Add LA-level council tax band-count targets and calibrate on them

[vahid-ahmadi]

Summary

Adds 350 LA-level ons/council_tax_band_d/{code} targets (Band D amount per billing authority) and 2,541 voa/council_tax/{code}/{band} targets (dwellings per band A-H per LA) built from four public sources, covering all 296 English + 22 Welsh + 32 Scottish LAs in local_authorities_2021.csv. Wires the eight band-count targets into datasets/local_areas/local_authorities/loss.py so the LA reweighter trains on them. Follows the pattern of the regional land-value targets, one geography deeper.

What this PR does

Four public sources, one canonical CSV

storage/la_council_tax.csv (360 rows, 31 KB) joins:

Column	Source	Coverage
band_d_amount (England)	MHCLG Council Tax levels set by local authorities in England 2026-27, Table 10, column 17	296/296 English LAs
band_d_amount (Wales)	Welsh Government Council Tax levels April 2026 to March 2027, Table 1 "Overall average band D"	22/22 Welsh LAs
band_d_amount (Scotland)	Scottish Government Council Tax Assumptions 2025, "CT by Band 2025-26" Band D column	32/32 Scottish LAs
band_A … band_H + total_dwellings	VOA Council Tax: Stock of Properties, 2025, CTSOP1.0 summary table	295 English + 22 Welsh LAs

Column 17 in DLUHC Table 10 is the right one: "Average (Band D 2 adult equivalent) council tax for area of the billing authority including both local and major precepts" — i.e. the full amount households pay, inclusive of county, police, fire and parish precepts.

Code joins reconciled

• Post-2023 South Yorkshire E-codes (E08000038 Barnsley, E08000039 Sheffield) are remapped to the pre-2023 codes used in the reference LA list (E08000016, E08000019).
• Scottish source name quirks normalised: Argyll & Bute → Argyll and Bute, Dumfries & Galloway → Dumfries and Galloway, Perth & Kinross → Perth and Kinross, Shetland Islands (double space) → Shetland Islands, Edinburgh, City of → City of Edinburgh.

New module

targets/sources/la_council_tax.py:

• get_targets() returns Target objects at geographic_level=LOCAL_AUTHORITY.
• Band D targets tagged with year 2026 (England, Wales) or 2025 (Scotland) based on the source's latest release.
• Per-country reference URLs so downstream consumers can audit each value to its publication.
• Missing values (NI has no council tax, Scottish band counts) are filtered out of values rather than emitted as NaN so calibrators cleanly skip them.

Wires the band counts into LA calibration

datasets/local_areas/local_authorities/loss.py now adds eight voa/council_tax/{A..H} columns to the LA target matrix:

• matrix entry: per-household indicator 1[council_tax_band == B] from policyengine-uk.
• y entry: 360-vector of per-LA dwelling counts from the canonical CSV. For LAs without VOA data — Scotland (the VOA summary tables don't cover Scotland) and NI (no council tax) — the value falls back to national_count × la_household_share, matching the existing tenure block's fallback pattern.

Two targets are deliberately not wired in this pass:

• Band I — Wales-only and mostly null in the CSV.
• Band D £ amount (ons/council_tax_band_d/{code}) — a per-rate quantity, not an aggregate. Converting it to a calibratable total (e.g. council-tax revenue) needs Scotland-specific band ratios that diverge from England/Wales after the 2017 reform, and is worth a separate PR.

Documented coverage gaps

• Northern Ireland (10 LAs): NI uses domestic rates, not council tax. has_council_tax=False flag set; no targets emitted for NI; loss matrix uses national-share fallback.
• Scotland band counts (32 LAs): the VOA summary doesn't cover Scotland. Scottish Assessors publishes per-LA chargeable-dwellings separately; follow-up PR.
• City of London Band A: VOA suppresses this cell ([c]) for disclosure control; other bands populated.

Testing (the #371 lesson)

22 hermetic tests on the targets themselves plus 13 new tests on the loss-matrix wiring, all green locally:

Targets — test_la_council_tax_targets.py (22 tests)

• CSV structure (4): row count matches the reference LA list, expected columns present, four UK countries represented, every code matches the reference.
• Value plausibility (4): Band D amount in [£900, £3,500], total dwellings in [200, 800,000], explicit Isles of Scilly regression (total in [500, 5,000], guarding against the 2.49M outlier that slipped into Add LA-level household land value targets and calibrate on them #371), band totals sum to total dwellings within 20-property slack.
• Coverage expectations (4): every English / Welsh / Scottish LA has a Band D value; Northern Ireland is explicitly flagged as no-council-tax.
• Spot-checks (2): Wandsworth and Westminster are the two lowest-Band-D LAs (catches row-swap bugs); Scottish average Band D is £500+ below English average.
• Target-API invariants (8): get_targets() returns non-empty without network, Band D target count matches CSV, band count target count matches Σ non-null band columns, every target carries LOCAL_AUTHORITY geo level + geo_code, correct units and is_count=True on count targets, every target has ≥1 year of values, every band count ≤ 500k.

Loss-matrix wiring — test_la_loss_council_tax.py (13 tests)

Light layer (no Microsimulation)

• Every LA code in local_authorities_2021.csv has a row in la_council_tax.csv.
• The eight count_band_{A..H} columns exist.
• E/W rows have band A-H counts populated (allowing for City of London disclosure suppression).
• Scotland band counts are uniformly null as documented.
• NI rows have has_council_tax=False.

Full-build layer (gated on the enhanced FRS fixture)

• All eight voa/council_tax/{A..H} columns present in matrix and y.
• y vectors length 360.
• All y entries finite and positive (Scotland and NI receive the fallback rather than NaN or zero).
• Matrix entries are 0/1 indicators with rows summing to ≤1 (each household sits in at most one band).
• y matches the CSV verbatim for an English LA (Hartlepool).
• Scotland and NI LAs receive a positive fallback estimate.

Sanity check — top 10 LAs by Band D amount (2026-27 where available)

LA	Country	Band D
Dorset	ENGLAND	£2,765
Lewes	ENGLAND	£2,756
Nottingham	ENGLAND	£2,755
Rutland	ENGLAND	£2,738
Wealden	ENGLAND	£2,728
Merthyr Tydfil	WALES	£2,594
Bridgend	WALES	£2,555
Hartlepool	ENGLAND	£2,560
Newport	WALES	£2,443
Blaenau Gwent	WALES	£2,452

Bottom of the table is dominated by the central-London LAs with the lowest precepts: Wandsworth (£1,028), Westminster (£1,050), City of London (£1,330).

Out of scope for this PR (follow-ups)

• Scottish Assessors per-LA chargeable-dwellings to fill the Scotland band-count gap.
• Wiring the Band D £ amount as a total-council-tax-revenue target (needs Scotland-specific band ratios).
• Council Tax Support caseload per LA (DWP StatXplore) — caseload-style, different target shape.
• Single Person Discount rate per LA (CIPFA Council Tax Base) — same.

Related

• #371 — Add LA-level household land value targets and calibrate on them — the regional methodology this PR parallels; also the origin of the disclosure-control bound-check test lesson.
• targets/sources/voa_council_tax.py already provides band targets at the REGION level; this PR adds the LA level alongside it without modifying the regional path.

[vahid-ahmadi]

Review — items to address before merge

Blocking

1. Welsh Band I is dropped. Wales has had 9 council tax bands (A–I) since its 2005 revaluation. _BAND_COUNT_COLUMNS in la_council_tax.py only maps A–H, and the CSV has no band_I column, so every Welsh Band I dwelling is silently missing from the emitted targets. If a calibrator consumes these to set the band distribution, Welsh Band I is implicitly targeted to zero.

Fix: add band_I column to la_council_tax.csv (populated for Welsh rows, NaN elsewhere), add "I": "band_I" to _BAND_COUNT_COLUMNS, update the module docstring, and add a test asserting non-null Band I for the larger Welsh LAs (Cardiff, Swansea, Vale of Glamorgan, Monmouthshire).

2. total_dwellings is Σ(A..H) for Welsh rows, not VOA "All properties". Verified from the CSV:

• Cardiff (W06000015): 5320 + 20420 + 34920 + 37370 + 31540 + 22250 + 10590 + 2850 = 165,260 — matches CSV exactly.
• Monmouthshire (W06000021): 520 + 3450 + 7140 + 9460 + 7660 + 8100 + 5520 + 1790 = 43,640 — matches CSV exactly.

Because both sides of test_band_counts_sum_to_total are derived from the same A–H sum, the test passes trivially and does not validate fidelity to the published total. Fix: re-source total_dwellings from the VOA "All properties" column. Once done, the test becomes a real fidelity check with the existing 20-property slack.

Coherence

3. variable field diverges from voa_council_tax.py. The existing regional source uses variable="council_tax_band" with no breakdown_variable; this PR uses variable="council_tax_band_count" with breakdown_variable=f"council_tax_band_{band}". Downstream loss.py wiring will then need to know two names for the same conceptual quantity. Prefer aligning — either update voa_council_tax.py to the new scheme or match its convention here.

4. CSV column asymmetry. band_A, band_B, band_C, band_D_count, band_E, band_F, band_G, band_H — the _count suffix only on D to avoid clashing with band_d_amount. Cleaner to use a count_band_X prefix scheme, or rename band_d_amount → amount_band_d so the count columns stay symmetric. This also simplifies _BAND_COUNT_COLUMNS to a plain prefix map.

Nits

5. _load_table() re-reads the CSV on every get_targets() call. @lru_cache matches the pattern in voa_council_tax.py and is free.

6. has_council_tax column is derivable from country == "NORTHERN_IRELAND". Arguably self-documenting, so leave if preferred.

7. Module docstring opens with "band A–H" — update once Band I is added.

PR description

With Band I added, the "2,541 band-count targets" figure will rise by the number of non-null Welsh Band I cells — worth regenerating the description from the final CSV before merging. My earlier arithmetic of 318 × 8 = 2,544 minus City of London band A = 2,543 gave 2 more suppressions than the description claimed; Band I coverage may resolve part of that gap too.

[vahid-ahmadi]

Pushed d2d2acf addressing the review items:

Blocking fixes

• Welsh Band I now sourced from VOA CTSOP1.0 for all 22 Welsh LAs (Cardiff 1480, Monmouthshire 670, Vale of Glamorgan 1060, etc.); emitted as voa/council_tax/{code}/I targets. New tests: test_welsh_las_have_band_i, test_english_las_have_no_band_i, test_cardiff_band_i_matches_published_figure.
• total_dwellings now sourced from VOA "All properties" for every English + Welsh row rather than derived from Σ(A..H). test_band_counts_sum_to_total is now a real fidelity check (max discrepancy 10, VOA rounding).

Coherence fixes

• Band-count targets renamed to voa/council_tax/{code}/{band} with variable="council_tax_band" (aligned with the existing regional voa_council_tax.py). breakdown_variable dropped to match. council_tax_band is the actual PolicyEngine-UK enum variable and has Band I in its possible values.
• CSV count columns renamed to the symmetric count_band_A..count_band_I scheme (old band_D_count outlier removed).

Nits

• _load_table() now @lru_cache-ed.
• Module docstring clarified: "A–H in England/Scotland, A–I in Wales".

Final target counts: 350 Band D amount + 2,563 band counts (up from 2,541: +22 Welsh Band I, +2 resolved [c] suppressions). 25/25 tests pass locally via uv run pytest.