Fix savings income calibration to match ONS D.41g

[MaxGhenis]

Summary

Calibrates savings income from ONS National Accounts D.41 (household interest received) instead of SPI-based data.

Changes

1. Remove savings from SPI-based targets

Removes savings_interest_income from INCOME_VARIABLES in:

• loss.py (national calibration)
• incomes_projection.py (income projections)

The SPI dramatically underestimated savings income (~£3bn) because it only captures taxable interest reported on tax returns.

2. Add ONS D.41 target

Adds a new ons/savings_interest_income calibration target in loss.py using ONS HAXV series:

Year	ONS D.41 (£bn)
2020	16.0
2021	19.6
2022	43.3
2023	86.0
2024	98.2
2025+	98.2 (held flat)

Source: ONS HAXV - Households (S.14): Interest (D.41) Resources

Why ONS D.41 is correct for savings_interest_income

The variable lineage shows savings_interest_income should include all household interest (taxable + tax-free):

In FRS data (frs.py):

# ISA interest (account type 21) 
pe_person["tax_free_savings_income"] = sum_to_entity(
    account.accint * (account.account == 21), ...
)

# Taxable accounts (1, 3, 5, 27, 28)
taxable_savings_interest = sum_to_entity(
    account.accint * (account.account.isin((1, 3, 5, 27, 28))), ...
)

# Total = taxable + tax-free
pe_person["savings_interest_income"] = (
    taxable_savings_interest + pe_person["tax_free_savings_income"].values
)

In policyengine-uk (taxable_savings_interest_income.py):

def formula(person, period, parameters):
    total_interest = person("savings_interest_income", period) + ...
    exempt_interest = person("tax_free_savings_income", period)
    return max_(0, total_interest - exempt_interest)

The model then subtracts tax_free_savings_income (ISA interest) when calculating taxable savings income. So savings_interest_income should match the ONS total.

Impact

Before: PolicyEngine estimated ~£0.02bn from a 2pp savings tax increase
After: Should estimate ~£0.4-0.5bn (closer to OBR's ~£0.5bn estimate)

Fixes #228

🤖 Generated with Claude Code

[MaxGhenis]

Closing this PR - the fix needs to be implemented in policyengine-uk instead by adding a new uprating parameter based on ONS D.41g household interest data. This will allow the uprating_factors.csv to be auto-generated correctly rather than manually edited.