Global Profit Shifting, 1975-2023 (earlier: Global Profit Shifting, 1975-2019)

Why it matters, in detail. For all multinationals the 2018-23 endpoint change is -0.04 pp/yr (-0.20 pp in total), with an OLS slope of -0.59 pp/yr (se 0.43). The 2019-20 values exceed 2017-18 by about 2 pp, and the whole 2015->2023 fall (-2.4 pp) comes in 2021-22. A TCJA-timed break exists only for US multinationals (-2.37 pp/yr after 2018, mostly in 2021-22). The timing that ties the decline to the TCJA is therefore unsupported at the global level.

Evidence. GWZ2026 abstract; pp. 2, 6, 8-9; Table 1 (p. 16); Table A1 (p. 18); workbook GWZ2026Data.xlsx DataF3-F4 cols AV, AE, G, BA (rows 45-53). Adjudicator: 2018->2023 change -0.20 pp; OLS 2018-2023 slope -0.59 pp/yr (HC1 se 0.43, not significant) for all multinationals vs -2.37 pp/yr (p<0.01) for US multinationals (scripts/table_a4.py).

Why this severity. The stated form of C3, a peak at the TCJA followed by a decline, is absent from the paper's own global series: the post-TCJA change is about zero. Only a weaker claim survives: the share was 2-5 pp lower in 2021-23 than in 2015-16, concentrated in 2021-22 and partly reversed in 2023.

The authors' response. None located. The paper's own Sec. 4 (p. 8) describes 2015-2019 as stagnation; Sec. 3 calls the US analysis 'provisory' and says the trend differential 'does not capture the causal impact'.

Why it matters, in detail. Under the paper's own reconciliation logic, a negative outward-minus-inward gap is no evidence of unrecorded haven income. Correcting it lowers global shifted profits by 1.6% in 2021 (-0.5 pp) and 5.6-5.9% in 2022 (-1.75 to -1.85 pp), net of the B10 reallocation, which offsets part of Ireland's -$94 to -$100bn. The 2023 raw gap (-$13.6bn) implies a smaller upward bias in the headline 2023 value, roughly 0.3-0.7 pp by analogy. This cannot be verified because no 2023 outputs are public. Country results change markedly: after correction Ireland falls from first to fifth or sixth place in 2022 ($70-76bn), behind the Netherlands ($153bn), Singapore (~$145bn), Hong Kong (~$104bn) and Switzerland ($95bn).

Evidence. GitHub repo idanngidron2/stata-project-profit-shifting (commit 654dfcef): code/Table B11.do l. 82-84, 139, 145-148; raw-data/OECD_BOP_2025_update.csv; output/Part1-profit-shifting-2021/2022.xlsx sheet 'Table B11'; paper workbook TableB11_2022 (IRL F4 = 39,665 used, C4 = 0) and Table U1 R37 = 42.07. The adjudicator independently re-implemented the B10/U1 propagation, which reproduces the 2021 and 2022 baselines exactly (scripts/b11_scenarios.py). Dropping the flipped US gap gives 2021 shifted profits $932.1bn -> $916.8bn (31.19% -> 30.67%) and 2022 $956.0bn -> $902.8bn (31.34% -> 29.59%); dropping the Japan flip as well gives $899.5bn (29.48%) in 2022.

Why this severity. A confirmed coding error that moves the 2022 global share by about 1.8 pp and cuts Ireland's 2022 shifted profits from $170bn to $70-76bn. Correcting it lowers post-2015 levels, which strengthens rather than reverses C3 and slightly trims the C1 endpoint.

The authors' response. None located; the July 2025 user guide's description contradicts the implemented sign flip.

Why it matters, in detail. Loss = sum over source countries of tau_s*S_s. The change in global receipts absent shifting is that sum minus sum_h tau_h*S_h. Under the authors' haven-ETR convention the second term is 36-45% of the first in 2016-2022 (2022: 6.90% gross vs 3.78% net of receipts). It is probably smaller in reality, because average ETRs overstate tax on shifted income, but it is positive. CFC/GILTI would lower the net figure further. The pre-2015 series is a proportional rescaling, so C2's long-run rise is unaffected.

Evidence. GWZ2026 abstract (ii), p. 5, p. 8; workbook DataF3-F4 cols BA, BF, E; repo code/Table C4d.do; code/prepare data for atlas.do l. 48-75 (value = shifted_profits * corporate_ETR); output/Atlas_tax_havens_allyears.xlsx. The adjudicator reproduced the haven-gain/gross ratios: 0.19 (2015, TWZ inputs), 0.36, 0.36, 0.43, 0.45, 0.45, 0.42, 0.45 (2016-2022); 2022 net = 3.78% of receipts, or 3.75% with the corrected Swiss CIT (scripts/net_vs_gross.py). TWZ (REStud 2023) Sec. 6.1.2 is titled 'Tax revenue losses in source countries' and computes haven gains separately (Supp. App. B).

Why this severity. The gross source-country loss is the standard statistic in this literature and matches the paper's p. 5 definition, and the C2 trend is unaffected. But the explicit net counterfactual on p. 8, and the natural reading of 'revenue lost from global profit shifting', overstate the worldwide effect by roughly 35-45% in 2016-22 under the authors' own conventions.

The authors' response. TWZ (REStud 2023, Sec. 6.1.2) present losses as being 'in source countries', note the CFC caveat and compute haven gains separately (Supp. App. B). The 2026 paper's global statement does not carry these caveats.

Why it matters, in detail. The implemented rule ties non-US shifting intensity to US intensity, while the described rule ties global dollars to US dollars; the two diverge because the US weight in foreign profits fell sharply. Under the described rule: 1975 8.3% (vs 4.3%), 1980 20.4% (vs 11.4%), a pre-2015 slope of 0.31 pp/yr (vs 0.54; the text says 0.7), and a counterfactual 2023 share of about 38.7% (vs '>40%'). C2's 1975 value scales with it (about 1.4% vs 0.7%).

Evidence. GWZ2026 p. 7 and fn. 17; workbook DataF3-F4 notes AO4/AP4 and formulas AO5:AO45, AP5, AV5; Table A1 (43%). Adjudicator: implemented 1975 value 4.25%. Under the stated rule (US share of global shifted dollars fixed at its 2015 value of 42.96%): 1975 = 8.26% (7.10% at 50%); pre-2015 OLS slope 0.31 vs 0.54 pp/yr; 1975->2023 rise 4.1x vs 7.9x; trend-break t-statistic -4.12 vs -5.45 (scripts/backcast_stated_trend.py, oil_benchmark.py).

Why this severity. The rule the text describes roughly doubles the 1975 starting value (8.3% vs 4.3%) and almost halves the pre-2015 trend. The large long-run rise and the statistical break after 2015 survive.

The authors' response. The implemented rule is described in workbook note AO4; the paper says the pre-2015 series 'have some margin of error'.

Why it matters, in detail. AE_t = AG_t * max(1 - AK_t/AJ_t, 0). With oil in the benchmark (AK 1.42 vs AJ 1.57 in 1975), almost no US haven profit counts as excess. Excluding oil is defensible, because oil rents inflate non-haven profitability, but the vintage change moves the abstract's starting value from under 1% to 4%. A net-of-tax basis, which removes petroleum levies, raises 1975 to about 10%. The narrative of a 2000s plateau followed by a 2010-15 'leap' is equally sensitive: under the Fig. A1 upper bound the 2000s share is already 32-33%.

Evidence. Workbook DataF3-F4 note AK4 and cols AK, AJ, AG, P, H, L, M, AW-AZ; GTER 2024 figure workbook 'Data Fig 2.7-2.8' col AK; WZ2022.xlsb.xlsx 'Data F4' col U; GWZ2026 p. 8 and Fig. 4 note (p. 15); NBER w30673 p. 8 ('virtually no profit shifting') and Fig. 4 note ('about 2%'). scripts/oil_benchmark.py, vintage_compare.py.

Why this severity. Defensible variants put 1975 anywhere from about 0.6% to 10-12%, against 34-38% in 2023. The size of the rise (3x to more than 50x) is therefore assumption-driven and was changed without disclosure; its direction is robust.

The authors' response. Workbook note AK4 documents the oil exclusion; the paper (p. 1) warns that pre-2015 series 'have some margin of error'.

Why it matters, in detail. Same-year revision ranges are up to 1.7 pp in the share in full-data years (2016-2020) and 3.9 pp for 2021-22 (projection replaced by data). For the tax-loss ratio they are 0.8-1.6 pp in 2016-2020, mostly from the denominator. Compare the C3 changes of -2.4 pp and -0.8 pp. On the 2023-vintage basis the tax-loss ratio rose after 2015 (8.4% -> 9.1-10.0% in 2016-19); on the 2026 basis it fell. The sign of the post-2015 revenue-loss trend therefore depends on the vintage.

Evidence. WZ2022.xlsb.xlsx 'Data F4' cols U/W and 'Table1'; GTER 2024 figure workbook 'Data Fig 2.7-2.8'; GTER 2024 text (Executive Summary 'Finding #2'; ch. 2 sec. 2.1 and Table 2.2 notes); TWZ 'Missing Profits of Nations: 2018 Figures' (Aug 2021) Tables 1-2 and p. 2; NBER w30673 abstract; GWZ2026 Tables 1, A2, A3. scripts/vintage_compare.py, vintage_components.py.

Why this severity. The revisions do not show the 2026 numbers are wrong. They do show that changes of 1-4 pp in this series lie within the range of data and method revisions, the same size as the 2.4 pp (share) and 0.8 pp (tax-loss) declines on which C3 rests. A reversal of the policy conclusion should have been disclosed.

The authors' response. The July 2025 user guide in the authors' repository documents several method and data changes (holding-sector removal, improved Irish FDI data, imputation for 'rest of world' countries); the paper does not discuss the revisions.

Why it matters, in detail. With the released numerators and denominators, the ETR fell 3.2 pp (-14%) over 1975-2023. Part of the gap with Bachas et al. (-24% to 2018) is the 2015/16 CIT splice (F7), which lifts the 2016+ ETR to about 19-20%. The 'stagnation' claim conflicts with a 17% rise in CIT/income. The 'quadrupling' of foreign profits since 1992 does hold on the post-tax series (4.2% -> 18.0%).

Evidence. GWZ2026 pp. 1, 2, 8; workbook DataF3-F4 cols B, C, E, I, J, AV, N; Bachas, Fisher-Post, Jensen & Zucman (2022) NBER w29819 Sec. 4.1 ('fell from 27% to 19%' for HICs); NBER w30673 p. 8. scripts/dataf34_table.py.

Why this severity. A secondary finding, presented as the Introduction's first main result, is overstated by about a factor of two and rests on a misattributed source figure; the profit-shifting series are unaffected.

The authors' response. None located.

Why it matters, in detail. Single country-years swing by $40-60bn or 20-70 pp of revenue, which signals input or allocation breaks in the source-country allocation (high-risk payments). The Fig. A3 relation cannot be distinguished from zero without the US, and part of it is mechanical, since loss = tau*S.

Evidence. GWZ2026 Tables A2-A3 (pp. 19-20) and Fig. A3 (p. 23); workbook data-FigureA3; scripts/fig_a3.py.

Why this severity. Affects secondary country-level results and the claim that countries cutting their tax rates saw falling losses; no effect on the global series.

The authors' response. The paper describes Fig. A3 as 'a mere correlation, not necessarily a causal relationship'.

Why it matters, in detail. The tax-loss ratio divides a numerator and a denominator built from different sources, so a source change moves it directly. The 2015/16 CIT break (about +11% on a level basis, measured by comparing vintages for 2016) is as large as the 2015-2023 relative fall in that ratio (-9.5%). The shifted share is less exposed, since one pipeline produces both its numerator and denominator. But the 2015 values embed TWZ's treatment of special purpose entities, unrecorded profits and offshore funds, which the pipeline changed. 2015 also anchors the whole pre-2015 backcast and the Table A4 break test.

Evidence. Workbook DataF3-F4 rows 44-46 (cols E, I); GTER 2024 vintage col E; repo raw-data/TWZ_original_results_2015.xlsx TableC4d vs output/Part1-profit-shifting-2016.xlsx 'Table U1'; GWZ2026 Table A2 'Rest of the world' row; repo code/Table U1.do l. 535-560; July 2025 guide step 9a. scripts/cit_splice.py, vintage_components.py, omf_floor.py.

Why this severity. The tax-loss part of C3 (8.4% -> 7.6%) crosses a denominator break of about 11%: rebasing 2015 by the 2016 ratio of new to old receipts gives about 7.6% for 2015, i.e. no decline. The share part also crosses a method change. A decline does survive within the new pipeline (share 37.1% -> 33.8% and tax loss 8.3% -> 7.6% over 2016-23), so the issue qualifies C3 rather than removing it.

The authors' response. Workbook note BA4 discloses that the denominator changed from non-haven to global receipts, and the guide documents other method changes; the change of CIT source at the splice is not discussed.

Why it matters, in detail. Global change = change in w_US*s_US + (1 - w_US)*s_nonUS; composition explains -0.89 of -2.39 pp. For US multinationals the haven share of profit-type return (Wright-Zucman list, excluding Puerto Rico) moves 43.0% -> 39.9% over 2015-23, but 45.9% -> 44.9% ex-mining. Flat haven dollars after 2018 are consistent with some policy effect relative to 2008-18 growth ($124bn -> $301bn), but the share metric mostly records non-haven profit growth. Attributing the decline to BEPS or the TCJA would need a counterfactual for haven dollars, which extrapolating the share trend does not supply.

Evidence. GWZ2026 pp. 2, 6-7, 9 and fn. 2; Tables A1-A2; workbook 'Wright-Zucman' sheet (cols H, O, P); BEA AMNE Part II Tables F1-F2 (2015r, 2020r-2022r, 2023p), Wright-Zucman haven list excluding Puerto Rico (scripts/bea_mining.py); shift-share from Auditor B (codex calculations.json), recomputed by the adjudicator.

Why this severity. The paper explicitly disclaims causal identification and calls the US analysis provisory. But once composition (about 37% of the global fall) and the non-haven denominator (about two-thirds of the 2015-23 fall in the US haven-booking share) are accounted for, little remains to support the policy reading.

The authors' response. The paper states that the trend differential 'does not capture the causal impact' and that the US analysis is 'provisory' (pp. 6-7).

Why it matters, in detail. S = sum over havens of [profits_f - wages_f * (local profits/local wages)]. A higher local benchmark (TWZ +45 pp: -18%), face-value haven data (-18% per TWZ; about -50% per the companion audit) or regression counterfactuals (CbC: 13-22%) lower 2015-type levels by roughly 18-65%. Since the backcast is anchored on 2015, proportional corrections would keep C1's shape but end it somewhere between the high teens and high twenties rather than at 34%. Corrections that vary over time (F2, F8) would also affect C3.

Evidence. GWZ2026 Secs. 2-3 (no post-2015 bounds); TWZ (REStud 2023) Table 2 cols 3-8 and Sec. 5.2; corpus entries offshore-havens-09 to -14; repo outputs (imputed profits 33-44% of shifted profits in 2016-22); Auditor B report 'Assumptions and Sensitivity'.

Why this severity. The level uncertainty is large: defensible estimates for a 2015-type year range from about a third of TWZ's value to the full value. But it applies to all years in a similar way, so it qualifies the 34% and 8% endpoints rather than the long-run rise.

The authors' response. TWZ (REStud 2023, Secs. 3.3 and 5.2) defend the counterfactual (excess profitability survives controls for capital, R&D and industry) and the upgrades (haven statistics are incomplete); Garcia-Bernardo & Jansky (2024) obtain comparable magnitudes.

Why it matters, in detail. Shifted profits move at least one-for-one with imputed profits R, and more for OECD havens, where R is netted from the local residual. R as a share of foreign profits halved (20.6% -> 10.9%) while recorded haven profits rose as reporting improved. Whether the net effect on S is neutral depends on whether the 2015 R measured truly missing income. Individual reconciliation choices in 2021-22 move the share by 0.2-1.9 pp (F2; setting the Netherlands' 2022 data gap to zero is worth only 0.16 pp after reallocation).

Evidence. TWZ replication TableA6 (World total: col D 306.39, col H 350.55, col E 1,703.2); repo outputs Part1-profit-shifting-2016...2022 'Table U1' (World unrecorded profits and foreign profits); raw-data/OECD_BOP_2025_update.csv (IRL inward from USA); July 2025 guide pp. 24-25; scripts/imputed_share.py. Blouin's 2025 conference slides (drh360.de, slide 25) describe $306bn of TWZ's 2015 estimate as imputed, consistent with TWZ Table A6 col D.

Why this severity. If part of the 2015 imputation was over-attribution (e.g. US statistics assigning to Irish or Dutch holding companies income earned elsewhere, a possibility TWZ concede), part of the measured decline reflects better statistics rather than less shifting. The evidence cannot settle this, so the issue qualifies C3's magnitude without showing a bias.

The authors' response. TWZ (REStud 2023, Secs. 4.2.3 and 5.2.4) defend raising haven data to partner data, because haven statistics are incomplete (e.g. Cayman excludes its offshore sector), and report an 18% face-value sensitivity for 2015; no post-2015 sensitivity is reported.

Why it matters, in detail. Before 2015, C2 = C1 x foreign share x a constant, so it carries no independent information, and any error in the backcast (F4, F5) passes one-for-one into C2. Allowing for the larger fall in statutory than effective rates raises mid-period losses by roughly 30-50% and flattens the 2010-15 jump.

Evidence. Workbook formula DataF3-F4!BA5 '=+$BA$45*(AV5*N5)/($AV$45*$N$45)', reproduced exactly by the adjudicator; statutory-rate variant from Auditor A (claude/scripts/backcast_sensitivity.py), not re-run by the adjudicator.

Why this severity. C2's endpoints barely move (1975 about 0.7-0.9%), but the mid-period shape, including the 2010-15 'leap', depends on the proxy, and readers may take the 1975-2015 losses as estimates rather than assumptions.

The authors' response. None located.

Why it matters, in detail. Intensity (S/F) fell 2.4 pp while the scale measures (S, S/global profits, S/global income) rose. Any 'measurable impact' concerns intensity relative to a trend counterfactual, not the amount shifted.

Evidence. GWZ2026 abstract and Sec. 3 p. 6; Table 1; workbook DataF3-F4 cols B, C, G, L.

Why this severity. The share of foreign profits is the paper's (and TWZ's) consistent measure of intensity, and the text discloses the dollar increase; this is a framing problem in the abstract rather than an error.

The authors' response. Sec. 3 (p. 6) reports the 69% dollar increase.

Why it matters, in detail. S_h = profits_f - wages_f*(local profits/local wages); with negative local profits, S_h exceeds foreign profits, and a near-zero Swiss benchmark classifies almost all Swiss foreign profits as shifted. Both point to mismatches between balance-of-payments-based foreign profits and national-accounts totals.

Evidence. Workbook Table U1 rows 36-38 (LUX local profits -4.17, profit/wage -0.28; CHE 0.013); DataF1 H5 and H9 (2023: -0.067 and -0.030); repo Part1-profit-shifting-2022 'Table U1' (same values); scripts/u1_havens.py.

Why this severity. The values are economically impossible for two large havens, but plausible fixes lower the 2023 level by only about 0.5-1.2 pp, slightly strengthening C3's decline and trimming C1's endpoint.

The authors' response. None located.

Why it matters, in detail. The 2022 denominator is understated by $26.2bn (0.75%); shifted profits are unaffected. A haven-gain calculation based on this workbook (F3) would use a 0.03% Swiss ETR.

Evidence. Workbook Table U1 L36/M36; repo output/Part1-profit-shifting-2022.xlsx 'Table U1' (CHE 26.179); GWZ2026 Table 1 (2022 receipts 3,488). These are the only cells where the repository's world totals differ from the workbook's in 2016-2022 (adjudicator comparison).

Why this severity. A transcription error that changes one year's revenue-loss ratio by 0.05 pp.

The authors' response. None; the repository version is correct.

Why it matters, in detail. The double deduction lowers Belgian and Swiss operating-unit equity income by about $1.4bn (2022). Table B10 reassigns it to non-EU havens, so global shifted profits change by far less than 0.1 pp.

Evidence. Repo code/FDI equity income operating units.do l. 31, 78, 125-153; raw-data/FDI_INC_AGGR_raw.csv (CHE 2019: T_D4P_F 100,214.0; T_D4S_F5 98,877.1; T_D4Q_FL 1,336.9; verified by the adjudicator); July 2025 guide p. 24 describes the deduction. The 2022 amounts are Auditor B's and were not re-run.

Why this severity. A conceptual slip worth about 0.05% of foreign profits in 2022 that moves small amounts between havens.

The authors' response. The guide (p. 24) describes removing interest income as intended; the double deduction is not addressed.

Why it matters, in detail. The post-2015 slope change is robustly significant, but the test compares 41 years of a series built from US data plus an assumed non-US rule with 9 directly estimated points that include a method splice (F7). The standard errors do not reflect generated-data or vintage uncertainty (F6).

Evidence. GWZ2026 Table A4 (p. 24), pp. 2 and 6; workbook DataF3-F4 cols AV, AE; scripts/table_a4.py.

Why this severity. The statistical break survives every check: HAC errors, disjoint windows (p about 3e-9 per Auditor B) and the backcast rule stated in the text (t = -4.1). The problems are presentation and interpretation slips.

The authors' response. The table note warns that post-2015 estimates rest on few observations.

Why it matters, in detail. Isolated formula and text-synchronisation errors left over from the data update; the 2015-2023 changes for six countries are mis-signed.

Evidence. GWZ2026 pp. 1-2, 5, 8, 15, 18-19, 21-22, 24; workbook TableA2 L44:L49 (formula '=K-C'); DataF3-F4 AP4/AP5; scripts/dataf34_table.py.

Why this severity. Labelling and transcription slips that do not affect the core findings.

The authors' response. None.

Why it matters, in detail. The 2023 country estimates (e.g. Ireland $184bn) cannot be verified, nor can it be checked whether the 2023 run shares the errors in F2 and F15, without re-running the code on fresh downloads.

Evidence. curl checks on 2026-10-25 (Dropbox page title 'File Deleted'; Atlas HTTP 404); GitHub API (repository, commits, tree); adjudicator comparison of repository outputs with the workbook; workbook DataF3-F4 AK53 = 0.54.

Why this severity. Everything except 2023 can be reproduced from public, if uncited, materials, and 2016-2022 reproduce exactly. But the headline year cannot be verified, and readers who follow the paper's own links find nothing.

The authors' response. The paper promises annual updates on the Atlas website.