Day-ahead vs real-time price integrity.

1. What this document is (and is not)

This is a methods-transparency document. Its purpose is to let a technically-fluent reader understand how we measure the integrity of the day-ahead vs real-time (DA–RT) electricity price relationship — the exact sources, windows, statistics, and null construction — well enough to judge the work, without exposing the engine’s source code.

It is not a sales sheet, a price or spread forecast, a trading signal, a securities or commodity research report, or a recommendation. It makes no directional call and no claim about future prices, spreads, or returns. It is not an Independent Market Monitor product and makes no claim to a regulated monitoring role. It is not an accusation: a systematic day-ahead premium is a market-structure feature, not evidence of manipulation or misconduct by any participant, venue, or ISO. It measures how a market-structure relationship behaves; tradability, cost, risk, and market impact were not measured and are not claimed, in either direction.

2. The problem we measure

A trading desk, utility, or gentailer builds models and hedges on the DA–RT relationship; a data or benchmark provider publishes analytics built on it. Each must be able to demonstrate — to an auditor, a board, a model-risk function, or an institutional client — that the relationship, and the models built on it, are behaving and calibrated. Independent, reproducible evidence of this is hard to obtain: internal QA is not independent; the ISO’s Independent Market Monitor reports on the market as a whole, not on your book or model, and its standing is regulatory rather than a second opinion you can commission; data vendors sell the feed and analytics, not an integrity attestation of the relationship; and bespoke consultancy opinions are not built for push-button external reproduction.

The position this document occupies: an independent, reproducible, method-transparent measurement of DA–RT relationship integrity, on free public ISO data, with honest reporting against a proper null — a second opinion / validation layer that complements the market monitor and never claims its role.

3. What we measure

1. DA–RT divergence — the behaviour of the gap between day-ahead and real-time prices over defined windows.
2. The systematic day-ahead premium — the average amount by which day-ahead prices sit above real-time, reported as a measured market-structure property, consistent with a forward risk premium (hedging pressure, forecast-error and weather uncertainty), and not attributed to inefficiency, gaming, or any participant’s conduct.
3. Serial dependence (autocorrelation) of the premium and its persistence — reported strictly as a calibration/health property of the DA–RT relationship: how the relationship behaves over time.
4. Null-validated throughout — every claim is tested against what a proper null produces.

Tradability, profitability after transaction costs, risk, and market impact are not measured and are claimed in neither direction.

4. Method transparency (the reproducibility core)

Method is disclosed faithfully; no engine source code is exposed.

4.1 Data sources. Free public ISO data: NYISO (New York City zone, with the Capital zone as a within-grid check) and CAISO (the SP15 and NP15 hubs). Single-provider-derived analysis, not redistributed raw data.

4.2 Sampling and windows. NYISO: six complete monthly windows, February–July 2026 (winter through summer). CAISO: July 2026, two hubs (SP15 and NP15). The series is deseasonalized before analysis — hour-of-day and weekday/weekend structure stripped — with negative prices and price spikes kept (not winsorized away), so the measured structure is the genuine residual, not an artifact of cleaning.

4.3 The statistics.

• Day-ahead premium: the median real-time-minus-day-ahead settlement, in $/MWh and as a percentage of the day-ahead price, with a bootstrap 95% confidence interval per window. (A negative RT−DA is a positive day-ahead premium: real-time settles below day-ahead.)
• Serial dependence: the lag-1 autocorrelation of the deseasonalized residual, with a high-pass that removes all weekly-and-slower structure applied before the autocorrelation, so smooth low-frequency structure cannot masquerade as hour-to-hour dependence.

4.4 Null construction and decision rule. A value-permutation null: the residual’s values are reshuffled i.i.d., destroying serial dependence while preserving the value distribution. (This — not an increment-shuffle random-walk surrogate — is the correct null for a stationary, deseasonalized power-market residual.) The decision bar is pre-registered and fixed before the windows are analysed.

4.5 Artifact guard. A stale / outlier / low-information guard is applied; low-information periods are masked and the test re-run. Negative prices and spikes are retained as genuine market events, not masked.

4.6 Independent reproduction — two distinct senses, kept separate.

• (a) Cross-grid replication (external validity). The instrument, run unchanged on a second, independent grid, produced the same qualitative result — see §5.4. This is replication across grids, not adversarial hardening.
• (b) Different-generative-family reproduction (adversarial hardening). The result also survived reproduction against a genuinely different family of synthetic generators — stochastic-volatility clustering, Student-t tails, and power-price spikes and negatives — importing the fixed core and the published harness unchanged. Both the premium and the serial dependence survived. This is distinct from, and additional to, the cross-grid replication in (a).

5. Worked demonstration — the DA–RT relationship on NYISO and CAISO (a completed, reproduced worked study)

5.1 Setup. NYISO New York City zone, six monthly windows Feb–Jul 2026 (Capital zone as a within-grid check); CAISO SP15 and NP15 for July 2026; deseasonalized residual; value-permutation null; high-pass applied before the autocorrelation.

5.2 The day-ahead premium (NYISO). Real-time settled below day-ahead — a systematic positive day-ahead premium — in all six windows; the bootstrap 95% confidence interval excludes zero in every one. The magnitude ranges from about −2.7% (April) to −6.9% (March) of the day-ahead price.

Month	RT−DA median ($/MWh)	95% CI	% of DA	lag-1 ACF	z vs null
Feb 2026	−4.28	[−5.57, −2.60]	−5.2%	+0.65	+17.8
Mar 2026	−3.14	[−4.13, −2.21]	−6.9%	+0.35	+10.3
Apr 2026	−1.03	[−1.39, −0.47]	−2.7%	+0.35	+9.6
May 2026	−2.00	[−3.06, −1.29]	−5.3%	+0.36	+9.9
Jun 2026	−1.80	[−2.66, −1.23]	−4.0%	+0.41	+11.9
Jul 2026	−2.00	[−3.03, −0.91]	−3.7%	+0.50	+15.2

The premium is reported as a measured market-structure property, consistent with a forward risk premium — not attributed to inefficiency, gaming, or any participant’s conduct.

5.3 Serial dependence (the calibration/health property). The deseasonalized premium exhibits measurable serial dependence (lag-1 autocorrelation ≈ 0.35–0.65 every window, z +9.6 to +17.8 against the value-permutation null) that persists beyond an i.i.d. reshuffle of its values and survives a high-pass removing all weekly-and-slower structure in all six windows, with the autocorrelation essentially unchanged — indicating genuine hour-to-hour forecast-error persistence rather than seasonality or slow drift. This is reported strictly as a calibration/health property of the DA–RT relationship — how the relationship behaves over time. Tradability, profitability after transaction costs, risk, and market impact were not measured and are not claimed, in either direction. No forecast, no directional read, no period-profitability ranking, no bid-shaped output.

5.4 Cross-grid replication (CAISO). The same qualitative result — a positive day-ahead premium with hour-to-hour serial dependence — appears on a second, independent grid: CAISO, July 2026, on two hubs. SP15 carried a premium of roughly −9% of the day-ahead price with lag-1 autocorrelation of about 0.30; NP15, the replication hub, showed a smaller premium (about −4%) with stronger persistence (about +0.70). The persistence survives the weekly-and-slower high-pass on both hubs. This is a directional cross-grid agreement, not a pooled or meta-analyzed statistic. The two grids differ: NYISO is milder in percentage terms (≈ −3% to −7%) with somewhat higher persistence (≈ 0.35–0.65) than CAISO (≈ −9%, ≈ 0.30). An import-constrained-NYC explanation for the difference is an offered hypothesis, not something this study adjusts for. Note also the asymmetric coverage: NYISO carries the six-month temporal axis; CAISO is a single month across two hubs.

5.5 Different-family reproduction. As §4.6(b): both the premium and the serial dependence survived reproduction against a different family of synthetic generators (stochastic-volatility clustering, Student-t tails, power-price spikes and negatives), core and published harness unchanged.

6. Honest reporting against the null

Every claim here is tested against a proper null (§4.4). The value-permutation null could have returned “no serial dependence beyond chance,” in which case that is what we would report; here it returned strong, consistent dependence, and that is reported honestly with its statistics. A method that only ever “finds something” is unfalsifiable; testing against a null that can return nothing — and reporting it when it does, across the wider body of work — is the evidence that a positive finding is real and not manufactured.

7. What this is not

• Not an Independent Market Monitor — we hold no regulated monitoring mandate; we are an independent second opinion / validation layer that complements the monitor.
• Not a data vendor — we do not sell the feed; we measure whether the relationship and the models built on it are genuine and calibrated.
• Not black-box surveillance — we pre-register, reproduce on free public data, and report against a proper null; the method is auditable.
• Not a forecast, signal, or recommendation — no directional read, no spread or premium forecast, no period-profitability ranking, no bid-shaped output, no “alpha.”
• Not an accusation — a systematic day-ahead premium is a market-structure feature consistent with a forward risk premium; nothing here implies misconduct by any named participant, venue, or ISO.

8. How an engagement works

1. Free first read — a calibration/integrity snapshot on a subset of one market’s public DA/RT data, at no cost and no commitment, so your team sees the method live.
2. Integrity / validation assessment — a fixed-scope assessment on defined market(s) and windows, delivered as a report formatted for a model-validation file, with a per-engagement reproducibility appendix (sources, windows, statistics, null — no engine code).
3. Monitoring (optional, later) — a recurring integrity/calibration re-run once an assessment has landed. Deeper-scope engagements (additional grids, additional integrity diagnostics) use client-supplied or engagement-specific data, procured only for that engagement.

9. Scope and limitations

1. Tradability, cost, risk, and market impact were not measured — no exploitability claim is made in either direction. (This study ran no cost-band clearing; the “sub-cost / not tradable” claim available in other domains is deliberately not made here.)
2. The day-ahead premium is consistent with a forward risk premium and is not attributed to inefficiency, gaming, or any participant’s conduct.
3. Attribution is weather / forecast-error confounded — what drives the premium is not isolated; the finding is the presence and structure of the premium, not its cause.
4. Six recent months (Feb–Jul 2026), winter through summer — not a year-round claim.
5. One grid plus a correlated companion zone, not a battery of independent grids. NYISO’s Capital zone shares the system energy component with NYC and is a within-grid corroboration, not an independent test. The two genuine axes are temporal (six NYISO months) and cross-grid (NYISO vs CAISO).
6. CAISO is a single month (July), two hubs (SP15, NP15) — the cross-grid result is a directional match, not a multi-month CAISO claim and not a pooled cross-grid statistic.
7. This is exploratory, single-provider-derived analysis of public ISO data — measurement, not investment or trading advice.
8. Serial dependence is a statistical property of the historical residual, reported as calibration/health — it is not a forecast.
9. Not an Independent Market Monitor function; not an accusation of manipulation against any participant.

10. Roadmap — not demonstrated in this document

Additional grids, additional months, and additional integrity diagnostics are natural extensions, each measurement-framed and claimed only once built and independently reproduced. Nothing in this document claims them.

---

Delta Motion Analytics is a cross-border analytics studio providing measurement and reproducibility infrastructure. This document describes method, not source code, and does not constitute investment or trading advice or a recommendation regarding any security, commodity, or asset.