Land value calibration targets not converging

[vahid-ahmadi]

Summary

The ONS land value calibration targets added in #292 are correctly wired into the national loss function, but the calibration cannot bring aggregates close to the targets through reweighting alone.

Current state (v1.44.0)

Variable	Current	ONS Target	Gap
Household land (owned_land)	£0.05tn	£5.06tn	~99% under
Corporate wealth	£6.96tn	£2.04tn	~241% over
Property wealth	£8.08tn	£7.10tn	~14% over

Root cause

The underlying household-level values for owned_land are near zero across nearly all households. No amount of weight adjustment can scale £0.05tn to £5.06tn — the base values are too small. The calibration optimiser includes these targets in the loss function (confirmed in calibrate.py lines 86-93, 113, 124), but the gradient signal cannot move weights far enough.

This likely needs an imputation step (similar to how wealth variables are imputed from WAS) rather than relying on reweighting alone.

Relevant code

• policyengine_uk_data/targets/sources/ons_land_values.py — target definitions
• policyengine_uk_data/targets/build_loss_matrix.py lines 293-300 — dispatch routing
• policyengine_uk_data/utils/calibrate.py — national targets in loss function
• policyengine_uk_data/tests/test_land_value_targets.py — tests marked xfail

Source

ONS National Balance Sheet 2025: https://www.ons.gov.uk/economy/nationalaccounts/uksectoraccounts/bulletins/nationalbalancesheet/2025

[MaxGhenis]

Current status update after #326:

• This is still a real issue, but the shape of the problem has changed since the original report.
• Land targets are now wired into the calibration loss function, and we also have regional ons/household_land_value/{REGION} targets.
• Even so, corporate_land_value remains weakly identified relative to the rest of the target suite.

Latest evidence:

• In PR Refresh admin target vintages #326 (Refresh admin target vintages), the only failing CI assertion was test_land_value_aggregate[corporate_land_value].
• land_value and household_land_value still passed.
• corporate_land_value came in at £3.43tn against a target of £2.06tn, i.e. +66.4%.
• The PR itself did not touch land targeting; it refreshed other admin targets, which changed calibration weights enough to push this metric from barely passing to failing.

That suggests:

1. this is not just a missing-target problem anymore; the target exists
2. the current corporate-land target is too weak/noisy to stay stable under unrelated calibration changes
3. household land is much better anchored than corporate land because it also has regional targets

So the likely next options are:

• increase effective weight on ons/corporate_land_value in calibration
• improve the corporate-land target concept or data model
• add a regional corporate-land decomposition if a defensible source exists
• or explicitly treat corporate land as validation-only until it is strong enough to calibrate against reliably

I would not treat this as blocking #326, but I do think this issue should stay open until we decide which of those paths we actually want.

[MaxGhenis]

There is now a companion country-package fix in flight at PolicyEngine/policyengine-uk#1583.

That PR changes corporate_land_value to allocate aggregate corporate land by the current weighted distribution of corporate_wealth, rather than scaling through a stale corporate-wealth parameter total. It also refreshes the aggregate land parameters to the 2024 ONS split used by uk-data.

Local check against enhanced_frs_2023_24 at 2025 on that branch:

• household_land_value: £4.737tn
• corporate_land_value: £2.058649tn
• land_value: £6.796tn

So the corporate component is materially improved and now aligns to the current ONS corporate-land target level. The remaining residual gap is mostly on the household side, not the corporate side.