06-reference/research

pjm peak load forecast power cycle corroboration

2026-07-28·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

PJM peak load vs forecast, 2017-2026 — does it corroborate the ERCOT-only anchor?

The question

"Does PJM's peak-load-vs-forecast data (2017-2026, via Wayback Machine) corroborate the ERCOT-only evidence anchoring the power-cycle thesis?"

Context: power-cycle-v1.1 has $20,000 of paper capital deployed. Its phase-marker #1 is stated as a PJM + ERCOT test, but the anchor dataset backing it contains ERCOT rows only — [[2026-05-17-power-anchors-backfill-notes]] flagged the PJM gap and never closed it.

Headline: PJM does NOT corroborate the ERCOT reading. The two ISOs point in opposite directions on realized peak load, and both directions are mostly weather artifacts. On the thesis's own marker-1 threshold, PJM's verdict flips depending on whether you use raw or weather-normalized peaks — meaning marker-1 as currently written is not decidable, not merely unconfirmed.

What we already know (from the vault)

What the web says

Retrieval note: Wayback was not needed. PJM keeps every Load Forecast Report vintage live at a stable URL pattern, so all ten vintages (2017-2026) were fetched directly from pjm.com in this run and parsed locally. Every number below comes from a PDF or XLSX I downloaded and read on 2026-07-28.

Table 1 — PJM RTO summer peak: vintage-N forecast for year N vs realized

Error = actual − forecast. Positive = PJM under-forecast. All forecast figures are the PJM RTO row of Table B-1 (summer peak, MW, non-coincident unrestricted) in the January report of that vintage. All actuals are Table F1-Summer of the Jan-2026 tables workbook (PJM restates prior-year WN values each vintage; using one vintage throughout avoids mixing restatements).

Vintage (Jan) Target yr Forecast MW Actual MW Err MW Err % WN actual MW Err vs WN %
2017 2017 152,999 145,434 −7,565 −4.94% 149,573 −2.24%
2018 2018 152,108 150,573 −1,535 −1.01% 149,707 −1.58%
2019 2019 151,358 151,302 −56 −0.04% 149,088 −1.50%
2020 2020 148,092 144,320 −3,772 −2.55% 147,175 −0.62%
2021 2021 149,224 148,433 −791 −0.53% 148,912 −0.21%
2022 2022 148,938 147,361 −1,577 −1.06% 149,000 +0.04%
2023 2023 149,059 146,799 −2,260 −1.52% 149,830 +0.52%
2024 2024 151,247 152,314 +1,067 +0.71% 151,025 −0.15%
2025 2025 154,144 160,709 +6,565 +4.26% 153,503 −0.42%
2026 2026 156,373 not retrieved not retrieved

2026 actual is not retrieved: the 2026 summer season is still in progress as of 2026-07-28, and PJM has not posted a summer-2026-peaks-and-5cps file (URL probed, returns /not-found). Do not evaluate the 2026 row until PJM's Jan-2027 report or the summer-2026 5CP sheet publishes. This is the same partial-year trap the backfill notes flagged on ERCOT's 2026 row.

All 9 non-2026 rows are backed by a document retrieved and parsed in this run. No figure is interpolated, estimated, or recalled.

Table 2 — Forecast for fixed target year 2030, by vintage

Vintage 2030 forecast MW Δ vs prior vintage
Jan-2020 157,132
Jan-2021 153,536 −3,596
Jan-2022 153,775 +239
Jan-2023 157,899 +4,124
Jan-2024 167,873 +9,974
Jan-2025 183,883 +16,010
Jan-2026 183,008 −875

Jan-2017 through Jan-2019 vintages do not reach 2030 (11-year horizon) → not applicable, not missing.

Table 3 — ERCOT side, for comparison (vault-sourced, NOT re-verified in this run)

Year ERCOT actual peak MW ERCOT-Adj forecast MW YoY vs forecast
2023 85,464 not in dataset +6.78%
2024 85,199 not in dataset −0.31%
2025 83,679 84,686 −1.78% −1.19%
2026 75,572 (partial-yr) 93,504 do not use

Source: iso-demand-growth-2017-2026.csv, scraped 2026-05-17 by power-anchors-scraper.py. I did not re-fetch ERCOT primary data in this run — these rows are carried from the vault as-is. ERCOT forecast columns are blank pre-2025 because the Peak Demand Scenarios workbook only publishes forward years.

Convergences and contradictions

Synthesis for RDCO

The 30-minute follow-up was worth more than 30 minutes. The PJM backfill does not turn marker-1 from "unconfirmed" into "confirmed" or "broken" — it shows marker-1 is specified badly enough that it cannot be evaluated as written, and that the ERCOT-only version was giving a falsely clean signal in the bearish direction just as a PJM-only version would have given a falsely clean signal in the bullish direction. Marker-1 says "annual peak-load growth vs forecast" without naming a basis. On raw non-coincident unrestricted peaks, PJM clears the 3.3% bullish threshold in both 2024 (+3.76%) and 2025 (+5.51%), and ERCOT trips the bearish flip on a literal read (materially below 9.6% for two consecutive annual cycles). On weather-normalized peaks — the basis PJM itself publishes precisely so that year-over-year comparisons mean something — PJM fails the bullish threshold in both years (+0.80%, +1.64%), and the whole apparent regime change evaporates: PJM's forecast error against WN actuals has been within ±2.3% for a decade with no directional break. Two metrics, two opposite verdicts, on a marker whose v1 consequence was closing a $20,000 bucket. That is the finding.

The second finding is more actionable and cuts against the position's timing, not its direction. PJM's Jan-2026 vintage trimmed target years 2026 through 2032 versus the Jan-2025 vintage (−1.61% for 2026, −2.27% for 2027, −2.60% for 2028, −2.59% for 2029) and raised 2033-2035 (up to +3.31% for 2035). PJM did not lose faith in the data-center load; it moved roughly one to two years of it to the right. power-cycle-v1.1 sizes the basket on being in phase 2 heading into phase 3, and claims that transition historically pays +60-120pp "in 12-18 months." A one-to-two-year rightward shift in the largest data-center ISO's own load curve is a direct hit on that 12-18-month clock — not on whether phase 3 arrives, but on whether it arrives inside the window the sizing assumes. The v1.1 anchor list (five founder-review triggers) has no trigger that would catch this, because none of them watch forecast-vintage deltas; they watch queue levels, PPA cadence, GEV orders, 13F deltas, and power prices. A "delay, not cancel" revision is invisible to all five.

The third finding is about evidence quality. The strongest-looking PJM datapoint — a +19.0% upward revision to the 2030 forecast in four vintages — is the weakest evidence in this brief, because PJM's forward forecast is built from large-load interconnection requests submitted by developers, which is the same input stream the thesis's own bear case identifies as speculative and double-counted. When PJM's forecast rises and ERCOT's forecast rises, that is one signal observed twice, not two signals. The genuinely independent PJM contribution is the realized weather-normalized series, and it says load growth accelerated from +0.06% (2022) to +0.56% (2023) to +0.80% (2024) to +1.64% (2025) — a real, monotonic, four-year acceleration, but running at roughly half the rate marker-1 calls bullish. The honest position: the demand story is real and accelerating, and it is arriving slower than the thesis priced.

None of this is an exit signal. The basket's return case in [[2026-05-18-power-cycle-v1-v2-honest-rerun]] rests on structural supply scarcity (queue depth ~13 years at ERCOT's clearing rate) far more than on realized peak-load prints, and nothing here touches the queue. But the PJM backfill removes the option of citing marker-1 as supporting evidence in either direction until the basis is fixed, and it puts a dated, PJM-sourced number on a timing risk that the current anchor set cannot see.

Why this is in the vault

This closes the PJM gap explicitly named in [[2026-05-17-power-anchors-backfill-notes]] and directly changes how phase-marker #1 of [[2026-05-18-power-cycle-v1.1]] must be evaluated at the next quarterly review of the $20,000 power bucket — specifically, it establishes that the marker needs a weather-normalization basis before it can trigger anything, and it surfaces a PJM-sourced timing risk (demand pushed 1-2 years right) that none of v1.1's five founder-review triggers currently monitor.

Open follow-ups

Related

Sources

Retrieved and parsed in this run (2026-07-28):

Probed and confirmed unavailable:

Vault (not re-verified in this run):

Paywalled/skipped: none encountered.

Wayback Machine: not used. All ten forecast vintages were available live on pjm.com, which is a stronger provenance than an archive snapshot. Noted because the question presumed Wayback would be required.