[WesIngwersen]

Part of #1270 (to-do 4) and #1194 (medicinal ingredients). Stacked on #1292. It changes Step 5 inputs, so it reaches the tables only in the next rebuild.

What was wrong

The export column (F04000) is Census domestic exports, mapped to BEA commodities (#762), and nothing conditions it. Imports, by contrast, are tied to BEA's summary totals (#785), and the computers family's imports are anchored on BEA's 2017 cells (#763).

For computers (3341), Census domestic exports run well above BEA's. In 2017, $bn:

Commodity	BEA published exports	Census domestic exports	Output
334111 computers	7.14	7.89	8.32
334112 storage devices	2.22	2.60	4.16
334118 terminals and peripherals	6.06	9.22	8.14
family	15.42	19.71 (1.28×)

Terminals export more than they produce: 1.01 times output in 2022 and 1.11 in 2024. A row that exports more than it produces leaves its imports nowhere to go. Total use excluding exports falls below imports, so q = L_dom y fails for terminals and the #1282 guard flags the row (2024: imports $0.76bn above use).

It isn't re-exports. Terminals are heavily re-exported, $16.4bn in 2017 against $19.9bn of Census imports, but #762 already leaves re-exports out. The excess is in the domestic exports themselves. On the import side, re-exports don't distort the family either. Its 2017 imports are BEA's published cells (#763). After that they move with Census import growth, which is almost the same with or without re-exports in 2024 (1.50 against 1.44 times 2017).

The change

bedrock/transform/iot/nowcast_export_anchoring.py, used in derive_initial_Y_pur: each 3341 commodity's export cell is its published 2017 F04000 cell moved by its own Census growth. That is the #763 construction for imports, applied to exports:

X_t = X_pub,2017 × (X_census,t / X_census,2017)

The same edge rules as for imports apply. A commodity with no published 2017 value keeps Census. A commodity with no Census 2017 value holds its published value. Every commodity outside ANCHORED_EXPORT_FAMILIES keeps Census.

Effect on the export column (before the balance)

Terminals' exports as a share of output:

	2017	2018	2019	2020	2021	2022	2023	2024
Census	1.13	1.16	1.22	1.21	1.25	1.40	1.36	1.52
anchored	0.74	0.76	0.80	0.80	0.82	0.92	0.89	1.00

Computers and storage devices stay below output in every year (0.70-0.92 and 0.32-0.84).

⚠️ 2024 terminals is still on the edge, $6.85bn of exports against $6.86bn of output. Census exports grow 13% from 2017 while terminals' output falls from $8.1bn to $6.9bn. Whether that fall is real is a question for the output side; this PR doesn't touch it.

Pharmaceuticals: anchored on the family's growth

Mo found the same gap for pharmaceuticals on #1286. Census puts pharmaceutical preparations onto medicinal ingredients (325411) and diagnostics (325413). In 2017, $bn:

Commodity	BEA published exports	Census domestic exports
325411 medicinal ingredients	0.08	7.30
325412 preparations	30.13	16.69
325413 diagnostics	0.12	6.57
325414 biologics	22.57	19.72
family	52.90	50.28

Medicinal ingredients exporting $14.0bn in 2024, more than its $13.2bn of output, is what drove its Phi to 0.04 (#1194).

Here a row's own Census line is no mover, because the rows are misclassified. With own-row growth, biologics would export $52.7bn in 2024 against $48.0bn of output. So FAMILY_GROWTH_EXPORT_FAMILIES = {'3254'} holds BEA's 2017 split of the family and moves the whole family by its Census growth (1.81 times 2017 by 2024):

X_t = X_pub,2017 × (Σ_family X_census,t / Σ_family X_census,2017)

All four rows stay below output in every year 2017-2024. Biologics peaks at 0.87, medicinal ingredients and diagnostics are at 0.01, and preparations at 0.35. Computers keep own-row growth (ANCHORED_EXPORT_FAMILIES = {'3341'}), because their rows' own Census lines are sound. anchored_exports applies both modes, and every row outside the two families keeps Census.

Not added: automobiles (336111). Its export input is already right ($23.7bn against $23.5bn published), and the pair is already split by BEA's 2017 export mix. The doubling happens in the Step 5 balance (#1296).

Also in this PR

• rebuild_nowcast.yml runs on ubuntu-24.04, as ci: pin ubuntu-24.04 so setup-python can resolve 3.11.13 #1297 did for CI. On ubuntu-latest the rebuild fails at "Set up Python".
• The rebuild script's comment says 2017 anchors exports as well as imports.

Verified on a CI build of 2022 and 2024

Run 43303427722 built Steps 5-7 from this stack. Nothing was uploaded. On the balanced tables, after redefinitions:

Terminals 334118	8365fa4	this stack
2022 exports / output	1.01	0.67
2024 exports / output	1.11	0.73
2024 imports above use	+$0.76bn	none

q = L_dom y passes in both years. (The same build also carried a car import split, which didn't work and isn't in this PR. See #1296.)

Checks

black, ruff, mypy, and the unit tests in test_nowcast_export_anchoring.py (6 passed): both modes, 2017 reproducing BEA, rows outside the families, and the edge rules. The effect on the balanced tables needs a rebuild.

🤖 Generated with Claude Code

[bl-young]

The mechanism is right and mirrors #763. I checked the export ratios against the 8365fa4 trade FBS. 334118 exports/output goes 1.13 -> 0.74 (2017, equal to BEA), 1.40 -> 0.92 (2022), 1.36 -> 0.89 (2023) and 1.52 -> 1.00 (2024). 334111 and 334112 stay below 0.9. Every rebuild year job already builds Trade_Exports_2017 via build_trade({2017, year}), so no workflow change is needed.

Notes:

1. 2024 terminals land at exactly 1.00, so the 334118 xfail and the #1282 warning likely survive the next rebuild. Worth an issue for the output side (terminals' output falling from $8.1bn to $6.9bn).

2. Mo's other export findings on #1286 (325411 pharma, 336111 doubling) look like the same Census-vs-BEA gap. Add them to ANCHORED_EXPORT_FAMILIES here, or file them separately?

3. rebuild_nowcast.yml uses the same setup-python on ubuntu-latest, so it will hit the Python 3.11.13 failure that's breaking CI here. Fix that before dispatching the rebuild that picks this up.

4. Nit: the comment in scripts/rebuild_nowcast_year.py says "2017 is the import anchor"; it's the export anchor now too.

[WesIngwersen]

Thanks @bl-young. Taking your notes in order, in 1331e58b:

1. Terminals' 2024 output: no issue filed, because the fall is BEA's own. BEA's gross output for industry 334118 goes from $8.76bn in 2017 to $7.43bn in 2024 (−15%), and our commodity output follows it ($8.14bn to $6.86bn, −16%). The 2024 tie at 1.00 is Census exports growing 13% against that. Anchoring on the family's growth instead would make it worse: the family grows 1.42 times, so terminals would reach $8.6bn. So own-row growth stays for computers. If the 2024 row still warns after the rebuild, the balance has $0.01bn to move.
2. Pharmaceuticals: added, with a family-growth mode. Own-row growth would have put biologics (325414) at $52.7bn of 2024 exports against $48.0bn of output, because the Census rows are misclassified. So 3254 holds BEA's 2017 split and moves by the family's growth (1.81 by 2024). All four rows stay below output in 2017-2024. Medicinal ingredients goes from $14.0bn of 2024 exports to $0.15bn, which is the Nowcast v1.0 margins: producer value collapses or turns negative while margins don't (Phi 0.011 for 336213, 0.143 for 325411) #1194 Phi case.
   Automobiles: not added. Its export input matches BEA ($23.7bn against $23.5bn), and the pair is already split by BEA's 2017 mix. The doubling is the Step 5 balance (Make cars (336111) and light trucks (336112) mirror each other until they are merged #1296), which the car import split I tried couldn't fix either.
3. rebuild_nowcast.yml: pinned to ubuntu-24.04. ci: pin ubuntu-24.04 so setup-python can resolve 3.11.13 #1297 didn't cover it.
4. The nit: the comment now says 2017 is the import and export anchor.

🤖 Generated with Claude Code