The day-ahead premium, reproduced on an independent grid (NYISO)

The question

In our CAISO study we found two things about the day-ahead versus real-time settlement price: the day-ahead price runs a small but real premium over real-time, and the gap between them — once its daily rhythm is stripped out — still carries hour-to-hour memory. That raises an honest question the CAISO data alone can’t answer: is this a quirk of one grid, or a property of the two-settlement market design itself? The clean way to tell is to hold the mechanism fixed and change the grid — run the identical instrument on an independent system operator and see whether the same structure appears.

Data

NYISO’s public Market Information System, zonal locational marginal prices (LBMP): day-ahead (hourly) and real-time (~5-minute) settlement prices for the New York City zone, over six complete calendar months (February–July 2026). NYISO is an independent grid with directly accessible public data — no account, no key. We also carry the Capital zone as a within-grid check (more on why that is a check, not a second grid, below). Every figure here is our own derived analysis, not redistributed raw data.

Method

Same engine, same estimator, same pre-registered gates and null as the CAISO study — only the data source changes, which is exactly what makes this a test rather than a re-parameterization. The basis is real-time minus day-ahead LBMP (the total price, $/MWh; we keep, but never subtract, the loss and congestion components). NYISO’s five-minute real-time prices are interval-ending, so each is assigned to the hour it falls within — a convention we self-test and which does not move the hourly result — and averaged onto the day-ahead hourly clock. We report a robust center (median) so scarcity spikes are kept but cannot dominate, remove the day-of-week-and-hour rhythm, then test the leftover for serial structure against a value-permutation null with every parameter fixed before any data was touched. Eastern time throughout, including the spring-forward hour.

Findings

Across all six months, the day-ahead premium is present and statistically significant. In New York City the median real-time-minus-day-ahead basis is negative in 6 of 6 windows — day-ahead richer — and its bootstrap 95% interval excludes zero in every one, ranging from roughly −7% of the day-ahead price in late winter to a small but still-significant −2.7% (about $1/MWh) in the mildest spring month.

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 deseasonalized basis shows clear positive persistence in every window — lag-1 autocorrelation around 0.35–0.65, far above anything value-shuffled versions of the same series produce (z from +9.6 to +17.8). In other words, after the predictable daily-and-weekend pattern is removed, the day-ahead-vs-real-time surprise still has memory from one hour to the next. The premium is strongest and most persistent in winter and milder in summer — consistent with Northeast cold-snap gas dynamics, where forecast errors and price divergence run largest.

The Capital zone tells the same story, but it is a within-grid corroboration, not an independent second test: two zones on one grid share the system-wide energy component and move largely together, so agreement between them guards against a zone-specific fluke and nothing more. The two real axes of this study are temporal (six months) and cross-grid (NYISO versus CAISO).

Set against CAISO, the direction matches: the same day-ahead premium, the same positive persistence. Read directionally — not as a pooled statistical test — NYISO’s premium is somewhat milder in percentage terms than CAISO’s (about −9%), while its persistence runs a little higher (~0.35–0.65 against CAISO’s ~0.30). A plausible reason is that New York City is import-constrained, so congestion and forecast-error divergence linger across hours; we offer that as a hypothesis, not something we have adjusted for.

What this means

The day-ahead premium and its persistence are not a CAISO artifact. Holding the two-settlement mechanism constant and swapping to an independent grid, the same structure appears — which lifts the finding from “true on one market” to “a property of this market design.” The honest boundaries: this is six recent months, winter into summer, not a year-round claim; one grid plus a correlated companion zone, not a battery of independent grids; and it is a measurement, not a trade. The day-ahead-vs-real-time spread is settled and, in principle, tradable through virtual bidding — but this study measures the premium and its predictability only. It makes no claim that either is capturable after costs, and offers no buy, sell, or allocation advice.

Reproducibility: produced by our own pipeline (fetch → normalize to a common hourly clock → day-ahead-premium characterization → deseasonalize → serial-structure vs value-permutation null), reusing the CAISO study's analysis code unchanged and touching the shared engine only through its alignment primitive. An offline self-test injects a known premium and a known efficient (no-structure) control and confirms the code recovers the first and rejects the second — alongside grid-specific guards for real-time interval alignment, Eastern-time and the spring-forward hour, and zone hygiene — before any real data is touched. The finding was then independently reproduced against a different family of synthetic generators (stochastic-volatility clustering, Student-t tails, and power-price spikes and negatives), and both the premium and the predictability survived. The hour-to-hour predictability survives a high-pass that removes all weekly-and-slower structure in every one of the six windows (leaving the autocorrelation essentially unchanged), so it is genuine forecast-error persistence rather than residual seasonality or slow drift. A methods note we banked from this: a value-permutation predictability test on a stationary residual should high-pass before the autocorrelation, or the null can read a smooth low-frequency component as predictability.