Stack HMRC size-band CGT donors and calibrate to per-band targets

[vahid-ahmadi]

Problem

The capital gains imputation samples from the Advani & Summers percentile table, which stops at p95 — so no imputed gain exceeds ~£2m and only ~1.6% of gains come from gains of £1m+ (HMRC Table 2.1a: ~61%). The aggregate targets added in #440 could not fix this: with no records carrying large gains, calibration had nothing to upweight, and satisfied the gains total by spreading gains even more thinly — the built enhanced_frs_2024_25 carries 1.86m CGT taxpayers against HMRC's 378k.

Approach — existing house patterns only

No imputation method is changed. This recombines two patterns already in the repo:

1. Zero-weight donor stacking (the SPI synthetic-household precedent in income.py, and the CG clone stack itself): 300 donor households per HMRC size-of-gain band (£12,300 → £5m+) are stacked after the capital gains imputation at household_weight = 0, flagged household_is_cgt_band_donor. Each donor's first adult carries the band's published mean gain. Donors are sampled with probability proportional to percent_with_gains at the first adult's total income, from the same Advani & Summers table the imputation already uses, so large gains land on plausible income profiles.
2. AEA-gated HMRC targets (the Add HMRC capital gains targets to constrain the gains distribution #440 pattern, extended per band): each Table 2.1a band becomes a taxpayer-count and a gains-amount target with custom_compute, gated on the annual exempt amount read from the parameter tree. Calibration — not any assumed distributional form — decides each band's weight. SPI-synthetic households demonstrably grow from zero initial weight in the published builds, so the mechanism is proven.

HMRC Table 2.1a is committed as capital_gains_size_distribution_hmrc.csv (stable-fetch reasoning in the header, following ons_demographics.py). Bands below £12,300 are skipped: HMRC's sub-AEA rows mix exemption regimes and the spline body already covers small gains.

Relationship to #443

Alternative to the Pareto-tail fit in #443: once band counts and amounts are targeted directly, the above-p95 shape assumption becomes unnecessary — the empirical band structure is the tail. The two should not both be merged as-is without re-checking the band targets against the Pareto-imputed distribution.

Tests

test_cgt_band_donors.py: band table integrity, per-band target values vs the published table, AEA/band gating of the custom computes, and full donor stacking on the base FRS (donor counts, zero weights, one gainer per household at exactly the band means, original weights preserved).

🤖 Generated with Claude Code