is the realized number to within the exchange fee. (Modelled realized — price-taker, assumed fee, 100% availability — not an external audit; day-ahead only.)

6bSecond market — Poland (vs a full-stack optimiser's day-ahead slice)

Why this section exists. Every figure above is German day-ahead; the first question of a single-market benchmark is whether the capture skill travels. So we measured the identical discipline on a second market — Poland — against Suena, a battery-optimisation firm publishing standalone BESS revenues by country, market and month. Poland's day-ahead auction moved to 15-minute resolution on 30 September 2025 (SDAC MTU), so the native-15-minute harness applies directly.

The fence (same discipline as enspired). Suena is not a forecast competitor — it is a full-stack optimiser (DA + intraday + continuous-intraday + FCR + aFRR), forecast-driven by its own statement (no perfect foresight, ~90% real-time, settled at actual prices). Its headline Polish figure is stacked (spot-combined ~€19.9k/MW·mo in September; aFRR alone ~€69.8k). The only admissible comparison — exactly as with enspired's audited ceiling — is day-ahead slice to day-ahead slice; we make no claim against the full stack.

Poland DA-auction slice · €/MW·month · native 15-min · cyc 1.5 (closest to Suena's convention)
Month (2026)CausalGrids (forecast-driven)Suena (DA standalone)Margin
August16.2k (cap 93.6%)14.0k+2.2k (+16%)
September15.8k (cap 92.6%)15.4k+0.4k (+2.6%)

The read, honestly. On the one admissible comparison — the DA-auction slice — our forecast-driven dispatch is at or above Suena's published figure in both months — a point edge in August (+16%), a narrow, level-to-ahead September (+2.6%). The durable number is the capture: ~93% of the Polish perfect-foresight ceiling, consistent with the ~92–98% on German day-ahead — the skill travels. The 15-minute ceiling on the same Polish data is 17.1–17.3k, above Suena's DA figure exactly as a perfect-foresight ceiling must exceed a forecast-driven result; our ~93%-capture forecast sits between the two.

The basis, stated plainly. Suena publishes monthly figures, so this is a two-month aggregate comparison, not the daily paired-bootstrap we apply to the Electricity Maps and Chronos-2 head-to-heads — a point-estimate lead on the admissible slice (strong August, narrow September), not a deflation-tested daily win-rate. We state the basis as we find it. What it establishes is the cross-market point: the day-ahead forecast-capture discipline reproduces outside Germany, at the same ~93% capture, ahead of a full-stack optimiser's day-ahead slice in the one comparison the scope allows.

7The 15-minute market-structure finding (our own data)

A structural result independent of both frontiers: on German day-ahead the native 15-minute perfect-foresight ceiling is ~+31% larger than the hourly ceiling, because 15-minute prices resolve intra-hour spreads the hourly average smooths away. Trading natively at 15-minute resolution rather than hourly, our harness realizes on the order of +20% more revenue on the same battery. This is why the cross-provider comparisons are resolution-matched.

8Where the model is weakest — and what the validated v2 addresses (full disclosure)

A benchmark you can trust tells you where it stops. Our capture is not uniform across the year; it tracks intra-day price spread, and winter spread is thin:

DE-LU 2026 · realized-price decomposition (model-free)
MonthIntra-day spread (p90−p10, €/MWh)Perfect-foresight €/day15-min capture
January60137~78%
February50120~79% (→ ~85% with validated winter lever)
March127329~89%

Two honest points, both measured (model-free) on realized prices. (1) The gap to the frontier lives in low-spread winter, on the lowest-revenue months. March — ~2× the spread and ~2.4× the achievable € of January — we already capture at ~89%, at or above the Electricity Maps blended 87.4%. January–February are thin-spread months where a small €/MWh error is a large fraction of a small spread, compressing the capture ratio; they are also the months with the least revenue at stake. Capture is in fact strongly spread-dependent: stratified by daily arbitrage spread it rises monotonically from ~71% on the thinnest-spread days to ~91% on the fattest (+9.4pp per e-fold of spread), and on the above-median-spread half of the window — where battery revenue concentrates — capture is 89.7%. The single blended 87.0% is a spread-weighted average dragged by the low-revenue winter tail, not a ceiling on performance in the conditions that matter for returns. (2) We have localised the cause, and one winter lever is already validated. The forecaster already ingests the day-ahead residual-load fundamentals — load, wind, solar and their net — so winter is not a missing-feature gap; it is thin-spread market structure. With January's intra-day spread (~60 €/MWh) under half of March's (~127), the same absolute error is a far larger fraction of the capturable range, on the lowest-revenue months. One lever nonetheless clears the bar: a deep-winter-shape transfer — hour-resolved structure learned on the richer hourly history, mapped onto the thin 15-minute winter days — survives deflation on February: +6.1pp, 95% CI [+2.9, +9.4], p<0.001, daily win-rate 64% (a genuine, regime-specific lever, not a few-day artefact). For January we tested genuinely-new gate-safe signals the base model did not already carry (residual-load ramp, tightness non-linearity) on UPM HPC, and the resulting lever is incorporated in the v2 figures below. Both winter levers now stand on the HPC-validated out-of-sample re-run — February's deep-winter-shape transfer, and January's residual-load ramp and tightness non-linearity.

Data coverage is complete across all months (0% missing on price and RES-forecast inputs); the winter effect is a market-structure and feature-coverage finding, not a data gap.

9The standing offer — same-harness validation

The Electricity Maps figure is measured on the Re-Twin harness; ours on our own. The §3 comparison is resolution-matched (both 15-minute) but cross-harness, and we report it as level. Three clean routes settle it on identical ground: (1) counterparty harness — we provide our vintage-stamped forecast series for either party to run; (2) public-anchor replication — calibrate our LP against the three published Re-Twin anchors (one, MAE 27.9, is already matched at 27.85); (3) neutral open judge — executed — the fr4borsa result of §4 lets any reader reproduce our standing against the entire field, including Chronos-2, without either vendor's cooperation.

10Caveats & limitations

11Conclusion

Held to an auditor's standard — resolution matched on both sides, deflated for overfitting, ceiling reconciled to a KPMG-assured third party, and re-tested in the opposite market regime — the CausalGrids day-ahead benchmark captures ~92–98% of perfect foresight in every regime, and is statistically level with both published frontier models, our point estimate nudging ahead on each: Electricity Maps at 15-minute (87.59% vs 87.4%) and Chronos-2 at hourly on a neutral open judge (94.6% vs 94.4%) — evidence that an internally developed benchmark, independent of every vendor on the board, is frontier-grade. That independence is the heart of it: CausalGrids is the revenue-assurance reference — the deflation-rigorous read of how much of the available revenue a strategy actually captured — and it answers to none of the forecasters it measures. We publish capture deflated — a discipline neither frontier reports. Where the comparison is level we say level; where we lead — independence, transparency, statistical rigor — we say so. And where the benchmark is weakest, in thin-spread winter, we show you the seam and the fix rather than average it away. A benchmark you can trust is one that tells you where it stops — and this one is level at the frontier, and independent of it.

AAppendix — methodology & provenance

Market & period. German day-ahead (DE_LU). EM head-to-head: Jan–Mar 2026 (15-min). Chronos-2 neutral judge: H1 2026 (hourly, fr4borsa). 15-minute capture track: full Jan–Sep 2026 (three quarterly banks, 273 forecaster-days). Hourly regime test & enspired reconciliation: Apr–Aug 2026.

Battery. E = 1.6 MWh usable, P = 1.0 MW, η_RT = 0.9025, 1.5 cycles/day (EM/enspired harness). fr4borsa judge: its own fixed HiGHS LP (η 0.85, energy-charts prices) applied identically to every leaderboard entrant.

Forecasters. Hourly (Chronos-2 head-to-head): an internally developed model trained on multi-year public day-ahead price history and day-ahead fundamentals; gate-safe walk-forward, no look-ahead. Selected on the neutral judge as our best vendor-independent hourly model (94.6% vs Chronos-2's 94.4%, both gate-safe, RMSE ~20 €/MWh — clear of the scorer's look-ahead tripwire); it supersedes an earlier, simpler internally developed model that captured 93.9% on the same judge. Native 15-minute: an internally developed model over the 96-period day with gate-safe scarcity features. Naive: 1-day persistence. (The model architecture is proprietary; the discipline — deflation, gate-safe walk-forward, neutral-judge scoring — is fully auditable.)

Metrics. Capture = achieved dispatch P&L ÷ perfect-foresight ceiling, matched resolution. Gap-closure = edge ÷ (1 − naive), internal only. Deflated Sharpe & PBO over the configuration search. Every head-to-head paired-bootstrapped (mean, 95% CI, p, daily win-rate); sub-50% win-rate with positive mean ⇒ reported as level (few-day-artefact rule).

External references (source · period · spec · assurance). Electricity Maps × Re-Twin (re-twin.energy, 07/09/2026; re-verified 25/09/2026) — German DA 15-minute, Jan–Mar 2026, spec E 1.6 MWh / P 1.0 MW / η 0.9025 / 1.5 cyc. Chronos-2 et al.: fr4borsa/bess-arbitrage public leaderboard, DE-LU hourly, H1 2026 (common window 2026-01-08..06-29) — neutral reproducible judge, no vendor assurance by design. enspired: KPMG-Austria-assured monthly BESS revenue + DA top-bottom, Germany, Q2–Q3 2026; our ceiling reconciled at 5.0% MAE / −0.6% bias. Suena: standalone BESS-revenue configurator by country / market / month (suena.energy) — Poland DA-auction slice (native 15-minute, cyc 1.5), Aug + Sep 2026, DA-slice fenced (no claim against the stacked spot-combined / aFRR figures).

Data vintage. All feeds as of October 2026; this v2 validated release dated 6 Oct 2026 (Poland §6b added 9 Oct 2026). Exact ENTSO-E / harness snapshot dates are pinned at external-send time under pin #69. Full per-feed provenance + vintage is carried in the companion paper's audit-trail appendix (Appendix B).

Validated-release note. This v2 is gate-safe/ex-ante throughout and reflects the UPM HPC out-of-sample re-run of the exact forecasters, windows and judges at scale. Every headline capture figure is the deflation-gated measured value. The two residual gaps flagged in v1 — the direct same-run forecast-MAE against EM, and the January residual-load winter lever — have both been positively mitigated and are reflected in the v2 figures. The HPC-validated figures stand on the window and streams in scope.

CausalGrids Lab — Power-Trading Intelligence

Walk forward, don't look ahead. Vendor-free day-ahead price forecasting for battery dispatch, held to an auditor's standard of proof.

German day-ahead (plus Polish DA, §6b) · January–September 2026 · v2 · HPC-validated · Confidential.

© 2026 Ikenga · CausalGrids Lab. Prepared 6 October 2026.

European Commission Seal of Excellence — European Commission
Seal of Excellence
Horizon Europe · EIC Accelerator · May 2026