Summary
Simulated alternative minimum tax validates 59% below IRS actuals. The mechanism is structural: PE-US's amt_income is taxable_income + amt_excluded_deductions + amt_separate_addition + exemptions — the SALT-addback channel — with no AMT preference items (incentive-stock-option exercise spread, private-activity-bond interest, depletion/intangible drilling), and the dataset carries no inputs that could feed them.
Evidence
On populace_us_2024.h5 (…f0af251…20260620), vs IRS SOI TY2023 (Pub 1304 Table 1.4):
|
populace |
SOI actual |
| AMT liability |
$1.13B |
$2.75B |
| AMT payers |
123k (weighted) |
150k returns |
Payer counts are close (−18%) but dollars are −59%: populace finds the SALT-driven AMT filers but misses the high-dollar cases. Post-TCJA, realized AMT is dominated by ISO exercises and specialized preferences — per-filer amounts far above the SALT-addback cases. Populace's by-AGI pattern shows the shortfall concentrated in the $200k–$1M bands where ISO exercises cluster.
Nature of the fix
Two-layer: (1) policyengine_us would need preference-item variables on Form 6251 (model gap, upstream); (2) even then, survey microdata has no ISO-exercise records, so the input would need imputation from SOI aggregates. Until either happens, AMT should be treated as a documented structural understatement in validation — expected ~40% of actual — rather than a calibration defect.
Found via the SOI-actuals baseline backtest (PolicyEngine/calibration-diagnostics reform-validation page).
Summary
Simulated alternative minimum tax validates 59% below IRS actuals. The mechanism is structural: PE-US's
amt_incomeistaxable_income + amt_excluded_deductions + amt_separate_addition + exemptions— the SALT-addback channel — with no AMT preference items (incentive-stock-option exercise spread, private-activity-bond interest, depletion/intangible drilling), and the dataset carries no inputs that could feed them.Evidence
On
populace_us_2024.h5(…f0af251…20260620), vs IRS SOI TY2023 (Pub 1304 Table 1.4):Payer counts are close (−18%) but dollars are −59%: populace finds the SALT-driven AMT filers but misses the high-dollar cases. Post-TCJA, realized AMT is dominated by ISO exercises and specialized preferences — per-filer amounts far above the SALT-addback cases. Populace's by-AGI pattern shows the shortfall concentrated in the $200k–$1M bands where ISO exercises cluster.
Nature of the fix
Two-layer: (1)
policyengine_uswould need preference-item variables on Form 6251 (model gap, upstream); (2) even then, survey microdata has no ISO-exercise records, so the input would need imputation from SOI aggregates. Until either happens, AMT should be treated as a documented structural understatement in validation — expected ~40% of actual — rather than a calibration defect.Found via the SOI-actuals baseline backtest (PolicyEngine/calibration-diagnostics reform-validation page).