Skip to content

Revert US certified dataset to Enhanced CPS (MP 2cdd45d loses clean-surface comparison) #391

@MaxGhenis

Description

@MaxGhenis

Problem

policyengine.py v4.14.1 (#390) promoted the microplex "MP eCPS" artifact (mp-ecps-2024-2cdd45d-20260605, HF revision a091769a) to the certified US dataset, replacing the production Enhanced CPS (policyengine-us-data 1.115.5).

That promotion was decided on a target surface that mis-scored the incumbent eCPS (unweighted MSRE 3.464; e.g. unemployment compensation +615%), so it cleared a content baseline gate rather than the numeric baseline-sanity gate.

Evidence

On a correctly-scored target surface (eCPS unweighted MSRE 0.174), two independent symmetric-refit comparisons agree to the digit that production eCPS beats MP 2cdd45d:

metric (lower is better) eCPS MP 2cdd45d
full PE-native loss 0.02662 0.04275
holdout loss 0.00769 0.00938
unweighted MSRE 0.174 0.290
target wins (of 2809) 2334 267 (208 ties)

MP is 14–49× worse on every income component (interest, retirement, dividends, capital gains, wages, self-employment) — the variables that drive AGI, brackets, and phase-outs in a tax model — and only better on SNAP.

Fix

Restore the certified US dataset to enhanced_cps_2024 (policyengine-us-data 1.115.5, revision d47fb547, model policyengine-us==1.715.2) via the supported scripts/refresh_release_bundle.py path, and cut a patch release. MP remains a research track, to be re-promoted only once it wins on a correctly-scored surface (after the CPS-passthrough income-component fix).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions