One advisor for the whole world?

CAppendix Tables and Figures

C.1Tables

Table A7The pre-registered hypotheses, their tests, and their outcomes
Statement Test and correction Status Result
H1 Equity recommendations differ systematically across countries (country fixed effects in Eq. (3)). Per-country wild-bootstrap tests of the twenty contrasts, corrected as a single Holm family. Primary Supported. Seven of twenty contrasts survive the Holm correction, all negative, from Brazil (−11.0 pp) to Mexico (−2.7 pp). Section 5.2, Table A8.
H2 Advice does not move with country fundamentals (the U.S.-default anchor): \(\kappa=0\) in the calibration regression (11). Wild bootstrap over the twenty non-U.S. countries, HC2 standard errors alongside. Second altitudea Descriptive read consistent with compression. \(\hat{\kappa}=0.13\) (HC2 se 0.11), wild \(p=0.32\) against \(\kappa=0\), and \(\kappa=1\) is rejected. Section 4.1, Appendix B.4.
H3 The U.K. contrast is equivalent to zero within a \(\pm 5\) pp margin, consistent with shared Anglo-American norms. Two one-sided tests: supported when the 90 percent wild-bootstrap CI for \(b_{UK}\) lies inside the margin. Primary Supported. \(\hat{b}_{UK}=+0.7\) pp, 95 percent CI \([-0.8,+2.2]\), inside the margin. Section 5.2, Table A8.
H4 Asking in the local language rather than in English shifts advice, holding the country name fixed (\(\lambda\neq 0\) in Eq. (5)). Single pre-specified coefficient, reported uncorrected. Primary Supported. \(\hat{\lambda}=-1.3\) pp, \(p=0.007\). Section 5.3, Figure 3.
H5 Naming local instruments and pension vehicles shifts advice beyond the country name alone (\(\phi\neq 0\) in Eq. (4)). Single pre-specified coefficient, reported uncorrected. Primary Not supported at the 5 percent level. \(\hat{\phi}=-0.7\) pp, \(p=0.069\). Section 5.3, Figure 3.
H6 The age-55 tilt toward bonds is smaller where pension replacement rates are higher (age-55 \(\times\) standardized replacement rate in Eq. (7)). Single pre-specified interaction, reported uncorrected. Secondary Not supported. −0.02 pp, \(p=0.964\). Section 6.2, Table A20.
H7 Country effects covary positively with the country’s GPS risk-taking score. Bivariate slope over the twenty non-U.S. countries, wild bootstrap over countries. Second altitudea Descriptive read unsupported. +0.77 pp per standard deviation, wild \(p = 0.272\). Section 5.5, Table A14.
H8 Country effects covary positively with the country’s financial literacy score. Bivariate slope over the twenty non-U.S. countries, wild bootstrap over countries. Second altitudea Descriptive read unsupported. +0.89 pp per standard deviation, wild \(p = 0.178\). Section 5.5, Table A14.
Table A7The pre-registered hypotheses, their tests, and their outcomes (continued)
Statement Test and correction Status Result
H9 The rate at which the explanation text invokes a given rationale differs across countries (Section 7). Country fixed effects per admitted category, one Holm family per category. Second altitudeb Supported where the design gives it variation. Local-context references are elevated in all twenty non-U.S. countries (+29 to +78 pp, nineteen Holm-surviving) and inflation reasoning in the high-inflation countries (Turkey +56 pp). Pension rates are flat, and horizon sits at its 100 percent ceiling. Section 7, Table 3.
H10 Version C raises the rate of local financial references in the explanation text relative to Version B (\(\phi_{\text{text}}>0\), Section 7). Single pre-specified coefficient, reported uncorrected. Second altitudeb Supported. +23.3 pp, \(p<0.001\), against the −0.7 pp allocation effect of H5. Section 7, Figure 10.
H11 A stated risk-tolerant investor is recommended more equity than a stated risk-averse one (\(\beta_{R}>0\) in Eq. (6)). Single coefficient, directional prediction tested with a two-sided \(p\)-value, reported uncorrected. Secondary confirmatory Supported. \(\hat{\beta}_{R}=+46.3\) pp, \(p<0.001\). Section 6, Table A18.
H12 A stated generous pension moves the recommendation (\(\beta_{P}\neq 0\) in Eq. (6), two-sided by design). Single coefficient, two-sided, reported uncorrected. Secondary confirmatory Not supported, and precisely so. \(\hat{\beta}_{P}=-1.0\) pp, \(p=0.107\). Section 6, Table A18.
H13 The stated horizon moves the recommendation, against the CRRA prediction of exact irrelevance (\(\beta_{H}\neq 0\) in the horizon analog of Eq. (6)). Single coefficient, reported uncorrected. Secondary confirmatory Supported. \(\hat{\beta}_{H}=+39.9\) pp, \(p<0.001\). Section 6, Table A18.
Notes: Notes: The table records the complete set of hypotheses in the pre-registration, independently of which ones the text discusses. The Result column gives the verdict and the headline estimate, and the exhibits it cites report confidence intervals and the full set of estimates. Estimates are in percentage points of the recommended equity share. Unless noted otherwise, \(p\)-values come from the wild cluster bootstrap (Webb six-point weights, 9,999 replications, clustered by model, \(G=13\)), and the H3 confidence interval is a 95 percent interval from the same procedure. All reported \(p\)-values are two-sided: the directional predictions (H7, H8, H11, H13) are stated as directional but tested with two-sided \(p\)-values, a conservative choice. The equivalence hypothesis H3 is the exception, tested by two one-sided tests as its statement requires. The second-stage tests (H2, H7, H8) instead bootstrap over the twenty non-U.S. countries.
(a) Registered as pre-specified but descriptive at filing.
(b) Registered as pre-specified but descriptive at filing, with the coding instrument frozen and its pre-registered reliability gates reported by the judge-validation amendment of record (filed 2026-07-15) before any coefficient of the explanation-text regression (Section 7) was estimated. Because it post-dates the pilot, these are pre-specified descriptive tests rather than blind confirmatory ones. Three of the four categories passed both gates. The horizon category is degenerate at 100.0 percent prevalence.
Table A8Country means and country effects on the recommended equity share (Question 1, Version B, age 40)
Recommended share Country effect relative to the U.S.
Country Mean SD Estimate 95% CI \(p\) (wild) \(p\) (Holm)
Brazil 0.46 0.14 −11.0 \([-15.0,\ -6.9]\) \(<\)0.001 0.002
Italy 0.48 0.14 \(-8.6\) \([-11.9,\ -5.3]\) \(<\)0.001 \(<\)0.001
Poland 0.50 0.14 \(-7.4\) \([-10.8,\ -3.9]\) \(<\)0.001 0.011
France 0.51 0.14 \(-5.4\) \([ -7.5,\ -3.4]\) \(<\)0.001 \(<\)0.001
China 0.52 0.10 \(-4.4\) \([ -8.8,\ -0.2]\) 0.041 0.371
Spain 0.53 0.12 \(-4.4\) \([ -6.3,\ -2.5]\) \(<\)0.001 \(<\)0.001
Switzerland 0.53 0.14 \(-4.3\) \([ -7.6,\ -1.1]\) 0.016 0.200
Netherlands 0.53 0.14 \(-4.2\) \([ -6.8,\ -1.7]\) 0.004 0.049
Germany 0.53 0.14 \(-4.0\) \([ -7.7,\ -0.4]\) 0.034 0.335
Turkey 0.54 0.11 \(-3.1\) \([ -6.5,\ -0.1]\) 0.044 0.371
Mexico 0.54 0.14 \(-2.7\) \([ -4.2,\ -1.2]\) \(<\)0.001 0.003
Finland 0.55 0.15 \(-2.4\) \([ -5.0,\ 0.4]\) 0.086 0.602
India 0.55 0.12 \(-1.9\) \([ -3.3,\ -0.4]\) 0.015 0.200
Canada 0.56 0.13 \(-1.1\) \([ -2.1,\ -0.1]\) 0.029 0.317
South Africa 0.56 0.12 \(-0.9\) \([ -2.4,\ 0.5]\) 0.219 1.000
Sweden 0.57 0.13 \(-0.2\) \([ -3.7,\ 3.4]\) 0.931 1.000
United States 0.57 0.13 omitted baseline
Australia 0.58 0.12 \(0.6\) \([ -0.7,\ 1.9]\) 0.343 1.000
United Kingdom 0.58 0.13 \(0.7\) \([ -0.8,\ 2.2]\) 0.352 1.000
Japan 0.58 0.14 \(1.2\) \([ -2.1,\ 4.8]\) 0.492 1.000
South Korea 0.58 0.11 \(1.4\) \([ -2.8,\ 5.7]\) 0.496 1.000
Notes: Notes: The first two columns report the mean and standard deviation of the recommended equity share by country on the primary cell (Question 1, Version B, age 40), pooled across the thirteen models with fifty repetitions per country–model cell. Each country has 648 to 650 responses after the pre-registered exclusion of responses without a parsed numeric recommendation or truncated at the token limit. The remaining columns report the country fixed effects \(b_{c}\) from equation (3), estimated on the same cell (\(N = 13{,}639\)) with model fixed effects and the United States as the omitted baseline. Estimates and confidence intervals are in percentage points of the recommended equity share, and countries are ordered by point estimate. \(p\) (wild) is the wild cluster bootstrap \(p\)-value (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\)), and the 95 percent confidence interval comes from the same procedure. \(p\) (Holm) applies the pre-registered Holm step-down correction across the family of twenty contrasts. Figure 2 plots these estimates.
Table A9Country effects on the mortgage and inflation questions (Version B, age 40)
Mortgage (Question 2) Inflation (Question 3)
Country Estimate 95% CI \(p\) (wild) \(p\) (Holm) Estimate 95% CI \(p\) (wild) \(p\) (Holm)
Japan −28.6 \([-42.5,\ -15.0]\) \(<\)0.001 0.002 −11.3 \([-13.7,\ -8.9]\) \(<\)0.001 \(<\)0.001
China −22.7 \([-32.5,\ -13.7]\) \(<\)0.001 0.002 \(-7.2\) \([ -9.6,\ -4.7]\) \(<\)0.001 \(<\)0.001
Switzerland −21.6 \([-29.8,\ -14.0]\) \(<\)0.001 0.002 \(-5.0\) \([ -7.3,\ -2.7]\) \(<\)0.001 0.006
Germany −21.1 \([-30.0,\ -12.0]\) \(<\)0.001 0.002 \(-2.0\) \([ -4.1,\ -0.1]\) 0.039 0.473
Spain −20.4 \([-28.6,\ -12.3]\) \(<\)0.001 0.002 \(-4.0\) \([ -6.6,\ -1.3]\) 0.007 0.091
Italy −19.5 \([-26.7,\ -12.5]\) \(<\)0.001 0.002 \(-5.0\) \([ -7.3,\ -2.8]\) \(<\)0.001 0.002
Finland −19.1 \([-27.1,\ -11.3]\) \(<\)0.001 0.002 \(-2.5\) \([ -5.2,\ -0.1]\) 0.041 0.473
South Korea −17.3 \([-24.4,\ -10.1]\) \(<\)0.001 0.002 \(-6.9\) \([-10.3,\ -3.5]\) 0.001 0.020
Australia −17.1 \([-24.8,\ -9.8]\) \(<\)0.001 0.003 \(-0.2\) \([ -2.2,\ 1.9]\) 0.862 1.000
Poland −16.2 \([-21.0,\ -11.5]\) \(<\)0.001 0.002 \(-1.8\) \([ -5.3,\ 1.6]\) 0.269 1.000
Netherlands −16.0 \([-23.2,\ -9.9]\) \(<\)0.001 0.002 \(-6.5\) \([ -8.9,\ -4.2]\) \(<\)0.001 0.002
France −15.7 \([-24.6,\ -6.6]\) \(<\)0.001 0.003 \(-4.3\) \([ -6.9,\ -1.8]\) 0.003 0.041
United Kingdom −14.3 \([-21.5,\ -8.1]\) \(<\)0.001 0.003 \(-1.5\) \([ -3.2,\ 0.1]\) 0.064 0.636
Canada −12.3 \([-18.4,\ -6.6]\) \(<\)0.001 0.003 \(1.6\) \([ -0.4,\ 3.6]\) 0.100 0.799
Sweden −10.0 \([-16.5,\ -4.1]\) 0.001 0.006 \(-0.1\) \([ -3.1,\ 2.7]\) 0.930 1.000
Indiaa \(-3.8\) \([ -6.7,\ -0.8]\) 0.015 0.060 \(0.8\) \([ -2.5,\ 3.9]\) 0.618 1.000
South Africaa \(-2.4\) \([ -4.3,\ -0.7]\) 0.002 0.008 \(2.3\) \([ -0.3,\ 4.9]\) 0.082 0.739
Mexicoa \(-1.9\) \([ -5.3,\ 1.7]\) 0.262 0.787 \(0.1\) \([ -3.3,\ 3.4]\) 0.973 1.000
Brazila \(-1.7\) \([ -6.2,\ 3.0]\) 0.452 0.904 \(2.6\) \([ -1.0,\ 5.9]\) 0.139 0.972
Turkeya \(-1.6\) \([ -8.4,\ 5.3]\) 0.628 0.904 \(-0.9\) \([ -4.5,\ 2.7]\) 0.596 1.000
Notes: Country fixed effects from the specification of equation (3), with model fixed effects and the United States as the omitted baseline, in percentage points. The mortgage columns report the recommended fraction to invest in a broad stock market index fund rather than prepay a 4 percent mortgage (\(N = 13{,}619\)). The inflation columns report the recommended fraction of savings to move out of a 1 percent savings account in the face of 3 percent inflation (Question 3, the registration’s Q4, \(N = 13{,}637\)). The numeric fraction is the primary Question 3 outcome because the collection produced no refusals and the pre-specified refusal contingency never bound (Section 3.4). Countries are ordered by the mortgage estimate. Inference is the model-level wild cluster bootstrap (Webb six-point weights, 9,999 replications, \(G = 13\)), and \(p\) (Holm) applies the step-down correction across the twenty contrasts of each question, which form separate families as pre-registered. Figure 4 plots these estimates, and Table A13 reports the categorical companion to the inflation columns.
(a) Pre-specified premise-acceptance group. The prevailing mortgage rate at the filing date exceeded twice the prompted 4 percent, so the mortgage contrasts for these countries test whether the model accepts or corrects the premise rather than whether it calibrates to local fundamentals.
Table A10The mortgage question without the stated rate: country effects
Prevailing 4% stated No rate Change in gap
Country rate (%) Estimate Estimate 95% CI \(p\) (Holm) Estimate \(p\) (Holm)
Panel A: Country effects relative to the United States
Switzerland 1.4 −17.2 1.5 \([-2.9,\ 6.2]\) 1.000 18.7 0.005
Spain 2.8 −18.9 −5.8 \([-11.0,\ -0.6]\) 0.267 13.1 0.004
Finland 2.8 −17.9 −5.9 \([-11.5,\ -0.4]\) 0.269 12.1 0.010
Japan 2.8 −26.9 3.0 \([-0.6,\ 7.0]\) 0.728 29.9 0.004
China 3.1 −25.3 −19.8 \([-33.1,\ -7.5]\) 0.005 5.5 0.468
France 3.1 −16.7 −0.6 \([-4.4,\ 3.2]\) 1.000 16.1 0.006
Sweden 3.4 −5.4 −1.3 \([-6.0,\ 3.8]\) 1.000 4.1 0.468
Italy 3.4 −20.2 −5.8 \([-10.5,\ -1.2]\) 0.176 14.3 0.005
Netherlands 3.6 −14.5 −3.2 \([-8.1,\ 1.7]\) 0.979 11.4 0.002
Canada 3.8 −12.3 −11.2 \([-18.4,\ -4.6]\) 0.015 1.0 1.000
Germany 3.8 −21.2 −5.4 \([-11.1,\ 0.2]\) 0.431 15.8 0.006
South Korea 4.3 −17.6 −10.7 \([-16.2,\ -5.8]\) 0.007 7.0 0.070
United Kingdom 4.8 −14.3 −10.4 \([-17.5,\ -4.4]\) 0.016 3.9 0.468
Australia 5.7 −17.1 −18.6 \([-29.5,\ -8.3]\) 0.016 −1.5 1.000
Poland 6.0 −12.0 −20.9 \([-32.7,\ -9.0]\) 0.007 −8.9 0.468
United States 6.2 baseline
Indiaa 8.1 −3.8 −10.8 \([-17.6,\ -5.0]\) 0.014 −7.1 0.303
South Africaa 10.2 −2.4 −26.7 \([-39.9,\ -13.3]\) 0.002 −24.3 0.010
Brazila 10.7 −1.8 −28.7 \([-43.4,\ -14.2]\) 0.002 −26.9 0.009
Mexicoa 11.4 −1.5 −20.1 \([-34.3,\ -6.2]\) 0.031 −18.6 0.070
Turkeya 36.2 3.5 −5.7 \([-16.6,\ 4.8]\) 1.000 −9.2 0.584
Panel B: Pooled estimates
Estimate 95% CI \(p\) (wild)
Field check: asking for the assumed rates 2.79 \([0.91,\ 4.66]\) 0.008
Gap change per point of rate above 4% −1.02 \([-5.06,\ -0.33]\) \(<\)0.001
without Turkey −4.63 \([-5.93,\ -3.54]\) \(<\)0.001
Per point of assumed mortgage rate −1.59 \([-2.79,\ -0.40]\) 0.017
Per point of assumed inflation 0.78 \([-0.19,\ 1.91]\) 0.102
sum (zero if only the real rate matters) −0.82 \([-2.25,\ 0.33]\) 0.135
Notes: Panel A: country fixed effects from equation (3) with model fixed effects, in percentage points of the fraction invested rather than used to prepay. “4% stated” is the English Question 2 cell of the main collection (July 2026). “No rate” deletes the rate clause (November 2026). “Change in gap” is the difference of the two contrasts from the stacked sample with model-by-collection fixed effects, which removes drift common to all countries. Wild cluster bootstrap over models (Webb, 9,999 replications), Holm over the twenty non-U.S. countries in each column. Prevailing rate from the pre-registered rate supplement. Panel B, in percentage points: the field check against the same question without the assumption fields (five countries, country and model fixed effects), the change in gap regressed on the prevailing rate minus 4 across countries (HC2, wild bootstrap over countries), and the share regressed on the assumed rates within countries (country and model fixed effects). Exploratory, post-filing.
(a) Premise-acceptance group.
Table A11The mortgage question without the stated rate: assumed rates by country
Mortgage rate (%) Inflation (%)
Country Prevailing Assumed Model range 2024 Assumed Model range
Switzerland 1.4 2.0 1.5–2.7 1.1 1.4 1.0–2.0
Spain 2.8 3.3 3.0–3.6 2.8 2.2 2.0–2.5
Finland 2.8 3.5 2.5–4.1 0.9 2.1 2.0–2.5
Japan 2.8 1.1 0.6–1.5 2.6 1.8 1.3–2.5
China 3.1 4.1 3.7–4.7 0.4 2.1 1.5–2.5
France 3.1 2.8 1.5–3.5 2.4 2.1 2.0–2.5
Sweden 3.4 4.0 3.4–4.5 2.8 2.1 2.0–2.5
Italy 3.4 3.4 3.0–3.6 1.2 2.1 2.0–2.5
Netherlands 3.6 3.7 3.4–4.1 3.2 2.2 2.0–2.5
Canada 3.8 5.0 4.3–5.5 2.4 2.3 2.0–2.5
Germany 3.8 3.4 2.7–3.8 2.4 2.1 2.0–2.5
South Korea 4.3 4.1 3.5–4.5 2.3 2.4 2.1–2.7
United Kingdom 4.8 4.5 4.1–5.0 2.6 2.4 2.0–2.9
Australia 5.7 6.0 4.7–6.4 3.2 3.0 2.5–3.5
Poland 6.0 6.7 4.5–7.5 3.8 3.7 2.6–4.7
United States 6.2 5.1 4.0–6.5 2.9 2.6 2.0–3.0
India 8.1 8.4 7.2–9.0 4.8 5.3 5.0–6.0
South Africa 10.2 10.9 8.9–11.8 4.6 5.3 5.0–5.6
Brazil 10.7 9.7 6.7–11.3 4.5 4.5 4.0–4.9
Mexico 11.4 9.5 6.5–11.5 4.7 4.3 3.9–5.0
Turkey 36.2 26.6 15.2–41.8 58.3 40.6 26.8–61.0
Notes: Assumed rates are the values each model reports in the assumed_mortgage_rate and assumed_inflation_rate fields after its recommendation in the no-rate cell. “Assumed” is the mean over all responses, and “Model range” is the smallest and largest of the thirteen model means. Every model reported the rates as decimals. The prevailing mortgage rate is from the pre-registered rate supplement, and inflation is 2024 consumer price inflation from the Global Macro Database. Countries are ordered by the prevailing mortgage rate.
Table A12The mortgage question without the stated rate: by model
Share Field check (pp) Assumed (%)
Model invested (%) Diff. (SE) Mortgage Inflation Corr.
claude-haiku-4.5 70.3 0.8 (0.7) 4.4 2.8 0.91
claude-opus-4.6 56.2 −2.1 (3.2) 5.3 3.1 0.93
claude-opus-4.8 63.4 1.4 (4.0) 5.4 2.8 0.94
claude-sonnet-4.6 69.6 −0.2 (2.1) 5.3 3.0 0.96
deepseek-v4-flash 67.1 6.8 (2.6) 4.8 2.7 0.91
deepseek-v4-pro 72.1 5.5 (4.2) 5.3 2.9 0.91
gemini-3-flash-preview 75.6 −1.2 (4.7) 5.2 2.7 0.85
gemini-3.1-pro-preview 64.0 5.9 (4.1) 5.2 2.9 0.83
gemini-3.5-flash 65.8 4.9 (2.4) 5.2 2.7 0.88
gpt-5-mini 70.8 5.8 (1.6) 4.7 2.6 0.90
gpt-5.5 52.1 0.8 (3.0) 5.4 2.8 0.92
grok-4.3 66.5 5.4 (1.5) 4.9 2.7 0.92
mistral-medium-3-5 73.8 2.5 (0.7) 4.6 2.6 0.90
Notes: “Share invested” is the mean recommended fraction invested in the no-rate cell across all twenty-one countries. “Field check” is the difference in percentage points between the no-rate cell and the same question without the assumption fields in the five field-check countries, from a regression with country fixed effects within each model, with standard errors clustered by country. “Assumed” gives the model’s mean assumed rates over the twenty countries other than Turkey. “Corr.” is the correlation across the twenty-one countries between the log of the model’s mean assumed mortgage rate and the log of the prevailing rate. Descriptive.
Table A13Primary-instrument recommendation on the inflation question by country (Question 3, Version B, age 40)
Country Equities Infl.-linked Gold Real estate Other
Brazil 13.1 86.7 – – 0.2
Turkey 17.5 66.0 16.3 – 0.2
Mexico 41.9 58.1 – – –
Poland 61.1 38.6 – – 0.3
South Africa 82.0 16.3 – – 1.7
India 88.5 7.8 3.7 – –
Italy 93.2 6.8 – – –
South Korea 94.0 6.0 – – –
Spain 96.4 3.6 – – –
United States 96.8 2.9 – – 0.3
Canada 97.2 2.6 – – 0.2
France 97.8 2.2 – – –
United Kingdom 98.0 2.0 – – –
Australia 98.3 1.5 – – 0.2
Switzerland 95.2 0.5 – 4.3 –
Netherlands 99.5 0.3 – – 0.2
Japan 99.7 0.2 – – 0.2
Sweden 99.8 0.2 – – –
China 96.3 0.2 – – 3.5
Germany 88.6 0.2 – 11.2 –
Finland 100.0 – – – –
Notes: Share of responses (in percent) recommending each instrument category as the primary way to protect savings against inflation (Question 3, Version B, age 40), pooled across the thirteen models. Each response is assigned a single primary instrument by a deterministic multilingual keyword classifier, and the five categories sum to 100. “Other” collects mixed, bond, and money-market funds and the rare response with no reported instrument. Countries are ordered by the combined inflation-protection share (inflation-linked bonds plus gold), descending. A dash denotes a category recommended in under 0.05 percent of responses. This is the categorical companion to the numeric outcome of Table A9, Panel B, discussed in Section 5.4.
Table A14Country effects and country characteristics, by question
Allocation (Q1) Mortgage (Q2) Inflation (Q3)
Characteristic Slope \(p\) \(R^{2}\) Slope \(p\) \(R^{2}\) Slope \(p\) \(R^{2}\)
@@inputTables/tab_country_correlates_body.tex
Notes: Each cell reports a separate regression, across the twenty non-U.S. countries, of the country fixed effects on one country characteristic. The dependent variables are the country effects of Appendix Tables A8 and A9 (Version B, age 40, United States omitted), in percentage points. Continuous characteristics are standardized within the sample, so a slope is the change in the country effect per standard deviation of the characteristic. Interest rates, inflation, and GDP per capita enter in logs. The English-prompt indicator is a dummy. The home-index rows cover eighteen countries due to a lack of data on the MSCI country series for China and Mexico. The benchmark row uses the benchmark share for each question, Benchmark S for Question 1 (Section 4.1) and \(\alpha^{Q2}_{c}\) for Question 2 (Section 4.2), in deviation from the U.S. value. Question 3 has no benchmark. \(p\) in brackets is the wild bootstrap \(p\)-value over countries (Webb weights, 9,999 replications). Holm-adjusted \(p\)-values within each question are reported in the text. The rows on GPS risk-taking and financial literacy are the pre-registered H7 and H8 slopes in standardized units, and the Question 1 benchmark row is the calibration slope \(\kappa\) in standardized units. All other cells are post-filing exploratory analyses.
Table A15Collection-robustness cells: temperature bookends and numeral magnitude
Country Mean (Version B) Mean (B-USD) Shift 95% CI \(p\) (wild)
Germany 0.53 0.54 \(1.2\) \([ -0.4,\ 2.7]\) 0.130
Japan 0.58 0.55 \(-2.8\) \([ -5.4,\ -0.4]\) 0.010
Sweden 0.57 0.58 \(1.2\) \([ -0.8,\ 3.4]\) 0.255
Notes: The table reports the two pre-registered collection-robustness cells (Section 5.6), re-collected on July 6, 2026 alongside the main corpus (Section 3.4). Panel A re-estimates the country fixed effects of equation (3) separately at each sampling temperature on the five bookend countries (Question 1, Version B, age 40, United States omitted), in percentage points of the recommended equity share. The \(T=0.7\) column uses the primary-collection cells and the bookend columns use the re-collected cells, so the \(T=0.7\) estimates differ from Appendix Table A8 only in the estimation sample, which is restricted to these five countries. GPT-5-mini, GPT-5.5 and Claude Opus 4.8 do not accept a temperature parameter (Section 3.4), so their bookend cells re-collect the provider default, and Panel A estimates the contrasts on the other ten models (\(G=10\)). The dispersion rows report the mean across country–model cells of the within-cell standard deviation of the recommended share, in fraction units, separately for the ten models and for the three without a temperature parameter. Panel B compares Version B, whose stake is a round native-currency amount, with an otherwise identical prompt stating the amount as $50,000, for three pre-registered countries with country and language held fixed. The mean columns are recommended equity shares, and the shift is the B-USD minus Version-B difference in percentage points, estimated with model fixed effects. All \(p\)-values and confidence intervals come from the wild cluster bootstrap (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\) in Panel B and \(G = 10\) in Panel A). The pooled estimates quoted in Section 5.6 average over the countries shown here.
Table A16Country contrasts in the exploratory re-collections: stake at purchasing-power parity, translated instructions, and the elicitation probe
Estimate 95% CI \(p\) (wild)
Stake restated at purchasing-power parity
Pooled shift, PPP minus filed stake (20 countries) \(1.0\) \([ 0.5,\ 1.7]\) \(<\)0.001
Per doubling of the real stake \(-1.2\) \([ -2.0,\ -0.6]\) \(<\)0.001
net of a common shift \(-0.9\) \([ -1.4,\ -0.4]\) 0.003
Translated minus English instructions (October minus July)
All twenty-one countries \(-1.2\) \([ -2.7,\ 0.2]\) 0.095
English-prompting countries (identical prompts) \(-0.6\) \([ -1.8,\ 0.4]\) 0.257
Non-English countries \(-1.4\) \([ -3.2,\ 0.1]\) 0.084
Language component (non-English minus English) \(-0.8\) \([ -1.8,\ 0.2]\) 0.130
Table A17Model-tracking waves: stability of the advice across model generations
All models New models Carried-over models
July September September July September
Panel A: Level and composition of the variance
Models 13 21 14 7 7
Responses kept 13,645 11,011 7,339 7,347 3,672
Repetitions per cell 50 25 25 50 25
Mean equity share (%) 54.9 49.4 45.8 55.4 56.6
Variance share, model (%) 78.0 70.5 42.9 73.9 76.7
Variance share, country (%) 2.8 4.7 14.8 1.5 1.5
Variance share, model \(\times\) country (%) 4.3 6.7 10.6 2.1 5.1
Variance share, within cell (%) 14.9 18.1 31.7 22.5 16.7
Range of country means, pooled over models (pp) 9.4 10.3 13.1 7.0 6.5
SD of country means net of noise, average over models (pp) 3.2 3.5 3.8 2.2 2.9
Non-U.S. cells below the model’s own U.S. cell (%) 67 77 79 62 74
Slope of country means on \(\alpha^{S}_{c}\) 0.118 0.081 0.080 0.073 0.084
(robust standard error) (0.091) (0.113) (0.143) (0.070) (0.054)
Panel B: Carried-over models, same identifier
Mean, July Mean, September Change (se) \(\Delta\) SD of country means Largest cell change
DeepSeek V4 Flash 46.3 54.8 +8.46 (1.08) +2.47 15.2
Grok 4.3 54.5 57.0 +2.43 (0.37) −0.70 5.4
GPT-5-mini 37.2 37.5 +0.22 (0.26) +0.09 2.2
Claude Haiku 4.5 72.5 72.6 +0.06 (0.12) +0.02 1.8
Mistral Medium 3.5 70.1 70.0 −0.07 (0.05) −0.08 0.6
Gemini 3 Flash 59.4 59.0 −0.44 (0.57) +2.53 9.2
DeepSeek V4 Pro 47.4 45.0 −2.31 (0.52) +0.44 6.6
Panel C: Slope on release date, percentage points per year (27 models, 34 points)
Estimate Cluster se \(p\) (wild)
SD of country means net of noise +0.98 0.53 0.19
with wave fixed effects +0.92 0.59 0.30
Notes: Question 1, age 40, one English prompt per country (Version B in the six English-prompting countries, Version B-cross in the other fifteen), temperature 0.7. The July columns are the main collection of July 2–4, 2026 cut to this slice, and the September columns the tracking wave of September 19–20, 2026. New models are the fourteen September models absent from the paper panel. Carried-over models are the seven paper-panel models re-queried under the same identifier. Panel A: variance shares are sums of squares relative to the total on response-level data, as in Section 5.1. With 25 rather than 50 repetitions per cell, sampling noise raises the between-cell shares by a negligible amount. The SD of country means net of noise subtracts the average sampling variance of the cell means, \(\overline{s^{2}_{mc}/n_{mc}}\), from the variance of a model’s twenty-one country means. The benchmark slope regresses the twenty-one country means, pooled over the models in the column, on the standalone benchmark of Appendix B.1. It carries no model fixed effects, so it is a descriptive analog of \(\kappa\) and its changes across columns are compositional. Panel B: mean recommended equity shares in percent. The change is the mean of the twenty-one within-country changes in cell means, with its standard error across countries. The SD change and the largest absolute cell change measure reallocation across countries at a given mean. Seven comparisons, uncorrected. Panel C: each model-by-wave outcome regressed on the model’s OpenRouter listing date, with carried-over models entering once per wave at the same date, standard errors clustered by model, and wild cluster bootstrap \(p\)-values (Webb weights, 9,999 replications). Release date is close to collinear with the wave, so only the wave-fixed-effects row separates vintage from collection date. Post-filing exploratory deviation recorded in the design lock, in no Holm family.
Table A18Stated-investor arm: cell means, pre-registered effects, and exploratory balance-sheet probes
Contrast Estimate 95% CI \(p\) (wild)
Pension bond-like annuity (Merton: +) \(4.8\) \([ 0.1,\ 9.6]\) 0.042
Pension equity-invested (Merton: \(-\)) −16.9 \([-21.8,\ -12.1]\) \(<\)0.001
Fund currency-hedged (at mean \(\sigma_c\)) \(0.4\) \([ -1.5,\ 2.2]\) 0.659
Fund unhedged (at mean \(\sigma_c\)) \(-1.3\) \([ -3.2,\ 0.4]\) 0.149
Hedged \(\times \sigma_c\) (Merton: +) \(23.8\) \([ 6.1,\ 42.1]\) 0.010
Unhedged \(\times \sigma_c\) (Merton: \(-\)) \(36.0\) \([ 15.8,\ 55.5]\) 0.002
Stake = all liquid wealth (US only) −23.3 \([-30.4,\ -16.1]\) \(<\)0.001
Stake = half of wealth (US only) \(1.4\) \([ -0.4,\ 3.3]\) 0.109
Stake = a third of wealth (US only) \(2.5\) \([ -0.1,\ 4.9]\) 0.057
Linear B/L slope, per step (US only) \(7.6\) \([ 5.4,\ 9.8]\) \(<\)0.001
Table A19Stated-investor effects by country
Pre-registered cells Exploratory probes
Country Risk tolerance Horizon Generous pension Pension in stocks Pension annuity Hedged \(-\) unhedged
Japan 48.9 36.9 0.5 −20.0 1.7 0.5*
Turkey 49.7 35.2 0.3 −16.3† 6.3* −2.4*
South Korea 48.2 34.6 0.6 −19.9 3.0 0.3
Spain 48.0 41.8 1.0 −16.2† 4.9 1.3
Poland 48.2 44.5* 0.9 −13.5† 6.1 3.1
Italy 47.8 42.5 1.4 −14.2† 5.7 3.0
China 48.8 34.1 0.6 −19.0 5.3 3.1
Brazil 48.3 37.6 1.0 −13.8† 7.5* 4.0
Mexico 49.9 34.3 0.9 −19.7* 3.9 1.2
France 47.1 44.9* 1.2 −15.4† 5.1 2.0
Germany 47.7 43.8* 0.5 −14.8† 5.1 0.9*
India 47.5 38.0 0.8 −18.9* 4.1 2.7
South Africa 46.7* 38.5 −0.1 −19.0 4.9 2.1
Switzerland 47.4 44.2† 0.6 −16.5† 5.8* 2.4
Netherlands 47.1 41.5* 0.9 −15.0† 5.2 1.9*
Canada 47.3* 41.1 1.2 −20.6 3.9 0.9*
United States 48.2 35.8 0.9 −21.7 3.5 3.1
United Kingdom 48.3 43.1* 0.9 −15.4† 3.4 1.2*
Australia 47.9 36.5 0.9 −19.9 3.4 1.2*
Finland 48.2 42.2* 1.0 −14.5† 4.7 2.7
Sweden 48.5 46.0† 0.6 −10.8† 3.9 0.8*
Mean of country effects \(48.1\) \(39.9\) \(0.8\) −16.9 \(4.6\) \(1.7\)
Range \(46.7\) to \(49.9\) \(34.1\) to \(46.0\) \(-0.1\) to \(1.4\) −21.7 to −10.8 \(1.7\) to \(7.5\) \(-2.4\) to \(4.0\)
Median CI half-width \(\pm 5.4\) \(\pm14.4\) \(\pm 1.4\) \(\pm 5.3\) \(\pm 5.0\) \(\pm 1.6\)
Countries differing from U.S. (Holm) 0 of 20 2 of 20 0 of 20 12 of 20 0 of 20 0 of 20
Notes: Effect of each stated-investor treatment on the recommended equity share, in percentage points, estimated separately for each country (Question 1, age 40, statements in English). Risk tolerance is the risk-tolerant minus risk-averse difference pooled over the two pension statements, and generous pension the generous minus limited difference pooled over the two risk statements, both from the four crossed cells of equation (6). Horizon is the twenty-year minus one-year difference on the two horizon cells. Pension in stocks and pension annuity are the probe cells of Section 6.3 against the country’s own no-statement reference, and hedged \(-\) unhedged is the difference between the two hedging cells, estimated on those cells alone. Each column comes from one regression that interacts the treatment with country indicators and absorbs model fixed effects, so each entry is the treatment effect within that country. \(^{\dagger}\) marks a country whose effect differs from the U.S. effect at the five percent level after Holm correction within the column’s family of twenty contrasts, and \(^{*}\) a difference significant at five percent before correction only. The summary rows give the unweighted mean and range of the twenty-one country effects, the median half-width of the per-country 95 percent confidence intervals, and the number of Holm-significant contrasts. All inference is the wild cluster bootstrap (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\)). No joint test of equal effects across countries is reported, because the model-clustered covariance of twenty interaction coefficients has rank at most twelve with thirteen clusters. The per-country breakdown is a post-filing exploratory deviation recorded in the design lock.
Table A20The age gradient and its interaction with pension generosity (Question 1, Version B)
(1) (2) (3)
No interaction Continuous \(\tilde\rho\) Median split
Age 30 (vs. 40) 6.24 6.24 6.24
(0.57) (0.57) (0.57)
Age 55 (vs. 40) -11.13 -11.13 -11.34
(1.90) (1.90) (1.96)
\(\times\) std. replacement rate (\(\psi\)) -0.02
(0.33)
\(\times\) above-median pension 0.44
(0.61)
Country fixed effects Yes Yes Yes
Model fixed effects Yes Yes Yes
Wild \(p\): Age 30 0.000 0.000 0.000
Wild \(p\): Age 55 0.000 0.000 0.000
Wild \(p\): interaction 0.964 0.489
Observations 40,910 40,910 40,910
Model clusters (\(G\)) 13 13 13
Notes: Estimates of equation (7) on Question 1, Version B, pooled across the three prompted ages (30, 40, 55), the twenty-one countries, and the thirteen models. The outcome is the recommended equity share in percentage points, and every column includes country and model fixed effects with the United States and age 40 as the omitted baselines. Column (1) fits the age effects alone. Column (2) adds the pre-registered interaction of the age-55 indicator with the standardized OECD net pension replacement rate \(\tilde{\rho}_{c}\) (the coefficient \(\psi\) of hypothesis H6). Column (3) replaces it with the coarser interaction of age 55 with an above-median replacement-rate indicator. Parentheses report CR1 cluster-robust standard errors clustered by model. The wild cluster bootstrap \(p\)-values (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\)) are the inference of record and are reported in the lower panel.

C.2Figures

Figure A1Age glide path by country (Question 1, Version B)
Notes: Mean recommended equity share by stated investor age (30, 40, or 55) within each country, Version B of Question 1, with 95 percent confidence intervals on the cell means. The downward glide path has nearly the same level and slope in every country, the raw pattern behind the null interaction of age with pension generosity in equation (7).