Summary
The experimental UK firm generator merged in #223 was ported from PolicyEngine/firm-microsim-paper before that repo's net-VAT liability correction, so it inherits the same defect: per-firm net VAT liability is computed as turnover − input instead of 0.20 × (turnover − input) (the standard rate applied to value added), with the input/output ratio drawn on [0.6, 1.5] so a large minority of firms carry negative value added.
Consequences (as measured downstream)
In firm-microsim-paper this inflated per-firm liabilities ~5× and left ~47% of firms with negative value added. The multi-target weight optimiser compensated on the band totals (aggregate calibration scores barely move — populace's 93.8% validator score is not evidence of correctness here), while distorting the near-threshold density in ways that manufactured spurious bunching structure. Full diagnosis: PolicyEngine/firm-microsim-paper#15; fix and re-run: PolicyEngine/firm-microsim-paper#21.
Fix (mirrors firm-microsim-paper#21)
vat_liability = STANDARD_VAT_RATE * (turnover − input) with STANDARD_VAT_RATE = 0.20 defined once- Input/output ratio recentred:
0.2 + Beta(4,2) * 0.6 + sector noise, clamped to [0.1, 0.95] (mean ≈ 0.6, value added strictly positive, net rate ∈ [1%, 18%])
- Apply the same formula in generation, calibration target matrix, and validation
- Note in docs that the model abstracts from net-repayment traders (negative liability requires rated-structure detail the band aggregates don't resolve)
Verified reference values after the fix (2023-24 vintage, firm-microsim-paper)
ρ mean 0.60, VA share 0.40, zero negative-VA firms; registered base (≥£85k) £184.7bn; overall calibration 89.4%.
🤖 Generated with Claude Code
Summary
The experimental UK firm generator merged in #223 was ported from
PolicyEngine/firm-microsim-paperbefore that repo's net-VAT liability correction, so it inherits the same defect: per-firm net VAT liability is computed asturnover − inputinstead of0.20 × (turnover − input)(the standard rate applied to value added), with the input/output ratio drawn on [0.6, 1.5] so a large minority of firms carry negative value added.Consequences (as measured downstream)
In firm-microsim-paper this inflated per-firm liabilities ~5× and left ~47% of firms with negative value added. The multi-target weight optimiser compensated on the band totals (aggregate calibration scores barely move — populace's 93.8% validator score is not evidence of correctness here), while distorting the near-threshold density in ways that manufactured spurious bunching structure. Full diagnosis: PolicyEngine/firm-microsim-paper#15; fix and re-run: PolicyEngine/firm-microsim-paper#21.
Fix (mirrors firm-microsim-paper#21)
vat_liability = STANDARD_VAT_RATE * (turnover − input)withSTANDARD_VAT_RATE = 0.20defined once0.2 + Beta(4,2) * 0.6+ sector noise, clamped to[0.1, 0.95](mean ≈ 0.6, value added strictly positive, net rate ∈ [1%, 18%])Verified reference values after the fix (2023-24 vintage, firm-microsim-paper)
ρ mean 0.60, VA share 0.40, zero negative-VA firms; registered base (≥£85k) £184.7bn; overall calibration 89.4%.
🤖 Generated with Claude Code