Skip to content

Commercial buyers' kWh follow their bills, which move on the carry (bills split to #1313) #899

Closed

Description

@WesIngwersen

Split 2026-11-13. This issue now covers commercial buyers' kWh, which #1305 completes. Seeding the commercial bills on an annual source is beyond Nowcasting Phase 3's scope, so it moved to #1313. The original text is below for the record; its task and acceptance criteria now belong to #1313.

What #1305 does (merged 85258d8, flag commercial_kwh_follow_intermediate_inputs, off until the combined re-pin): each commercial buyer's kWh share follows its 2017 bill, moved on SAS Purchased electricity through 2022 and then on its other intermediate inputs, or on those inputs throughout. Bills, the class's EIA kWh, and other classes don't change. With #1306's AIES hold, every commercial group's delivery margin sits between 8.8 and 14.2 ¢/kWh, and the median N_d volatility for commercial commodities falls from 0.106 to 0.100.


Revised 2026-11-06 for the plan in #902. This no longer blocks the electricity work. The plan keeps bills as they are and puts delivery in a new 42ELEC margin sector, so a commercial seed is an improvement to the bills, not a prerequisite. It still matters: commercial bills run about 1.15× EIA revenue in 2024, and that excess now lands in commercial buyers' delivery margins, overstating them. It no longer inflates their emissions, because emissions follow kWh. Priority is therefore lower than when this was filed, and it's a follow-up to #1217 rather than an input to it.


Phase 1 of Electricity G/T/D v3. Depends on #896.

Manufacturing electricity is at least indexed on Census CSTELEC through nonmaterial_seed. The commercial band has no electricity seed at all and moves on the carry — and it is the larger half of the row and the worse-behaved one.

Evidence

The commercial band is where the row misbehaves. Against EIA-861 published retail revenue, the commercial gap runs +31.3% at 2017, balloons to +67.5% in 2021, and falls back to +22.3% by 2024. Residential, which is EIA-anchored through PCE, matches to +0.04% at 2017 and stays within a few percent throughout. The industrial band, which at least has the CSTELEC index, moves less than commercial in every year.

When the electricity row lost a fifth of its share of intermediate use across 2023-24 (#896), the collapse was broad across unrelated commercial purchasers — food and beverage stores −$3.06bn, limited-service restaurants −$3.01bn, hospitals −$2.39bn, general merchandise −$2.32bn, enterprise management −$2.07bn, hotels −$1.98bn. That spread is the signature of a row-wide rescale absorbed by whichever columns are free to move — which is exactly what an unseeded band is.

The task

Find an annual dollar source for commercial-sector electricity expense and seed the band, so the larger half of the row responds to something observed rather than to whatever the balance leaves it.

Candidates worth checking: SAS / AIES operating-expense electricity cells (EXPS_ELEC_VAL), which already serve services and transportation elsewhere in the nowcast; the AWTS/ARTS Business Expenses Supplement, which is what BEA itself uses for wholesale, retail and accommodation at the benchmark.

⚠️ CBECS cannot do this job. It classifies by principal building activity and its 2018 microdata codebook contains zero NAICS references. A building is not an establishment and a multi-tenant office cannot be resolved to NAICS by any mapping. CBECS prices the commercial class; it does not distribute one. See #895.

Acceptance

Activity

  1. WesIngwersen commented on Sep 24, 2026

    @WesIngwersen
    MemberAuthor

    Assigning @jvendries, to follow the reference-series fix rather than precede it.

    ⚠️ This issue's headline evidence is computed against a stale series. The commercial gap quoted above (+31.3% at 2017, +67.5% in 2021, +22.3% by 2024) is measured against EIA_861_REVENUE_BN, a hand-transcribed literal in electricity_row_control.py. Comparing it to the EPA Table 2.3 revenue data that EIA_ElectricPowerAnnual already extracts (filtered to Total Electric Industry), commercial disagrees in exactly the years in scope:

    year sector dict extracted
    2022 Commercial 173.42 172.60
    2023 Commercial 178.16 177.34
    2024 all four — exact match

    2024 matches to the cent and 2022/2023 do not, which reads as the literal having been transcribed from an earlier EPA edition since revised.

    ✅ The conclusion here survives — the error is ~0.5% against gaps of 22-67%, so the commercial band really is unseeded and really does misbehave. But the baseline should be corrected before a seed is built to close a gap measured against it, and the numbers in this issue will move slightly.

    Suggested order: (1) read the series from the existing EIA_ElectricPowerAnnual FBA instead of the literal — see the review on #992, where the same defect affects the #990 verdict; (2) recompute the commercial gap; (3) build the seed.

  2. WesIngwersen commented on Sep 24, 2026

    @WesIngwersen
    MemberAuthor

    @jvendries on a related note I just reassigned #895 to me which might be relevant here, especially CBECS. I took it over because as I'm working on improved the fuel use allocation it became necessary to get that MECS expenditure data.

  3. WesIngwersen commented on Sep 24, 2026

    @WesIngwersen
    MemberAuthor

    @jvendries Please review #995 first because it finds some suppressed electricity use data small for electricity but something that should be merged first for this to build on

  4. jvendries commented on Sep 24, 2026

    @jvendries
    Contributor

    Reproducing comment from analysis in PR #1005 below

    Summary (#899 analysis)

    Wholesale and retail buy electricity, but the nowcast still moves those cells on the 2017 carry — there is no survey seed for trade electricity.

    This PR adds a candidate seed and a grader, not a production change:

    1. Seed builder — takes 2017 trade×electricity dollars and moves them with an annual index:
      • A (preferred): one national commercial electricity revenue growth factor from EPA Table 2.3, applied uniformly to every wholesale/retail industry
      • B (backup): each industry’s own QCEW payroll growth
    2. Grade CLI — checks Slack bars (uniform moves, no weird outliers, electricity vs payroll consistency after column control)
    3. Write-up — About_trade_electricity_seed.md

    It does not wire the seed into composed_seed.

    Why only wholesale ∪ retail (not all of EIA “commercial”)

    EIA/EPA Table 2.3 commercial is a broad end-use class (offices, services, etc.), not a BEA industry band. This issue (#899) is scoped to the trade hole only: BEA wholesale and retail detail columns (WHOLESALE∪RETAIL), which today have no electricity seed and ride the carry.

    Other “commercial-looking” IO sectors are out of scope here on purpose:

    So Candidate A borrows the Table 2.3 commercial revenue index as a uniform mover for trade cells, but it does not claim to seed the full EIA commercial class. physical_proxy_cv is the within-trade cross-check against that broader proxy.

    Grade conclusion (2017–24, MUT v0.3.0_4276083)

    Neither candidate clears the bars → do not wire; leave trade electricity on carry.

    Candidate Result
    A (uniform EPA commercial) Uniformity OK by design; fails only physical_proxy_cv (~0.79 vs ≤0.35)
    B (QCEW payroll) Fails uniformity, idiosyncratic outliers, and CV

    A failed CV does not mean Table 2.3 is wrong — EPA “commercial” is broader than wholesale∪retail. B is not a viable fallback.

    Reproduce:

    python -m bedrock.analysis.nowcasting.trade_electricity_seed_grade \
        --csv --check --mut-vintage v0.3.0_4276083
  5. WesIngwersen commented on Sep 28, 2026

    @WesIngwersen
    MemberAuthor

    Added to the Nowcast US IOT Backlog (project 28) in the v0.5 scope cut. #1005 was closed: under theta = 1 the trade cells already move on price. What remains is pricing the electricity carry on EIA retail prices by customer class instead of BEA's industry price index, which runs about 17% above EIA's commercial price in 2021-22. That applies to every unseeded buyer, not only trade.

  6. WesIngwersen commented on Nov 6, 2026

    @WesIngwersen
    MemberAuthor

    @jvendries for your awareness: the revised plan in #902 keeps bills unchanged and puts delivery in a new 42ELEC margin sector, so this seed doesn't block that work. It would still improve the commercial bills, which run about 1.15× EIA revenue in 2024, and which set the level of the margins. Your call on whether to continue. It stays on the board.

  7. changed the title [-]Commercial band has no electricity seed and moves on the carry[/-] [+]Commercial band has no electricity seed and moves on the carry (follow-up to #1217, no longer a blocker)[/+] on Nov 6, 2026
  8. WesIngwersen commented on Nov 13, 2026

    @WesIngwersen
    MemberAuthor

    #1305 (merged 85258d8) splits commercial kWh on survey electricity through 2022, and on other intermediate inputs otherwise, with the flag commercial_kwh_follow_intermediate_inputs (off until the combined re-pin). That steadies buyers' generation and delivery shares, but it doesn't seed the commercial bills. That part, the subject of this issue, is still open, as Ben noted on #1305.

    🤖 Generated with Claude Code

  9. changed the title [-]Commercial band has no electricity seed and moves on the carry (follow-up to #1217, no longer a blocker)[/-] [+]Commercial buyers' kWh follow their bills, which move on the carry (bills split to #1313)[/+] on Nov 13, 2026
  10. WesIngwersen commented on Nov 13, 2026

    @WesIngwersen
    MemberAuthor

    Complete for the kWh split via #1305. The commercial bill seed is split to #1313, on the Nowcast US IOT Backlog.

    🤖 Generated with Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions