Scottish/Welsh council tax ~20% above devolved budget figures; WAS owned_land imputation unstable across releases #448

Activity

1. vahid-ahmadi

   added a commit that references this issue on Jul 26, 2026

   Net Scottish Water charges out of FRS council tax (fixes ~25% Scottis…

   ...

   d99a245
2. MaxGhenis commented on Aug 23, 2026

   

   MaxGhenis

   on Aug 23, 2026

   Contributor

   More actions

   Re-measured on the live line at 2025 (enhanced_frs_2024_25.h5 @ v1.56.14, microcosm#733's parity reference; policyengine-uk 2.91.0 @ c93e1a05). The two issues separate: the council-tax overstatement is much smaller on the live 2024-25 line against current official comparators; under-simulated Council Tax Reduction accounts for a material Scottish residual, while its Welsh contribution is unquantified (no CTRS-cost comparator pinned); the owned_land instability is a sparsity problem, excluded from microcosm's input-mass validation but still inside an active land-value target.

   1. Council tax

   Model vs the nearest official 2025-26 figures (household level, 2025). The three comparators are different measures — a requirement, an amount billed net of CTR, and a collectable amount ex-CTRS — so read the gaps as diagnostics, not one series:

   nation	gross	CTR	net	official 2025-26	net gap
   England	£42.9bn	£2.0bn	£40.9bn	council tax requirement £44.1bn (MHCLG)	−7.2%
   Scotland	£3.80bn	£0.25bn	£3.56bn	net billed £3.389bn, after CTR and excluding water and sewerage (gov.scot)	+4.9%
   Wales	£2.62bn	£0.01bn	£2.61bn	collectable ex-CTRS £2.524bn (gov.wales)	+3.4%

   Against the incumbent's own OBR Table 4.1 targets at 2025 (England £44.1bn, Scotland £3.05bn, Wales £2.40bn, net of CTR) the same model runs −7% / +17% / +9%, so the comparator moves Scotland by 11.6 points; which series to calibrate to is a decision to make once. Average gross bills are £1,640 / £1,405 / £1,657, and I could not reproduce the issue's band-weighting hypothesis in a national band-share check (the artifact's shares match the published stock). Water is not the mechanism either: both builds net Scottish water and sewerage out of ctannual before imputing (uk-data frs.py:823-884; microcosm frs_council_tax.py:71-103); council_tax is reported where the FRS has it (about a quarter of households, per the build's comment) and a region × band × single-adult cell mean otherwise.

   CTR, on two levels:

   • claims_all_entitled_benefits.py:13-28 tests its seven reported-benefit columns with .sum() < 1 over the whole array, a simulation-wide scalar: on this artifact it is False for every benefit unit (0 of 61,223, though 51,855 report none of the seven by the formula's own test). would_claim_council_tax_reduction is then reported-CTR only (would_claim_council_tax_reduction.py:4-16), and the same gate sits under housing benefit and the legacy credits. The code comment describes a simulation-level switch, so the intended dataset behaviour needs deciding before it is "fixed".
   • For the reporters that remain, the simulated devolved schemes (simulated_council_tax_reduction_benunit.py:17-94) zero most of them: of 341 real Welsh reporter records routed to the Welsh formula, 102 fail on imputed household savings > £16k and 159 elsewhere in the formula (119 on the applicable-income taper, 40 with no eligible liability recorded); Scotland 98 + 161 of 451. The 80 Welsh records that keep an award carry a mean weight of 148 against 614 for the zeroed ones. Modelled CTR lands at £249m for Scotland (2024-25 outturn ≈ £389m; 2025-26 not yet pinned) and £10m for Wales (no CTRS-cost comparator pinned). England already runs below its requirement and has no pinned scheme-cost comparator; the measured CTR shortfall is Scotland's, so the net council-tax overstatement is a devolved-nation story, not a UK one.

   Targets: microcosm's surface binds obr.council_tax (UK, £50.9bn, council_tax_less_benefit); obr.council_tax_england/scotland/wales are declared but held. Not for lack of facts: chronicle carries the OBR EFO March 2026 Table 4.1 nation rows (FY2025-26 England £44.12bn / Scotland £3.05bn / Wales £2.40bn — the incumbent's exact values) at UK geography K02000001, while microcosm's reference generator pins nation geography ids onto those three ids, so they match nothing. Activating them is a generator pin change (plus the source-projection policy the UK id already carries), registered for adjudication in microcosm#736; a non-OBR basis (the collection statistics above) would be separate chronicle packages.

   2. owned_land

   nation	total (v1.56.14)	records with owned_land > 0	share held by the 10 largest records	share of nation household_land_value
   England	£326bn	348	80%	7%
   Scotland	£62bn	16	100%	22%
   Wales	£18bn	56	97%	9%
   NI	£15bn	336	68%	27%

   UK £421bn on v1.56.14 and £418bn on v1.56.16 (the stored draws are near-identical across the two releases: 16 of 52,846 household records differ). With sixteen Scottish records carrying the whole Scottish total, a re-draw can move nation aggregates sharply: microcosm#733 reports the adjacent-seed swing at 2024-25 as 71.4% national / 112.9% max-region (37.7% / 2.41× London per the register's earlier receipt, which #733 dates to 2023-24; not independently re-run here) and renewed the input-mass exclusion to 2026-09-20 (input_mass_reviewed_exclusions.json:13-19). Mechanics: the incumbent imputes owned_land (WAS DVLUKValR8_sum, imputations/wealth.py:34-65) as one of eleven QRF outputs and passes no seed (wealth.py:272-312, utils/qrf.py:28-40); the pinned microimpute 1.8.1 then falls back to a fixed default seed, so the release-to-release swing in this issue comes from changed inputs and row order rather than an unseeded draw. microcosm draws owned_land first in its regime-gated QRF chain, then property_wealth, with declared seed 0 and donor-support clipping only (was_wealth.py:78, :194-201, :350-364, :397-398, :481-496).

   The exclusion covers the input-mass gate only. household_land_value = property wealth × regional intensity + owned_land (household_land_value.py:14-56), and ons.land.household_land_value (£4.56tn) is an active target bound to that variable (uk_national_targets.json:9198-9217, target_reference_membership.json:7358-7369), so the draw would enter a calibration run on the surface as currently bound.

   Suggested next steps

   1. Decide the intended microsimulation behaviour of claims_all_entitled_benefits (per benefit unit vs the current simulation-wide switch), encode it explicitly, and re-measure CTR, housing benefit and the legacy credits after.
   2. CTR's capital gate uses imputed WAS savings: validate that proxy or add a reported-capital route; this is separate from the pension-wealth issue on UC caseload runs ~40% below administrative counts, capping aggregate reform estimates #452.
   3. Decide the nation-level council-tax basis once: the OBR Table 4.1 ids need only their geography pin re-pointed to bind (microcosm#736); the collection statistics (Scottish net billed and CTR, Welsh collectable, MHCLG requirement) would be new chronicle packages on a different definition.
   4. owned_land: set a nation-level stability and fidelity acceptance test, then compare the candidates (winsorised donor, a separate imputation outside the wealth chain, expected value instead of one draw) against it; add the regression test this issue asked for; and decide explicitly whether household_land_value stays active while owned_land is excluded (hold the reference, or bind property wealth × intensity only) — microcosm#736's registry is the place for that decision. The exclusion is a bridge; it expires 2026-09-20.

   Basis: council-tax comparators are 2025-26 statistics on three different definitions; CTR outturns are 2024-25 where 2025-26 is unpublished; all model numbers are on the incumbent's artifacts, aggregates only.