06-reference/research

pjm jan2026 load forecast trim methodology vs slippage

2026-10-10·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

PJM's Jan-2026 near-term trim: slipped projects, or a new large-load haircut?

The question

"Did PJM's Jan-2026 near-term load forecast trim come from data-center projects actually slipping in-service dates, or from a Manual 19 / large-load-haircut methodology change?"

Context: [[2026-07-28-pjm-peak-load-forecast-power-cycle-corroboration]] established that PJM's Jan-2026 vintage cut target years 2026-2029 by 1.6-2.6% versus Jan-2025 while raising 2033-2035 by up to 3.3%, and left the cause open. power-cycle-v1.1 has $20,000 of paper capital deployed against a phase-2-into-phase-3 timing call that this cause materially affects.

Headline: mostly the haircut, with one genuine exception at the front edge. For target years 2027-2029, where the deepest cuts land, PJM's effective inclusion rate on large-load requests collapsed from 82-84% to 60-74% while the utilities' submitted volume held flat or rose sharply. That is a haircut, not slippage. For target year 2026 alone it is the reverse: submitters themselves cut their 2026 ask by 1,530 MW (-9.9%), and PJM's inclusion rate for that year actually ticked up, so the nearest year's large-load reduction is submitter-side, not PJM-side. Separately, the question's "Manual 19 revision" premise is false: the haircut rules lived in a standalone implementation guideline dated 2025-07-01 and did not enter Manual 19 until Revision 39, effective 2026-06-30, six months after the forecast published.

Confidence: high that the haircut is the dominant cause of the 2027-2029 cuts, and high on the Manual 19 timing correction. Medium on the precise split for target year 2026, where the two submitted-request snapshots differ in zone granularity (see Table 2 caveats).

What we already know (from the vault)

What the web says

All figures below were read from documents downloaded and parsed on 2026-10-10. Nothing is recalled.

Table 1 — Jan-2026 minus Jan-2025, by target year

All rows are the PJM RTO row of the named sheet in each vintage's Load Report Tables workbook, compared on identical target years. B-9b = "Total Load Associated with Adjustments to Summer Peak Load." B-9 = "Adjustments Above Embedded to Summer Peak Load," the incremental amount layered onto the forecast. B-1 = total non-coincident unrestricted summer peak.

Target yr B-9b 2025LF B-9b 2026LF Δ B-9b B-9 2025LF B-9 2026LF Δ B-9 Δ B-1 total Δ B-1 %
2026 12,554 11,469 −1,085 8,453 6,403 −2,050 −2,564 −1.61%
2027 17,721 15,856 −1,865 13,668 10,785 −2,883 −3,735 −2.27%
2028 23,418 21,280 −2,138 19,320 16,242 −3,078 −4,414 −2.60%
2029 29,483 27,361 −2,122 25,334 22,290 −3,044 −4,564 −2.59%
2030 37,050 38,805 +1,755 32,795 33,707 +912 −875 −0.48%
2031 45,361 46,638 +1,277 41,101 41,477 +376 −1,630 −0.85%
2032 52,767 53,949 +1,182 48,531 48,772 +241 −1,592 −0.79%
2033 55,683 60,258 +4,575 51,509 55,089 +3,580 +1,887 +0.92%
2034 57,822 65,041 +7,219 53,679 59,889 +6,210 +4,629 +2.23%
2035 59,413 68,968 +9,555 55,239 63,796 +8,557 +6,949 +3.31%

MW. Negative = Jan-2026 lower than Jan-2025 for the same year. Δ B-1 % is computed from the Table B-1 PJM RTO base values (Jan-2025: 158,937 / 164,186 / 169,981 / 176,094 / 183,883 / 192,647 / 200,507 / 204,197 / 207,253 / 209,923; Jan-2026: 156,373 / 160,451 / 165,567 / 171,530 / 183,008 / 191,017 / 198,915 / 206,084 / 211,882 / 216,872). The sign flip on both adjustment measures occurs at 2030.

On the B-9 above-embedded measure, the large-load adjustment accounts for roughly 67-80% of the total near-term trim (2,050/2,564 in 2026; 3,044/4,564 in 2029), with the EV down-revision contributing another 9-20% (218/2,564 and 913/4,564). Treat those shares as arithmetic indications, not a formal variance decomposition — PJM re-estimates the whole model each vintage and removes embedded data-center load from history before layering adjustments, so B-9, B-8c and B-1 are not strictly additive.

Table 2 — The slippage test: submitted versus submitted, across two cycles

This is the comparison that discriminates. Columns 1 and 2 are what the utilities asked PJM for in each cycle, before any PJM processing. Columns 3 and 4 are what PJM included. The inclusion rate is the haircut.

Target yr Asked, 2024 cycle Asked, 2025 cycle Δ ask Included 2025LF Incl. rate Included 2026LF Incl. rate
2026 15,398 13,868 −9.9% 12,554 81.5% 11,469 82.7%
2027 21,186 21,402 +1.0% 17,721 83.6% 15,856 74.1%
2028 27,809 32,391 +16.5% 23,418 84.2% 21,280 65.7%
2029 36,072 45,796 +27.0% 29,483 81.7% 27,361 59.7%
2030 51,943 60,117 +15.7% 37,050 71.3% 38,805 64.5%
2031 57,682 71,552 +24.0% 45,361 78.6% 46,638 65.2%
2032 61,255 81,463 +33.0% 52,767 86.1% 53,949 66.2%
2033 64,288 88,056 +37.0% 55,683 86.6% 60,258 68.4%
2034 66,296 93,554 +41.1% 57,822 87.2% 65,041 69.5%
2035 67,971 98,622 +45.1% 59,413 87.4% 68,968 69.9%

MW, PJM RTO summer peak. 2024-cycle asks are the sum of the eleven zone rows (DLCo, BGE, DEOK, PECO, PSEG, APS, ATSI, ComEd, PPL, AEP, DOM) in the 2024-10-25 LAS post-meeting workbook, whose header states "Below values show the amount requested for consideration not the amount included in the 2025 Load Forecast." That sheet has no RTO total row, so the sum is mine. 2025-cycle asks are the PJM RTO row of the "2026 DEMAND Requests" sheet in the 2025-09-16 LAS informational workbook (its own caveat: "Note these are supplied to PJM by EDC/LSE and are not Final Large Load Adjustments"). Included columns are Table B-9b.

Caveats, and which way they cut. The two snapshots differ in zone granularity: the 2024 workbook has one DOM row where the 2025 workbook splits DOM / NOVEC / ODEC / REC, and DAYTON, DQE, PEPCO, SMECO, JCPL and METED appear only in the later sheet. That makes the 2024-cycle column a likely undercount of the full 2025-cycle submitter set, which biases against the 2026 slippage finding rather than for it: a fuller 2024 baseline would make the 2026 ask decline larger, not smaller. The snapshots are also taken at different points in the cycle (October 2024 vs September 2025), and the 2025 workbook flags one later downward correction (REC-APS). B-9b includes a small amount of non-data-center adjustment, notably EKPC peak shaving at −10 MW and Dominion's negative voltage-optimization adjustment (Table B-9b, EKPC and DOM rows, 2026 workbook). None of that is material against a 9.7 GW swing in the 2029 ask.

Convergences and contradictions

Synthesis for RDCO

The useful finding is that the Jan-2026 near-term trim carries almost no demand information in the years where it looks worst. For target years 2027-2029, the utilities either held their ask flat or raised it by up to 27%, and PJM's effective inclusion rate fell from the low eighties to the sixties. The published number went down because the scorer changed, not because the projects moved. The one real exception is the nearest year: submitters cut their own 2026 ask by 1,530 MW while PJM's inclusion rate for 2026 ticked up, so the 2026 reduction is a genuine front-end slip. One year of front-end slippage inside a cycle where the ten-year ask rose 45% is not a demand-softening signal. For power-cycle-v1.1, the practical effect is that the 2.6% cut to 2028-2029 that looked like the most alarming datapoint in [[2026-07-28-pjm-peak-load-forecast-power-cycle-corroboration]] should be removed from the thesis's evidence base. The EV and macro residual, roughly 20-33% of the trim, stays in as an ordinary input revision; it is just not about data centers.

Recommendation: (b) add a sixth founder-review trigger. Not (a) change the thesis: direction is untouched, and the structural-scarcity core in [[2026-05-18-power-cycle-v1-v2-honest-rerun]] never ran through PJM's forecast vintage. Not (c) do nothing: that leaves a real instrumentation hole, because the parent brief's timing concern was raised against a number that turns out not to mean what it looked like, and the current anchor set cannot tell the difference. The sixth trigger should be constructed so it would not have fired on the Jan-2026 trim, which means keying it to the submitted-side and contract-backed series rather than the headline forecast. The three artifacts that carry the signal publish on a fixed annual cadence: the September LAS submitted-requests workbook, the November LAS preliminary summary with its Firm / Non-Firm split, and PJM's new May Data Center Accuracy Report. Reviewable events are a year-over-year decline in Firm (ESO/CC-backed) MW, a sustained fall in the inclusion rate below the 60-70% band this cycle established, or realized data-center load landing materially below the Firm-only forecast in the accuracy report. Headline forecast-vintage deltas should be explicitly disqualified as trigger inputs. No trade, allocation change, sizing change or deploy is proposed here, and none should be read into this brief.

There is a second-order consequence worth naming because a bearish reading of the trim would miss it. The rule is not only that non-firm load gets derated; it is that only Firm impacts RPM. PJM now procures capacity against a deliberately conservative near-term load number while the utilities' own submitted load runs 1.2x to 1.7x higher than what is included (Table 2, 2026LF columns). If even part of the discounted volume arrives on the submitters' schedule rather than PJM's, the auctions will have under-procured against it. That is a tightening mechanism, not a loosening one. It does not belong in the thesis as evidence, because it is a hypothesis about a rule's consequences rather than an observation, but it is the reason the honest read of this finding is "no signal" rather than "bearish signal."

The methodological lesson has the longest shelf life and generalizes past PJM. A forecast series the thesis depends on had a structural break inserted into it by a one-page guideline that was never a manual revision, never a tariff filing, and whose live URL has since 404'd; the manual caught up half a year later. Pre-2026 and post-2026 PJM vintages are not comparable instruments. Combined with the parent brief's finding that marker-1 is undecidable because it never named a weather-normalization basis, the power layer now has two separate instrument-definition failures, both found by reading primary sources rather than by any trigger firing. The cheap fix is a rule rather than a trigger: any anchor keyed to a published series must name the exact table and vintage, and must carry a documented check for methodology breaks in that series. That is a harness change rather than a thesis change, and it is probably worth more than the sixth trigger.

Why this is in the vault

This closes the first open follow-up of [[2026-07-28-pjm-peak-load-forecast-power-cycle-corroboration]] with a primary-source verdict, which does three specific things to the live $20,000 power-cycle-v1.1 position: it removes the 2027-2029 portion of the Jan-2026 trim from the thesis's evidence base as a scorekeeping artifact while preserving the 2026 slip as a real but small submitter-side signal, it supplies the exact published artifacts and threshold band (60-70% inclusion rate, Firm MW direction) for a proposed sixth founder-review trigger, and it corrects the unverified footnote-[2] claim in [[2026-05-17-power-cycle-v1]] by dating and sourcing it (guideline 2025-07-01, LAS presentation 2025-11-24, Manual 19 codification only at Rev 39 on 2026-06-30).

Open follow-ups

Related

Sources

Retrieved and parsed in this run (2026-10-10):

Probed and confirmed unavailable:

Vault (read, not re-verified):

Paywalled/skipped: none encountered.