Stop discounting LTT and LBTT twice

[MaxGhenis]

Fixes #2246

Summary

expected_ltt and expected_lbtt multiplied Land Transaction Tax (Wales) and Land and Buildings Transaction Tax (Scotland) by a 4.5% property sale rate. The liabilities already charge only the year's purchases, so this discounted them a second time. This PR follows the convention #2235 sets for SDLT: the purchase draw is the expectation, so each expected value equals its liability. It also removes property_sale_rate, which no longer has any reader.

National effect on certified data, 2026-27:

• gov_tax, household_tax, pre_budget_change_household_tax and gov_balance rise by £0.435bn if Stop charging property stocks as purchases when a household buys its home #2241 has landed. Without Stop charging property stocks as purchases when a household buys its home #2241 the rise is £2.273bn: £1.838bn of that is LBTT on property stocks that Stop charging property stocks as purchases when a household buys its home #2241 stops charging as purchases.
• Household net income falls by the same amount.
• Poverty rates don't change.
• Reforms to LTT or LBTT rates are now scored at 22 times their old value (1/0.045).

All of this is measured by real runs, below. Stacked on #2235; land after #2235 and after #2241 (see "Merge order").

The bug

• expected_ltt = property_sale_rate × land_transaction_tax and expected_lbtt = property_sale_rate × land_and_buildings_transaction_tax, with gov.hmrc.stamp_duty.property_sale_rate = 0.045. That parameter was a 2021 placeholder ("TODO add proper calculations").
• The liabilities are already annual. They charge the prices of properties bought in the year, plus the tax on each year's addition to cumulative rent:
  • main_residential_property_purchased = main_residence_value × property_purchased.
  • policyengine-uk-data draws property_purchased at a seeded 3.85% (datasets/frs.py:1874-1878 at tag 1.58.0: PROPERTY_PURCHASE_RATE = 0.0385, default_rng(0)).
  • On Count corporate SDLT once and let SDLT abolition remove all of it #2235's base, the additional and non-residential purchase prices are also set only on that draw. After Stop charging property stocks as purchases when a household buys its home #2241 they are inputs.
• So the 4.5% applied the purchase probability a second time. On 1.58.0 in 2026-27, LBTT is £2.310bn but expected_lbtt was £0.104bn. LTT is £0.070bn, and expected_ltt was £0.003bn.
• The totals sum the expected values. So LTT and LBTT entered gov_tax, household_tax and pre_budget_change_household_tax at 4.5% of their value, and so did any LTT or LBTT reform.

The fix

• expected_ltt adds land_transaction_tax and expected_lbtt adds land_and_buildings_transaction_tax, with no scaling. These are pure adds, the same convention expected_sdlt follows in Count corporate SDLT once and let SDLT abolition remove all of it #2235.
• gov.hmrc.stamp_duty.property_sale_rate is removed.
  • After Count corporate SDLT once and let SDLT abolition remove all of it #2235 its only readers were these two formulas.
  • An org-wide code search finds no other reader. The only other hits are a notebook under old/ in structural-reforms-demo, and policyengine-uk-chat's eval snapshot synced from policyengine-uk 2.90.2, which drops the case on its next resync.
  • A reform that sets it now fails with ValueError: Could not find the parameter gov.hmrc.stamp_duty.property_sale_rate (checked), where before it silently did nothing.
  • The seven remaining property_sale_rate: 0 overrides in the high value council tax surcharge tests are removed. Count corporate SDLT once and let SDLT abolition remove all of it #2235 removed the other five.
• fiscal_partition.yaml (from Count corporate SDLT once and let SDLT abolition remove all of it #2235):
  • The data_notes on heads land_transaction_tax and land_and_buildings_transaction_tax now give the measured 1.58.0 figures instead of describing the double discount.
  • The exclusion notes on the two liabilities now say "equal to the head; report the head".

Invariants

1. Expected is realised. expected_ltt is exactly land_transaction_tax, and expected_lbtt is exactly land_and_buildings_transaction_tax, for every household. So they also agree in weighted total.
2. Geography and sign. LTT is zero outside Wales and LBTT zero outside Scotland, and both are non-negative.
3. Once, at full value. Let Z set every LTT and LBTT rate to zero, including first-time buyer, additional dwelling, non-residential and rent rates. For each total, T(baseline) − T(Z) equals the liabilities land_transaction_tax + land_and_buildings_transaction_tax, independently of 1. The old scaling gave 0.045 × that; counting a head twice would give 2 ×.
4. Zeroing removes them and nothing else. Both expected values are 0 under Z and SDLT is unchanged, so with 3 nothing else in the totals moves.
5. Statute, differential. The model's expected LBTT and LTT equal an independent marginal-band calculation from the statutory tables, wherever the model's schedule is up to date. The schedules that are stale are recorded as strict expected failures.

#2235's partition invariants (heads = lists, traced overlap guard, coefficient one, identity) still hold.

Tests

File	Covers
tests/policy/baseline/gov/revenue_scotland/expected_lbtt.yaml (9 cases)	Hand-computed LBTT for purchases in Scotland, 2026; expected_lbtt = liability = household_tax = gov_tax. Cases: £400k, £1m, band tops £145k and £250k, a first-time buyer at £200k, non-residential £300k, plus zero outside Scotland.
tests/policy/baseline/gov/wra/land_transaction_tax/expected_ltt.yaml (8 cases)	Hand-computed LTT for purchases in Wales; same identities. Cases: main residence £300k and £2m and the band top £180k (2022), an additional dwelling at £200k (2022), non-residential £1.2m (2026), plus zero outside Wales.
tests/test_ltt_lbtt_expected_properties.py	Invariants 1–4 (Hypothesis: 12 examples of 3–5 households, all four nations, 2022–2030, purchases, rents, cumulative rents, earnings), a check that the zero-rate reform covers every schedule, and invariant 5: 6 statutory cases plus 3 strict xfails.

Rate bands cited (read on legislation.gov.uk 2026-10-11):

• LBTT: SSI 2015/126 schedule.
  • Table A residential: 0% to £145k, 2% to £250k, 5% to £325k, 10% to £750k, 12% above. In force again from 1 April 2021; SSI 2020/215 art. 2's temporary nil band ran "until the end of the day of 31 March 2021".
  • Table B non-residential, as substituted by SSI 2018/372: 0% to £150k, 1% to £250k, 5% above.
  • First-time buyer relief takes the first £175,000 out of charge (LBTT(S)A 2013 sch. 4A para. 3).
• LTT: WSI 2018/128 schedule.
  • Table 1 main rates as at 1 January 2022: 0% to £180k, 3.5% to £250k, 5% to £400k, 7.5% to £750k, 10% to £1.5m, 12% above.
  • Table 2 higher rates, as substituted by WSI 2020/1618: 4%, 7.5%, 9%, 11.5%, 14%, 16% on the same thresholds from £0.
  • Table 3 non-residential, also from WSI 2020/1618 and still in force: 0% to £225k, 1% to £250k, 5% to £1m, 6% above.
• The model samples parameters at 30 April (utils/parameters.py:113).
• Why the Welsh residential cases are in 2022. The model's main residential schedule still has the pre-10 October 2022 bands (Welsh LTT residential rates stale (missing 10 October 2022 main and 11 December 2024 higher-rate changes) #1796), and its LBTT additional dwelling supplement is 6% rather than 8% from 5 December 2024 (Scottish LBTT Additional Dwelling Supplement stale at 6% (should be 8% from 5 December 2024) #1795). So the residential LTT cases are in 2022, when the model matches the law. The 2026 schedules are strict xfails citing those issues: WSI 2022/1027 Table 1, WSI 2024/1311 Table 2, and the LBTT(S)A 2013 sch. 2A para. 4(2) supplement as amended by SSI 2024/367. Fixing either issue turns its xfail into a failure, which prompts moving that case into the statutory set.

Cross-check against Axiom. The test's statutory tables reproduce every value in rulespec-uk's companion tests, computed:

• uk/policies/govuk/lbtt.test.yaml: £200k → 1,100; £300k → 4,600; £500k → 23,350; £800k → 54,350; additional £300k → 4,600 + 24,000.
• uk/policies/govuk/ltt.test.yaml, current main rates: £300k → 4,500; £500k → 18,000; £800k → 41,750; £2m → 171,750. Higher rates: £300k → 19,950.

Red on base, green here. Run against #2235's head (41e970b):

• 9 of the 17 new YAML cases fail: every case with a non-zero tax.
• The property test fails at invariant 1: expected_lbtt 6,373.53 vs liability 141,633.98.
• With invariant 1 deleted, invariant 3 alone still fails ("household_tax: LTT and LBTT not counted once at full value").

On this branch all pass.

Local runs (Python 3.13, policyengine-core 3.33.0):

• The 17 new YAML cases: 17 passed.
• The 9 YAML files that test expected_lbtt, expected_ltt, the LTT/LBTT liabilities, gov_tax or household_tax: 60 passed.
• Stamp duty, WRA and surcharge YAML: 41 passed.
• test_ltt_lbtt_expected_properties.py: 8 passed, 3 xfailed.
• test_fiscal_partition.py and test_sdlt_counted_once_properties.py: 40 passed.

CI. The PR workflow runs only on PRs into main, so for this stacked PR I dispatched it on the branch with workflow_dispatch. It runs again when the PR is retargeted to main.

National impact (real runs)

Setup.

• Data: policyengine-uk-data 1.58.0, enhanced_frs_2024_25.h5, sha256 5ba399ea5ab9a6c7d4e9aaf6c623080beb34d9ad5f330f6244dbad72b8d960b6, the certified release.
• Code, four versions, each loaded from its own checkout (the script records the path): base = Count corporate SDLT once and let SDLT abolition remove all of it #2235 head 41e970b; head = this PR; and each merged with Stop charging property stocks as purchases when a household buys its home #2241 (904b991).
• Runtime: policyengine-core 3.33.0, policyengine-uk 2.126.8.
• 48 scenario-years: 4 versions × 2 scenarios × 6 years. Every figure is a weighted total from Microsimulation, with no scaling or interpolation.

Current law, £bn: on #2235 alone

	2025-26	2026-27	2027-28	2028-29	2029-30	2030-31
expected_ltt, base → head	0.003 → 0.065	0.003 → 0.070	0.003 → 0.076	0.004 → 0.082	0.004 → 0.088	0.004 → 0.095
expected_lbtt, base → head	0.099 → 2.211	0.104 → 2.310	0.109 → 2.424	0.114 → 2.539	0.119 → 2.654	0.125 → 2.780
Change in gov_tax, household_tax, pre_budget_change_household_tax, gov_balance	+2.174	+2.273	+2.387	+2.503	+2.619	+2.746
Change in household_net_income	−2.174	−2.273	−2.387	−2.503	−2.619	−2.746

Current law, £bn: with #2241 merged into both

	2025-26	2026-27	2027-28	2028-29	2029-30	2030-31
expected_ltt, base → head	0.003 → 0.065	0.003 → 0.070	0.003 → 0.076	0.004 → 0.082	0.004 → 0.088	0.004 → 0.095
expected_lbtt, base → head	0.017 → 0.367	0.017 → 0.385	0.018 → 0.405	0.019 → 0.427	0.020 → 0.448	0.021 → 0.471
Change in gov_tax, household_tax, pre_budget_change_household_tax, gov_balance	+0.413	+0.435	+0.460	+0.486	+0.512	+0.541
Change in household_net_income	−0.413	−0.435	−0.460	−0.486	−0.512	−0.541

In every run, land_transaction_tax, land_and_buildings_transaction_tax, expected_sdlt and hbai_household_net_income are unchanged by this PR.

Where the LBTT comes from, 2026-27.

• On Count corporate SDLT once and let SDLT abolition remove all of it #2235 alone, LBTT is all on residential purchases: £2.310bn. That includes the residential rates and the 6% supplement on £15.171bn of Scottish other-residential stock, which the purchase flag charges as additional-dwelling purchases (additional_residential_property_purchased treats a household's whole other-property stock as an additional purchase #2238).
• With Stop charging property stocks as purchases when a household buys its home #2241 it is £0.385bn, on £7.642bn of main-residence purchases.
• Non-residential LBTT is £0: four records, one household repeated, buy £118,171 of non-residential property, under the £150,000 nil band.
• No Scottish or Welsh record has an LBTT or LTT rent charge in any year.
• Welsh LTT is all on main-residence purchases: £3.121bn of prices, 120 purchaser records, 36 paying.

Reform scoring. An illustrative reform raises the top residential band by 1pp in both countries: LBTT over £750k from 12% to 13%, and LTT main rates over £1.5m from 12% to 13%. The change in gov_tax from current law, £bn:

Code	2025-26	2026-27	2027-28	2028-29	2029-30	2030-31
base (#2235)	+0.0013	+0.0015	+0.0018	+0.0021	+0.0023	+0.0026
head	+0.0282	+0.0335	+0.0396	+0.0457	+0.0517	+0.0584
base + #2241	+0.0006	+0.0007	+0.0008	+0.0008	+0.0009	+0.0010
head + #2241	+0.0143	+0.0156	+0.0171	+0.0185	+0.0200	+0.0216

Accounting checks (every year 2025-26 to 2030-31, current law):

Check	base	head	base + #2241	head + #2241
max per-household |expected_ltt − land_transaction_tax|	£25,277	£0	£25,277	£0
max per-household |expected_lbtt − land_and_buildings_transaction_tax|	£127,212	£0	£112,548	£0
households with ≥2 of SDLT, LTT, LBTT	0	0	0	0
max per-household |gov_tax − sum of its list| (float32)	£0.12	£0.12	£0.12	£0.12
non-finite outputs	none	none	none	none

Distribution. Head − base in household net income, weighted mean £ per household, 2026-27:

	D1	D2	D3	D4	D5	D6	D7	D8	D9	D10	All
#2235 alone	−8.49	−0.50	−1.33	0.00	−54.32	−1.10	−5.69	−80.25	−278.20	−305.07	−72.48
with #2241	−8.49	−0.50	−0.98	0.00	−54.32	−0.94	−5.69	−69.89	−2.97	−4.68	−13.86

The payers are few and the deciles are lumpy.

• With Stop charging property stocks as purchases when a household buys its home #2241: 11.3k Scottish households (42 records) pay £32,454 more each, and 7.4k Welsh households (36 records) pay £9,039 more each.
• On Count corporate SDLT once and let SDLT abolition remove all of it #2235 alone: 32.0k Scottish households pay £68,934 more each, with the higher-decile stock charges. The Welsh figure is unchanged.
• No household gains.

Poverty. BHC and AHC, absolute and relative, are unchanged to four decimal places in every year and every pair. hbai_household_net_income does not move.

Evidence and scripts: ~/reviews/uk-ltt-lbtt-2246/ (impact.py, compare.py, impact.md, runs/, shas.txt).

Merge order

Land after #2235 (this PR's base) and after #2241.

Without #2241, this PR moves the £1.838bn a year of LBTT that #2241 removes from 4.5% into the totals at full value, until #2241 lands. That LBTT is on Scottish property stocks charged as purchases (#2238). Landing #2241 first means the totals only ever gain the £0.435bn on real main-residence purchases.

If #2241 lands first, the LBTT data_note stays accurate: it gives both figures. When #2235 merges, GitHub retargets this PR to main and CI runs on it.

Not in this PR

• Stale schedules, recorded as strict xfails:
  • Welsh main and higher rates (Welsh LTT residential rates stale (missing 10 October 2022 main and 11 December 2024 higher-rate changes) #1796). These affect the £0.070bn of LTT on 1.58.0, which is all main-residence purchases.
  • The LBTT supplement at 6% rather than 8% (Scottish LBTT Additional Dwelling Supplement stale at 6% (should be 8% from 5 December 2024) #1795). After Stop charging property stocks as purchases when a household buys its home #2241 no 1.58.0 household has an additional-dwelling purchase, so this has no effect on that data.
• LBTT lease rates. lbtt/rent.yaml charges 5% above £2m. SSI 2020/24 art. 2 made that band 2%. The model also applies the lease table to residential rent. On 1.58.0 no household has a rent charge, so the impact is zero there. Split into its own task.
• Component variables are not country-gated. ltt_on_residential_property_transactions and lbtt_on_residential_property_transactions are computed for every UK purchaser: £22.4bn and £29.2bn nationally in 2026-27. Only land_transaction_tax and land_and_buildings_transaction_tax multiply by ltt_liable / lbtt_liable. The totals are right, and the partition lists the components as exclusions, but don't sum the components.
• Purchase data. 1.58.0 draws its purchasers UK-wide at 3.85%, which gives 0.027m Welsh and 0.065m Scottish purchasers a year. Imputing purchases by nation belongs in policyengine-uk-data, with the d833 batch. See the benchmark below.

Against official receipts

These figures were read from the official pages on 2026-10-11. The model figures are 2025-26 from the runs above; the outturn is 2024-25.

	Model, expected_lbtt / expected_ltt before → after this PR (with #2241)	Outturn
LBTT on residential conveyances, excluding ADS	£0.017bn → £0.367bn	£484.6m
LBTT ADS (net)	£0 (no additional-dwelling purchases on 1.58.0 after #2241)	£204.3m
LBTT on non-residential conveyances and leases	£0	£197.2m and £34.3m
Scottish residential purchases	0.065m households	102,100 residential returns
LTT, residential	£0.003bn → £0.065bn a year	£67.2m in January–March 2025 alone

Sources:

• Outturn LBTT: Revenue Scotland, Annual LBTT Statistics 2024-25, declared due, revised 24 July 2025.
• Outturn LTT: WRA, LTT statistics January to March 2025. WRA publishes quarterly; I found no 2024-25 annual total.

After this PR, the model's main-residence LBTT is about three quarters of Revenue Scotland's residential (non-ADS) receipts; before it, it was about 3.5%. Welsh LTT stays well below outturn, because the data has few Welsh purchasers. The remaining gaps are the purchase data: ADS, non-residential and lease transactions, and the purchaser count. They are not this PR's convention.

axiom: n/a: microsim-only. This changes the expected-value convention by which the model's tax totals count LTT and LBTT. The statutory liabilities are unchanged; rulespec-uk encodes them in uk/policies/govuk/ltt.yaml and lbtt.yaml, and this PR's statutory tables reproduce their companion tests.

🤖 Generated with Claude Code