populace-us

The populace-built US population: a calibrated synthetic microdataset for PolicyEngine-US, built by the populace stack entirely from primary sources — the enhanced CPS appears only as the benchmark this file is scored against, never as a build input. It loads anywhere the enhanced CPS loads (an API-compatible alternative population), with its own calibrated weights — and its own strengths and gaps, both documented below.

Load it

pip install 'populace-data[us]'

from policyengine_us import Microsimulation
from populace.data import load

sim = Microsimulation(dataset=load("us", 2024))
sim.calculate("household_net_income", 2024).sum()

Or grab the H5 directly:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="policyengine/populace-us",
    filename="populace_us_2024.h5",
    repo_type="dataset",
)

How it is built

One HDF5 USSingleYearDataset per year. Every layer comes from a primary survey or administrative source:

source	provides
Census CPS ASEC	household structure, demographics, incomes, benefits, tenure, hours, occupation flags, health coverage at interview, retirement distributions (DST codes), childcare, prior-year income (longitudinal PERIDNUM join)
IRS SOI Public Use File 2015 (uprated)	tax detail: capital gains, dividends, interest, itemized-deduction inputs, QBI/SSTB components, partnership self-employment, estates, tuition
Fed SCF 2022	wealth: bank/stock/bond assets, net worth, mortgage balance hints
Census SIPP	tip income for tipped occupations; household vehicles (count and value)
CPS-ORG	hourly wage, paid-hourly status, union coverage
MEPS-IC parameters	employer-sponsored insurance premiums
Census ACS 2022	rent for renter households

Imputations use weight-aware quantile-forest models fit on each donor's own records, and every imputed value is clipped to that donor's realized range (the support guard) — nothing is anchored to the enhanced CPS. The result is calibrated to PolicyEngine's administrative target surface (3,704 IRS/Census/ program targets, plus a signed net short-term capital gains target so the optimizer cannot silently drive a net-negative aggregate to extremes) with a hard per-record weight bound (max_weight_ratio=50), so no aggregate leans on a handful of super-weighted records.

Acceptance gates

The build refuses to publish unless every gate passes; this file passed all of them:

• Parity 0: every PolicyEngine input layer the enhanced CPS populates non-degenerately, this file's simulation populates (169 reference layers checked at simulation level).
• Exported-nonzero: all 308 stored columns carry signal — no all-zero scaffolding that would silently mask engine formulas or defaults.
• Calibration: 94.66% of 3,704 targets within 10% (loss 0.022); max household weight 379,623 with zero records above 500k (the enhanced CPS ships 21, max 1.05M).
• Smoke aggregates through Microsimulation: 332.8M people, $98.0B SNAP, $175.5T net worth (Fed Z.1 ≈ $169T), net short-term capital gains −$77.4B against the −$76.8B PUF-anchored target, tips $53.1B, rent $759.7B.

Validation

Scored by the sound comparison — matched samples (41,314 households), symmetric weight refit on the full administrative target surface, held-out targets never seen by either side's refit. Lower is better.

metric	populace-us	enhanced CPS
training loss (2,965 targets)	0.176	1.089
held-out loss (739 unseen targets)	0.037	0.317
full-surface loss (3,704 targets)	0.213	1.406

Per individual target the incumbent still wins more often (2,528 of 3,704 to our 1,127, 49 ties): populace wins big where it wins and loses narrowly where it loses. Both facts are the story.

Known gaps

We publish the misses with the hits:

• Net worth runs ~4% above Fed Z.1 ($175.5T vs ≈ $169T): the calibration target ($160T) sits below Z.1 and the achieved total lands between them.
• Investment interest expense is thin ($7.2B against IRS SOI ≈ $24B): the PUF-residual rule populates the layer conservatively; a dedicated SOI calibration target is the roadmap item.
• Per-target wins vs the incumbent: see Validation — aggregate losses are what the comparison gates on, but per-target patterns differ between the two populations. Results are not interchangeable.

The dashboard at populace.dev/dashboard shows the full per-family calibration fit, the worst-fit targets by name, the weight distribution, and a live strip while a build chain runs. Methodology and evidence: populace.dev; loader and registry: github.com/PolicyEngine/populace (packages/populace-data).