Capital Gains Withholding

Why it matters, in detail. Per Table 2, withholding accounts for $1.62T of the abstract's '$2 trillion without raising rates' (-22%) and $1.82T of the $2.23T score at 39.6% (-19%). The non-withholding row is $417-506B at every threshold. That pattern fits constructive realization (plus the rate change), which applies to every estate above the exemption and every filer above $1M. So for about a fifth of the revenue, the statement that it comes only from the roughly 90,000 withholding taxpayers does not hold. At the $1B threshold, withholding is $442B, about 1.5-1.6 times projected estate-and-gift receipts, not 'over three times'. The rule PWBM used to split the rows is undocumented (F3), so withholding's exact marginal effect relative to the package without withholding is not identified. For the abstract's scenario, however, every reading of Table 2 gives about $1.6T.

Evidence. Table 2 (p.21) transcribed from the image and checked against PWBM's 20 Jan 2021 workbook: 48/48 published cells identical, and the components sum to the totals within $1B (scripts/adj_table2_checks.py -> out_adj_table2_checks.txt). Withholding share of the window total: - panel B, $50M: 77.9% - panel A, $50M: 81.3% - panel A, $1B: 47.8% CBO estate-and-gift baseline for FY2021-30 (CBO eval-projections dataset): $280.9B (Sept 2020) and $300.3B (Jan 2020). Against these, $925B is 3.3x (3.1x), but withholding alone is 1.57x (1.47x) (scripts/adj_source_checks.py). Slide 3: 'Limited revenue: $450B over ten'. Figure 1 (p.22) plots $5.07T ... $0.93T, i.e. the Table 2 totals.

Why this severity. The PWBM totals are reported accurately, and the split is disclosed on p.15 and in Table 2, so nothing is hidden. But the abstract credits withholding with about 22% more revenue than the paper's own decomposition gives it. That is just under the rubric's quarter-of-headline threshold for 'major'. The billionaire comparison more than doubles withholding's contribution.

The authors' response. None located. The paper itself discloses the split on p.15 ('$1.8 trillion of the $2.2 trillion score would be paid as withholding payments while $0.4 trillion would be paid as constructive realization').

Why it matters, in detail. Net in-window withholding ($1.6-1.8T) is identically the end-2030 balance of prepaid tax owed back as credits. The one-time 90% prepayment on the 2020 stock alone is about $1.5T at 23.8% (after a 15% haircut; $2.4T at 39.6%), the same order as the whole withholding row. After that, the steady-state net gain is the growth of the prepaid stock, g x 0.9 x tax rate x gains. The authors put it at $20-30B/yr (2020 economy, apparently at current rates); a stylised steady state gives about $60B/yr. Either is 10-35% of annual in-window receipts. If balances earn the Treasury rate i, the steady-state gain is about (g - i) x balance, roughly zero. The present value of the withholding component (about 80% of the headline) is then close to zero, apart from behavioural and political-economy effects. Without interest, the gain is the forgone interest on the balance, an implicit tax on the prepaying families. Loan-financed withholding lowers the scored deficit but not federal borrowing. Its share is unknown but potentially large: 27-71% of SCF-group withholding is owed by households whose prepayments exceed their liquid assets.

Evidence. Quotes: abstract; pp.5, 10-11, 16-17; fn 12 (later withholding amounts 'will likely be dramatically smaller than his 2021-2030 annual amounts which cover decades of unrealized gains'); slides 5, 8 ('Withholding naturally taxes the very wealthy's $8T of already accrued capital gains which generates enormous revenue') and 9. Arithmetic: 0.9 x 0.238 x $8,086B = $1,732B, or $1,472B after a 15% haircut; at 39.6% the figure is $2,882B. Stylised account model (scripts/adj_timing_model.py -> out_adj_timing_model.txt): once accounts reach the 90% cap, the recurring net gain is about growth x 0.9 x tax rate x gains, roughly $60B a year at 23.8% in a 2020 economy after a 15% haircut. That is a third or less of average in-window receipts. If interest is credited at rate i, the gain falls to (g - i) x balance. Liquidity (scripts/adj_scf_table1.py; this ignores borrowing capacity): households owing 27% of the SCF group's year-1 withholding owe more that year than their liquid financial assets. For the cumulative nine-year prepayment the share is 71%. PWBM's later note on Wyden's Billionaires Income Tax: 'more than half' of its $507B 'would come from a one-time transition tax on previously accrued, unrealized capital gains'.

Why this severity. The ten-year cash score follows standard scoring conventions, and the paper calls withholding a prepayment, so the number itself is not wrong. But most of it is tax collected early on gains that already exist. The recurring gain is several times smaller. Its present value depends on undecided interest and refund rules and on counting loan-financed payments as revenue. None of this is quantified alongside the headline.

The authors' response. Partly anticipated. The paper says the net effect 'slowly declines over time' (p.16), gives the $20-30B long-run figure (p.17), and presents the scoring of loan-financed payments as a deliberate feature (pp.5, 10; slide 9).

Why it matters, in detail. Without the model, gross withholding, credits, avoidance and loan treatment cannot be separated, and inferences from the published totals are fragile. If the 2021 row is gross, PWBM's 2021 withholding ($195B) is about 40% below a mechanical base of 3.96% x capped gains (about $323B in 2021). That implies a semi-elasticity of roughly 10-13 (linear vs log form), close to the -13 PWBM used for the Warren tax. PWBM's implied credit offsets also look low relative to SCF realization behaviour, but that inference depends on undocumented timing and realization responses. The headline is therefore a conditional model output with a wide plausible range (see F4-F5).

Evidence. PWBM workbook and PDF: /mnt/zdata/downloads/syz2021-claude/pwbm_cgw_data.xlsx and pwbm_cgw_doc.pdf. Ratios from scripts/adj_table2_checks.py: ten-year withholding is 9.0-10.2 times the 2021 amount in panel A and 10.2-11.0 times in panel B. Pure growth would give about 12 (ten liability years) or about 10.5 (nine). The design credits withholding only against later realizations (Example 2), so the 2021 row is probably gross of credits. If so, PWBM's credits offset only about 0-25% of gross withholding over the window. Stylised account models using the SCF's 3.2% realization rate and 1-1.5% mortality give 30-45% (scripts/adj_timing_model.py). Panel A withholding is only 7-16% above panel B, although the withholding rate is 66-98% higher. PWBM's behavioural response is therefore large and depends on the rate; the paper attributes the small gap to 'substantial assumed avoidance' (p.15).

Why this severity. The figures are exact transcriptions of a genuine independent score, so the problem is verifiability, not error. But the sensitivity of the core $2.1-2.2T to its key assumptions cannot be assessed. Nor can the decomposition on which the paper's own attribution rests be checked.

Why it matters, in detail. Correcting the scope alone puts the authors' estimate 32% above PWBM's comparable figure. Scaling their own semi-elasticity of 8 to the 3.96% rate cuts it to $1.91T. A 7-8x multiplier then gives about $1.5-1.7T, or $1.3-1.5T with a 40% haircut. The estate-tax benchmark (the Summers-Sarin/Kopczuk logic) implies a shortfall in the realized base of roughly 35-65%, if withholding valuations behaved like estate-tax valuations. That is far above 15% and brackets PWBM's implied ~40%. The authors have a reasonable counter-argument from the wealth-tax debate: estate-tax avoidance and weak enforcement need not carry over, and a creditable prepayment gives less reason to under-value than a final tax does. The paper neither makes nor quantifies it. Instead it describes withholding as operating 'like a 3.96% annual wealth tax' (pp.7-8).

Evidence. Reproduction of the authors' year-1 score (scripts/adj_scf_table1.py -> out_adj_scf_table1.txt; matches both auditors): $263.9B. That is mechanical $310.5B, less $12.9B removed by the 2x cap, less the 15% haircut. Times 9 it gives $2.375T. The 50%-of-wealth variant gives $258.9B. Haircut sensitivities: - 30%: $1.96T - 31.7% (semi-elasticity 8 x 3.96%): $1.91T - 40%: $1.68T Multiplier: scripts/adj_timing_model.py -> out_adj_timing_model.txt. Warren-letter quote: scripts/adj_source_checks.py. SOI estate returns filed in 2016 (16es01fy.xlsx): $20-50M estates $31.6B (1,073 returns) plus $50M+ estates $71.4B (434) = $103.0B. SCF realized gains (kginc) of non-billionaire households above $50M: $199.6B, against an unrealized stock of $6,316B (3.16%).

Why this severity. This concerns the authors' corroborating calculation, not the PWBM headline. But the claimed independent agreement does not hold. Done like-for-like, with a haircut and multiplier consistent with the rate, the authors' method supports roughly PWBM's withholding-only $1.8T or less, not the $2.2T package figure it is said to match.

The authors' response. No response specific to this paper. In the wealth-tax debate, Saez & Zucman ('Response to Summers and Sarin', 25 June 2019) argue that estate-tax data understate top wealth because of avoidance and collapsed enforcement, and that a comprehensively enforced tax would see little avoidance.

Why it matters, in detail. With full refundability (with interest), price declines generate refunds that lower and destabilise receipts. With carryforward only, the ultimate tax liability does depend on price paths, contradicting the abstract. The paper does not disclose which rule PWBM scored, so the revenue impact of the choice is unknown. The choice trades the fairness guarantee against in-window revenue.

Evidence. Quotes checked in the paper (abstract; pp.5, 9, 10-12) and in the slides (scripts/adj_source_checks.py). Codex's illustration: an asset worth $100M with $90M basis generates nine 3.96% installments on the $10M gain, $3.56M in total. If it is later sold for $90M, the final tax is zero and only a cash refund reconciles the account. Treasury's later Billionaire Minimum Income Tax specified refunds both when prepayments exceed the tax on unrealized gains and at death (FY2023 Greenbook p.35). The rule can therefore be specified.

Why this severity. It undercuts a central design claim in the abstract (the advantage over mark-to-market) and leaves the revenue score's treatment of refunds unknown. But the fix, specifying refundability, timing and interest, is straightforward and does not overturn the proposal.

Why it matters, in detail. About $1.0-1.2T of PWBM's $1.8T withholding row depends on marking private-business gains to (SCF-reported) market values every year. Because gains are value minus basis, valuation shortfalls are leveraged: a 30% valuation shortfall with basis at a third of value removes about 45% of private gains. Under-valuation shifts tax from the window to realization or death (with penalties), so it lowers the ten-year score rather than the ultimate liability. The BMIT comparison shows that an administrable rule which does not mark non-tradable assets to market would collect roughly a third of PWBM's rate-adjusted withholding. It does not show that the paper's formula-based approach would yield that little.

Evidence. Table 1 composition reproduced: private business is 65.5% of gains above $50M (scripts/adj_scf_table1.py); Table 2, $1B column. FY2023 Greenbook pp.34-36 and revenue table ($360,843M) (scripts/adj_source_checks.py). PWBM withholding at the $100M threshold and 23.8% rate is $1,214B; scaled to a 20% rate it is about $1,020B. Removing private-business gains (64% of gains above $100M) leaves about $367B, close to the BMIT score (Claude's calculation, re-checked). Illustration: if private holdings were valued at 60-80% of market and basis is about a third of value, private gains would fall by roughly 30-60% and the withholding base by roughly 20-40%.

Why this severity. The dependence is real and unexamined, and plausible under-valuation could cut withholding revenue by a fifth or more. But the impact on the headline is not demonstrated: PWBM's large implicit avoidance allowance may already absorb part of it; reconciliation at realization limits the gain from under-valuing; and the BMIT gap mainly reflects a deliberate choice to exempt non-tradable assets, not evidence that the proposed valuations would fail.

The authors' response. Anticipated in the paper (pp.9-11): formula and market-signal valuation, ex-post reconciliation with denied refunds and late-payment penalties, look-through and information reporting, and in-kind payment as an alternative.

Why it matters, in detail. The relative SE is about 13% for gains above $50M and about 17% above $100M, and higher for SCF-only cells. Revenue estimates built on these cells, whether PWBM's or the authors', inherit at least this uncertainty before any model or behavioural uncertainty is added.

Evidence. scripts/adj_scf_se.py -> out_adj_scf_se.txt. Method: 999 SCF bootstrap replicate weights (p19_rw1) plus between-implicate variance. Sanity check: mean net worth $746.8K with SE $16.4K. Results: - SCF part above $50M: $6,277B, total SE about $1,053B. - With Forbes added: $8,127B +/- 13% (95% CI about $6.1-10.2T). - Above $100M: $6,276B +/- 17% (95% CI about $4.2-8.4T). - Authors' year-1 score: $264B, SE about $36B (13%). - Distinct non-billionaire SCF families above $50M: 308.

Why this severity. Sampling noise is symmetric and leaves the qualitative conclusions intact. It does mean that the $8T base, and the revenue figures built on it, carry 95% intervals of roughly +/-25% from sampling alone.

Why it matters, in detail. The SCF household share above $50M is 0.081%, 1.6 times the stated 0.05%; the 0.05% is a tax-unit share. There is no revenue implication.

Evidence. scripts/adj_scf_table1.py (household counts and shares by group); Table 2 counts; pp.8, 15, 22.

Why this severity. A labelling inconsistency. The affected population is about 90,000-105,000 units under either convention, and no estimate changes.

Why it matters, in detail. The heading overstates the >$50M stock by a factor of 2.2; the note reverses the sign of the rate effect on taxpayer counts.

Evidence. p.2 heading; Table 1; Table 2 note (p.21); scripts/adj_table2_checks.py (A - B counts run from -324,788 to -23).

Why this severity. Both slips can be corrected from nearby numbers and change no estimate.

Why it matters, in detail. The footnote understates the SCF ratio by about 3 points (>$50M) and 6 points (>$100M); Table 1, not the footnote, is consistent with the microdata.

Evidence. scripts/adj_scf_table1.py; Claude's fn21_variants.py (raw, aged, over assets, mean of ratios, 2016 SCF).

Why this severity. The footnote's conclusion, that assuming 50% for Forbes is conservative, survives because the SCF ratio is at or above 50%. Nothing downstream changes.

Why it matters, in detail. The Bezos example overstates the prepaid balance by $7.9B: under the 90% rule the account holds $71.3B and $7.92B remains due at sale (see F13). Example 2 is conditionally correct.

Evidence. Arithmetic re-done by the adjudicator: tax on $2M under the fn 10 brackets is $559,170; 9 x $7.92B = $71.28B; 0.9 x 13.3% = 11.97% (also in Claude's misc_checks.py output).

Why this severity. These are illustration-level slips; the design rules themselves are internally consistent.

Why it matters, in detail. Using the proposal's own rate, the illustrative dilution is 45%, not 26%, and it starts from a stake about 3 points too high.

Evidence. Facebook 2020 DEF 14A beneficial-ownership table (scripts/adj_source_checks.py); 16.2 x 0.9604^15 = 8.84%.

Why this severity. An illustration in the discussion of economic effects; no estimate depends on it.

Why it matters, in detail. The residual tax at realization is at least 10% of the liability once the cap is reached, and higher for recent gains, so deferral retains some value.

Evidence. pp.6 and 13; slide 7 ('Prepays 90%'); remaining liability at the cap = 0.1 x tax rate x gains.

Why this severity. A secondary claim about economic effects; no estimate depends on it.

Why it matters, in detail. The $235B benchmark for realization at death is a package score, so it cannot be compared directly with constructive-realization rows; the other items are descriptive inaccuracies.

Evidence. FY2017 Greenbook pp.155-156 and revenue table; IRS expatriation-tax page; ATF/TPC note (15 Oct 2019); Forbes data (Claude's forbes_checks.py); Treasury (2014) pp.3-5 (scripts/adj_source_checks.py for (a)-(c)).

Why this severity. These are footnoted benchmarks and supporting facts. Correcting them changes no estimate or conclusion.

Why it matters, in detail. The case that taxing before realization is 'necessary' leans on official scoring elasticities. It omits that the cited source estimates much smaller long-run responses, which would make rate increases under the current system more productive.

Evidence. Agersnap & Zidar, NBER w27705 (rev. Dec 2020), abstract and p.2 (scripts/adj_source_checks.py); 27.4% implied rate re-derived by Claude from their eq.(6).

Why this severity. The claim is motivational; the package itself raises the rate, and no estimate depends on it.

Why it matters, in detail. The California withholding score at the state's 13.3% rate is never stated. The long-run $20-30B (federal) and $2-3B (California) figures cannot be reproduced from the stated inputs.

Evidence. pp.15-17, fn 22-23; Claude's soi_2016_checks.py output (SOI 2016 estates, non-cash contributions and net gains; SCF kginc shares).

Why this severity. This is a documentation gap in a secondary, state-level result. The deviations mostly make the authors' long-run figure conservative.

Why it matters, in detail. If realization were held to be constitutionally required, withholding on unrealized gains (as opposed to a creditable payment framed as a prepayment) could be challenged; the paper offers no analysis.

Evidence. Moore v. United States, 602 U.S. ___ (2024), Opinion of the Court n.2 and pp.22-24 (scripts/adj_source_checks.py).

Why this severity. This is a legal design assumption outside the paper's empirical core. It is flagged, not adjudicated.

The authors' response. The paper's implicit answer is that withholding is a creditable prepayment of a realization-based income tax. The paper predates Moore.