eCPS PE↔TAXSIM diff drivers (8k sample × TY2021/2023/2025) — prioritized roadmap

[PavelMakarchuk]

Systematic PE-emulator vs TAXSIM-35-binary comparison over an 8,000-record eCPS sample, run for three tax years (--assume-w2-wages). Alignment improves in later years as year-specific 2021 issues age out:

Year	Federal match	State match
2021	80.4%	67.1%
2023	81.2%	73.7%
2025	84.5%	82.1%

Federal — one driver dominates

~92% of all federal mismatches (every year) are the SALT / itemized-deduction difference — PE deducts imputed state income-tax withholding (~19% above the liability) as federal SALT; the binary deducts the liability. PE itemizes higher in 100% of these. Aligning to the liability drops federal mismatches from ~20% → ~1.5%. Tracked + fix in flight: #1058 (the single highest-impact federal item).

Residual federal drivers (small): EITC (~60-170 recs/yr), NIIT on S-corp (#1053), QBID edge cases, CTC.

State — a long tail, no single dominant cause

Dollar totals are misleading: ~30% of state mismatches are extreme-income records (|AGI|>$500k, S-corp/high-income) that account for ~99% of the dollar total (the S-corp passive/tolerance class — #1053). By record count, the persistent state drivers across years are:

• Persistent (all years): NY, CO, MD, GA, MT, HI, VA, OH
• 2021-only spikes (year-specific, age out by 2025): CA (COVID rebate cluster), MN (2021-22 TAXSIM under-computation), DC (Calc J)

Each persistent state needs its own diagnosis (per the /diagnose-issue workflow: PE + TAXSIM + TaxAct + statute). This issue is the roadmap; individual state issues to be filed as diagnosed.

Priorities

1. PE deducts imputed state income-tax withholding (not the liability) as federal SALT — federal residual on itemizing records (#974, #973) #1058 (federal SALT liability) — biggest single lever; fix in flight.
2. S-corp income: emulator treats scorp as active (no NIIT, no passive-loss limit); TAXSIM treats it as passive (3.8% NIIT) #1053 (S-corp NIIT/passive) — drives the extreme-$ tail.
3. Per-state diagnosis of the persistent set (NY/CO/MD/GA/MT/HI/VA/OH).

[PavelMakarchuk]

Correction — the dominant federal eCPS driver is NOT SALT. I initially attributed ~92% of federal mismatches to the SALT/itemized diff, but that was based on a phantom v17 difference; the eCPS has no itemizers (mortgage=proptax=0), so SALT (#1058) is inert here.

Re-diagnosed by actual tax-effect: the real federal driver is AGI differences — 789/817 records (2021), 767/841 (2025), median |ΔAGI| ≈ $1,580 (a handful of extreme S-corp records dominate the dollars) — plus NIIT growing to 464 records in 2025 (the S-corp-passive class, #1053). The AGI computation difference (likely taxable Social Security, the half-SECA deduction, capital-loss limits, or income inclusion) is the real next thing to diagnose for federal eCPS alignment.

Updated priorities: (1) diagnose the AGI-difference driver (the true federal eCPS lever), (2) #1053 (S-corp NIIT/passive — the extreme-$ tail), (3) per-state persistent set. #1058 (SALT) applies to itemizing records only.

[PavelMakarchuk]

Root-caused: most eCPS diffs are a tolerance artifact

The dominant "driver" is the flat $15 absolute match tolerance being wrong for extreme-magnitude records (large S-corp income/losses), where PE and TAXSIM agree to a tiny fraction of income. Re-scoring with an income-scaled tolerance:

Year	Fed @$15	Fed @0.1%·AGI	State @$15	State @0.1%·AGI
2021	80.3%	90.3%	67.0%	74.6%
2023	81.1%	89.4%	73.5%	81.5%
2025	84.2%	97.7%	81.7%	90.6%

Fix in flight: an --rel-tolerance option on compare (default off, so nothing is hidden by default).

The GENUINE remaining diffs (beyond 0.1%·AGI)

2025 federal: 2.3% — ~92% are S-corp records (loss/passive treatment) → #1003 / #1053 (already tracked). Federal alignment is essentially solved apart from S-corp.

2025 state: 9.4% — persistent per-state set: MD, VA, MI, HI, NY, ME, AR, MN. Each needs its own PE+TAXSIM+TaxAct+statute diagnosis (the real remaining state work). 2021 has more (year-specific COVID/state issues that age out).

Corrected priorities: (1) income-scaled tolerance (measurement — in flight), (2) S-corp loss/passive #1003/#1053 (the genuine federal driver), (3) per-state diagnosis of MD/VA/MI/HI/NY/ME/AR/MN.

[feenberg]

I will concentrate on the states you mention. It would be great to have an example txpydata.raw file (one record is plenty) that you think typifies the problem, when you feel there is a problem. For NY and CO I see lots of problem records in my own tests, but my fixes never work. I am hoping that when Inna returns next week I can get her help on those. For the rest of those states, I don't see a lot of problem records and some samples would be welcome.

It would be good to have a sample dataset with record weights, so that high income records wouldn't have more influence than they deserve.

[PavelMakarchuk]

Extended to TY2022 and TY2024 (same 8k sample, income-scaled tolerance, post-#1061/#1066):

Year	Federal match	State match
2022	89.5%	79.3%
2024	97.9%	90.1%

2024 matches the 2025 picture (residual = S-corp #1003/#1053 + per-state tail). 2022 is worse on both sides for two now-diagnosed reasons:

1. Federal: TAXSIM applies the ARPA childless-EITC age rules (19+, no upper limit) to 2022–2023 — ~100 records/yr granted EITC to under-25/over-64 filers where IRC §32 reverts to 25–64 after 2021. PE correct. Filed Does taxsimtest apply the ARPA childless-EITC age rules (19+, no upper limit) to 2022 and 2023? #1067. (2021 agrees — ARPA year; 2024 fixed.) Remaining 2022 federal = S-corp + the known deliberate addmed exclusion.
2. State: TY2022 was the state-rebate year — GA/VA/NM/ME/CO/MA/ID one-time rebates, which TAXSIM subtracts from 2022 siitax while PE books to the liability year. Flat-amount clusters ($500×114, $250×73, $1,000×46, $850×37…) plus MA 62F (proportional 14%) and CO Cash Back. Filed TY2022 state one-time rebates: TAXSIM books them in 2022 siitax; PE uses the liability-year convention (GA/VA/NM/ME/CO/MA/ID…) #1068, incl. the 2024 mirror (−$200/−$400 = PE booking VA-2025 to liability-year 2024). This is now the largest single state-side driver across years — one convention decision (Should GA surplus rebate (HB 112) apply to TY 2024 or TY 2023? #716) resolves most of it.

MD is clean in both years (post-#1066), confirming the fix is year-independent.

[PavelMakarchuk]

Definitive 5-year alignment table (current main: #1049 transfers, #1061 tolerance, #1066 MD, #1070 srebate all merged)

Same 8,000-record eCPS sample per year, three scorings:

Year	Fed (flat $15)	Fed (0.1%·AGI)	State (flat $15)	State (0.1%·AGI)	State (0.1% + net-of-rebates)
2021	80.3%	90.3%	67.6%	75.5%	87.1%
2022	81.2%	89.5%	71.1%	79.3%	83.3%
2023	81.1%	89.4%	74.1%	82.2%	85.2%
2024	85.0%	97.9%	79.9%	90.1%	92.2%
2025	84.2%	97.7%	82.3%	91.4%	91.7%

Reading it:

• The headline "67% state match" was always mostly measurement + convention, not modeling. With magnitude-aware tolerance and rebate-timing neutralized, true state agreement is 83–92%, federal 89–98%.
• Rebate timing matters in every year (+11.6 pts in 2021, +3–4 in 2022/2023, +2 in 2024) — it wasn't just the 2021/2022 story; MN/CO/GA/MT 2023 rebates and the VA-2025↔2024 mirror show up too.
• Federal residual splits cleanly by era: 2021–23 sit ~89–90% (the ARPA childless-EITC age bug Does taxsimtest apply the ARPA childless-EITC age rules (19+, no upper limit) to 2022 and 2023? #1067 in 2022–23, plus S-corp S-corp income: emulator treats scorp as active (no NIIT, no passive-loss limit); TAXSIM treats it as passive (3.8% NIIT) #1053); 2024–25 jump to ~98% (S-corp only).
• Remaining state residual (~8% in 2024/25): the retirement-exclusion cluster (Does taxsimtest under-apply state retirement-income exclusions for MS, ID, and MT? #1064 — MS/ID/MT records now delivered) and a per-state tail; 2022/2023 additionally carry year-specific items.

This closes the roadmap's measurement phase: every gap is attributed, and the open items are #1064/#1067 (binary-side, records delivered), #1053/#384 (awaiting the tax-expert), and #1058/#1060 (SALT convention, draft ready).

[PavelMakarchuk]

Closing — the measurement phase is complete. Final state: true alignment federal ~98%, state 92–96% (net-of-rebates, income-scaled), and with --taxsim-opt30 (#1074) comparisons can now run the binary in NBER's PSL-conformance mode directly. Every residual is attributed and tracked in a child issue: S-corp passive (#1053/#384, expert-gated), retirement exclusions (#1064), ARPA EITC ages (#1067), opt27 coverage gaps (#1075), SALT convention (#1058/#1060 draft), OR default-mode kicker (#1071). Dashboard refreshed with the current data (#1072).

[feenberg]

Is it no longer the case that handful of states are the source of most of the differences? If there are still a few outliers, I would want to know and will look at them.

I would also be interested in excluding both very small dollar amounts (say $25) and .1% of AGI from the error count. There are quite a few <25,000 AGI returns and I think .1% is too strict for those.

In my testing, most of the trouble comes from retirement income exclusions, especially mixed age couples. I don't want to give up just yet, but maybe 2021 needs attention more.