One advisor for the whole world?

6The Advice Rule

The country-level analysis asks whether the same investor receives different advice in different places. This section asks which rule generates the advice. I consider several candidate models of the advisor. Two (academic) advisors are motivated by the benchmark model in Section 4: the frictionless Merton planner and the life-cycle planner with background wealth. An additional candidate is motivated by the risk questionnaire of popular financial advice (Canner et al. 1997; Choi 2022). I will refer to these candidates as the Merton advisor, the Life-Cycle advisor and the Questionnaire advisor.

The goal of this section is to separate between these advisors by varying one output of the benchmark at a time. I first use a stated risk tolerance to target the risk-aversion coefficient in the denominator of the Merton rule (1). All three candidates predict that it raises the equity share. Second, I vary the investment horizon. Under CRRA preferences with i.i.d. returns the optimal share is the same at every horizon, so neither the Merton advisor nor the Life-Cycle advisor would respond. However, the Questionnaire advisor may respond, as investment horizon is a common question in popular financial advice. Third, I vary pension generosity, which targets the outside-wealth multiplier \(1+B_{c}/W\) in equation (2). This variable is only relevant for the Life-Cycle advisor (a generous pension raises the share). These tests were pre-registered. Afterwards, I also added several other exploratory tests that help further distinguish the advisors. I added a probe about the composition of pension wealth, which only the Life-Cycle advisor should respond to. I added a statement about the share of the total portfolio that the stake represents, which the Life-Cycle advisor should again respond to. Finally, I added a statement about currency hedging. A hedge removes exchange-rate risk for a non-U.S. investor, so both the Merton advisor and the Life-Cycle advisor predict that it raises the share. Table 2 at the end of the section lists the prediction of each advisor for every test.

Empirically, I hold the country fixed and vary the investor. I take the English Question 1 prompt at age 40 and insert a statement about the investor’s risk tolerance, pension, or investment horizon. The prompts are otherwise identical to Version B in the six English-prompting countries and to Version B-cross in the fifteen others, with the same local-currency amounts. I collect every cell in all twenty-one countries.

6.1Risk tolerance, horizon and pension generosity

The pre-registered arm has six treatment cells. The two risk-tolerance statements and the two pension statements are crossed, which gives four cells. The risk-tolerance statements are length-matched qualitative sentences: “I am comfortable with substantial fluctuations in my portfolio value, including potential short-term losses, if it means higher expected returns over time” versus “I am uncomfortable with large fluctuations in my portfolio value and would prefer to limit potential losses, even if it means lower expected returns.” The pension statements are “I will receive a substantial pension when I retire, projected to replace most of my pre-retirement income” versus “I will receive only a small pension when I retire, covering a modest fraction of my pre-retirement income.” The two remaining cells replace the phrase “over a five-year horizon” with a one-year or a twenty-year horizon and insert no other statement.

On the four crossed risk–pension cells the estimating equation is \[y_{icm} \;=\; \beta_{R}\,\mathrm{Tol}_{i} \;+\; \beta_{P}\,\mathrm{Gen}_{i} \;+\; \beta_{RP}\,\mathrm{Tol}_{i}\,\mathrm{Gen}_{i} \;+\; \phi_{c} \;+\; \lambda_{m} \;+\; \varepsilon_{icm}, \tag{6}\] where \(\mathrm{Tol}_{i}\) and \(\mathrm{Gen}_{i}\) indicate the risk-tolerant and generous-pension statements, the omitted cell is the risk-averse investor with the limited pension, and \(\phi_{c}\) and \(\lambda_{m}\) are country and model fixed effects. For the horizon question, I regress the recommended share on an indicator for the twenty-year horizon with the one-year horizon omitted, yielding \(\beta_{H}\). The coefficients \(\beta_{R}\), \(\beta_{P}\), and \(\beta_{H}\) are three pre-registered tests, each without multiplicity correction, and the interaction \(\beta_{RP}\) is descriptive. Inference is the same model-level wild cluster bootstrap as the primary regressions.

The estimates yield three findings. Appendix Table A18 reports the cell means and the coefficient estimates. First, the advice responds strongly to stated risk tolerance. The mean recommended share rises from 0.30 in the risk-averse cells to 0.78 in the risk-tolerant ones. The coefficient, which measures the effect at the limited pension, is \(\hat{\beta}_{R}=46.3\) percentage points (confidence interval \([41.1, 51.4]\)). In addition, Appendix Table A19 shows that the response is almost the same in every country. All twenty-one country responses to stated risk tolerance lie inside a band of 3 percentage points, and no country’s response differs from the U.S. response after Holm correction. All three advisors predict this response, so the cue does not separate them. Inverting the Merton rule (1) at the paper’s calibration puts the implied relative risk aversion at roughly 2 for the stated-tolerant cells, 5 for the stated-averse cells, and 3 for the no-statement reference, all inside the conventional range.

Second, the advice responds strongly to the stated horizon, which ought to mostly matter to the Questionnaire advisor. A caveat is that mean reversion in returns (Campbell and Viceira 2002) could rationalize the horizon response for a Life-Cycle or Merton advisor. The mean equity share is 0.36 at the one-year horizon and 0.76 at the twenty-year horizon, and the coefficient is \(\hat{\beta}_{H}=39.9\) percentage points (confidence interval \([25.7, 53.4]\)). The horizon moves the share substantially in every country, but its cross-country span of 11.9 percentage points is wider than the 3-point band for stated risk tolerance. The response is largest in Sweden, France, Poland, and Switzerland, at 44 to 46 points, and smallest in China, Mexico, South Korea, and Turkey, at 34 to 35 points.

Third, the advice responds at most weakly to the stated pension. At the risk-averse baseline the generous pension lowers the recommended share by 1.0 percentage point (\(\hat{\beta}_{P}=-1.0\), confidence interval \([-2.3, +0.2]\)). For the risk-tolerant investor it raises the share by about 2.6 points, the sum of \(\hat{\beta}_{P}\) and the risk-by-pension interaction of \(+3.6\) points (\([2.0, 5.2]\)). Averaged over the two investors, the effect is about \(+0.8\) points. The pension response is equally small in every country, with a cross-country range of 1.4 points. This invariance result is consistent with both the Questionnaire advisor and the Merton advisor, but is not consistent with the Life-Cycle advisor. For the Life-Cycle advisor, a stated generous pension is a bond-like annuity claim that should crowd financial-portfolio equity in (Cocco et al. 2005), yet the estimated response is at most a few points.

6.2Age and the pension system

The country/language arm provides a second test of whether the advice takes pensions into account. In the main arm, I varied the age of the investor between 30, 40, and 55 in Version B of Question 1 in the local language. I pre-registered a test of whether the age profile interacts with the national pension system. The intuition again comes from the Life-Cycle advisor, for whom a generous pension is akin to a large bond-like claim whose weight in total wealth grows as retirement nears. To test this, I estimate \[y_{imca} \;=\; \alpha_{m} + b_{c} + \delta_{a} + \psi\,\tilde{\rho}_{c}\,\mathbf{1}[\text{age}=55] + \varepsilon_{imca}, \tag{7}\] where \(\tilde{\rho}_{c}\) is the OECD net pension replacement rate standardized across the twenty-one countries. The coefficient \(\psi\) measures how the age-55 tilt toward bonds changes with pension generosity. The hypothesis is that the equity share of the liquid portfolio falls less steeply with age where pensions replace more income. However, there is little support for this hypothesis. The models tilt toward bonds with age, but they apply the same tilt whatever the pension system. Appendix Table A20 reports the results. Relative to the 40-year-old, the models recommend 6.2 percentage points more equity to the 30-year-old and 11.1 points less to the 55-year-old. A one-standard-deviation higher replacement rate changes the age-55 tilt by an insignificant \(-0.02\) points. The three advisors make different predictions here. The Merton advisor predicts no age gradient at all, so the gradient by itself rules it out. The Life-Cycle advisor predicts a gradient that flattens where pensions are generous, and the Questionnaire advisor predicts the same gradient in every country. The models produce the gradient without the flattening.

The age results favor the Questionnaire advisor in two further ways. First, Figure 8 plots the advice at each prompted age against the 70, 60, and 45 percent that a 100-minus-age rule prescribes. Love (2013) shows that this rule prescribes less equity than the optimum at both tails of the life cycle, so the models’ departures from it run opposite to the Life-Cycle advisor. The rule matches the modal advice at 40, where exactly 60 percent accounts for 31.7 percent of responses in the six English-prompting countries and 30.5 percent elsewhere. Pooled advice falls by 0.70 percentage points per year of age against the rule’s 1.00.

The second way in which the results by age point toward the Questionnaire advisor is the lack of variation in the country age profiles. Appendix Figure A1 shows very similar glide paths in all twenty-one country panels. This will be important to keep in mind for later, as it implies that the models are implementing a similar rule in all countries. The insignificant coefficient would leave open the possibility that the advice differs across countries for some other reason. The similarity of the glide paths across countries suggests that the models are imposing a similar age rule everywhere. In their explanations, the models apply a glide path in every language but name it almost only in English. Explicit citations of the heuristic (“rule of thumb,” “100 minus age”) appear in 15 to 17 percent of primary-cell explanations in the United States and India and in under 1 percent outside the English-prompting countries. We will return to the question of heterogeneity in AI financial advice by country in more detail in Section 7.

Figure 8The advice against the 100-minus-age rule, by prompted age
Notes: Distribution of the recommended equity share at each prompted age (Question 1, Version B, all twenty-one countries and thirteen models, parsed non-refusals, \(N = 40{,}910\)). Dashed lines mark the 100-minus-age prescription, and the highlighted bar is the share of responses exactly on it. Essentially all responses sit on multiples of five percentage points, so each panel is a discrete bar chart as in Figure 1. Exploratory exhibit, not pre-registered.

6.3Which balance-sheet cues register

I now describe three added exploratory probes, intended to determine whether the models respond to other parts of the balance sheet. The earlier results indicate that the advice is not adapted to pension wealth in general, but it is still possible that the models respond to other balance-sheet cues that are relevant for the Life-Cycle advisor. To find out, I added three probes after examining the pre-registered results. Each probe is a single English sentence in the same statement slot as before. I focus on three cues: pension composition, currency hedging, and the share of wealth that the stake represents.

The composition probe states either “My pension is a guaranteed, inflation-linked annuity that will pay a fixed real income in retirement regardless of how financial markets perform” or “My pension is invested almost entirely in stocks, so its eventual value will rise and fall with equity-market performance.” The hedging probe states either “The stock fund I am considering is currency-hedged to my home currency, so its returns are not exposed to exchange-rate movements” or “The stock fund I am considering is a global fund with no currency hedging, so its returns also move with exchange rates.” The wealth ladder states that the stake is all, about half, or about a third of the investor’s liquid wealth. I first collected all seven probe cells for the United States at fifty repetitions per model and cell (4,550 queries), and then extended the composition and hedging probes to the other twenty countries at twenty-five repetitions (26,000 queries). The wealth ladder covers the United States alone. Both collections ran in August 2026. The estimates use the no-statement Version B and B-cross cells of the July collection as the reference. For a full set of results by country, see Appendix Table A19.

The Life-Cycle advisor makes a prediction for each probe, while the Merton advisor treats the stake as a standalone portfolio and responds only to the hedging probe. For the composition probe, the Life-Cycle advisor predicts a tilt toward equity when the background claim is bond-like and a tilt away from equity when the background claim is equity-like.

For the wealth ladder, with \(W\) the stake and \(D\) the stated safe savings beside it, the Life-Cycle share of the stake is \(\theta^{*}(W+D)/W\), which rises monotonically as the stated safe cushion grows. For the hedging probe, the hedge removes exchange-rate variance for a non-U.S. investor and should raise the Merton share, so both the Merton and the Life-Cycle advisor predict a hedged–unhedged gap that widens with local-currency volatility \(\sigma_{c}\). I estimate each probe cell against the no-statement reference with country and model fixed effects and the same wild cluster bootstrap as the rest of the arm. Figure 9 plots the estimates and Appendix Table A18, Panel C, reports them.

Told that the pension is invested almost entirely in stocks, the models cut the recommended equity share by 16.9 percentage points (confidence interval \([-21.8, -12.1]\), \(p<0.001\)). Told that it is a guaranteed annuity, they raise the share by 4.8 points (\([0.1, 9.6]\), \(p=0.042\)). Hence, the response to background risk is more than three times the response to background safety. The size of the cut varies across countries, although the asymmetry still holds. When told that the pension was invested in stocks, the models cut the share by 19 to 22 points in the United States, Canada, Australia, Japan, South Korea, Mexico, China, South Africa, and India, and by 11 to 17 points in the ten European countries, Turkey, and Brazil. The annuity response is small everywhere. This pattern is consistent with neither academic advisor. The Life-Cycle advisor predicts a response in both directions, and the Merton advisor predicts no response to either statement.

The wealth ladder shows the same asymmetry. Told that the stake is all of the investor’s liquid wealth, the models cut the risky share by 23.3 points (\(p<0.001\)). Told the stake is half or a third of liquid wealth, the models raise the risky share by only 1.4 and 2.5 points. The Merton advisor predicts no response on any rung. The Life-Cycle advisor predicts a cut when the stake is all of liquid wealth and a rise when it is a third or half. In contrast, the currency-hedging probe moves the advice by very little. At the mean local-currency volatility, the hedged statement changes the share by 0.4 points (\(p=0.66\)) and the unhedged statement by \(-1.3\) (\(p=0.15\)). The gap between them, estimated on the two hedging cells alone, is 1.8 points (\(p<0.001\)). As the hedging motive is more relevant in some countries than others (a U.S. investor does not need to hedge against the dollar, while a Turkish investor may want to do so), I also report the results by country. However, the effect of the hedging probe does not differ by country. Across countries the estimated gap lies between \(-2.4\) and \(+4.0\) points, it does not grow with currency volatility \(\sigma_{c}\) (in fact the slope is negative, with \(p=0.26\)), and it is negative in Turkey, the country with the highest volatility. Both the hedged and the unhedged statements raise the share more where \(\sigma_{c}\) is higher, by roughly two and a half and four points across the panel’s range, but they rise together. The models respond to the mention of currency exposure, not to what a hedge does to the variance. This is inconsistent with both the Merton and the Life-Cycle advisor, which predict a gap that widens with \(\sigma_{c}\).

Figure 9Exploratory balance-sheet probes: large de-risking, small re-risking
Notes: Effect of each probe statement on the recommended equity share, in percentage points against the no-statement Version B and B-cross reference (Question 1, age 40), with country and model fixed effects. The upper group collects statements that lower the investor’s risk exposure and the lower group statements that raise it, so the Life-Cycle advisor would place the lower group to the left of zero and the upper group symmetrically to the right. The Merton advisor would place every probe except the hedge at zero. The pension-composition and hedging cells pool all twenty-one countries and the wealth rungs are U.S.-only. Whiskers are 95 percent wild cluster bootstrap confidence intervals (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\)). All estimates are exploratory and post hoc, in no Holm family.

Both the pension composition and wealth ladder show the same one-sided pattern, where there is an asymmetric response depending on whether the prompt raises or lowers risk exposure. The Merton advisor ignores outside wealth and would respond in neither direction. The Life-Cycle advisor would respond in both, since safe outside wealth crowds equity in just as risky outside wealth crowds it out. The one-sided pattern instead matches the Questionnaire advisor. Retail suitability rules require advisors to ask whether the client can afford to lose the money. U.K. guidance calls this capacity for loss and asks advisors to recommend only the risk a customer is both willing and able to take (Financial Services Authority 2011). The MiFID II guidelines treat the client’s ability to bear losses as part of her financial situation, separate from her risk tolerance (European Securities and Markets Authority 2022). Since the recommendation must satisfy both conditions, a screen of this kind can lower the recommended risk but not raise it.

Moreover, the questions the models ask are those of the Questionnaire advisor. In one of the robustness checks, I allow the models to ask questions before providing the answer. These questions are very close to a suitability questionnaire. The responses ask about the same four things in every country: the investor’s tolerance for losses (in every response), the purpose of the money (93 percent), whether a separate emergency fund exists (93 percent), and whether the stated horizon is firm (84 percent). Three quarters of the responses ask all four. Risk tolerance is the first question in half of the responses, and the purpose of the money or the firmness of the horizon in most of the others. The order depends on the model and not on the country. Six models open with risk tolerance in more than 80 percent of their responses and six in fewer than 10 percent, while the share ranges only from 45 to 63 percent across countries. The question is often framed behaviorally, as how the investor would react to a drop of thirty percent. The thresholds are the same across countries: the thirty-percent drawdown and the three-to-six-month cash buffer appear in all countries, from Turkey to Japan. Instead, country-specific questions enter in narrow ways. For countries with weaker currencies, the models ask whether the investor would accept exposure to the home currency or prefer international diversification. Such questions appear in 24 percent of the responses, from none in the United States and the United Kingdom to 89 percent in Mexico. Questions about inflation or sovereign risk appear in 12 percent, from under 10 percent in fifteen countries to 81 percent in Turkey. Where a named tax account exists, such as an IRA or a Swedish ISK, the models ask whether the money sits in it.

6.4Invariance and responsiveness

Table 2Invariance and responsiveness: the evidence against the three advisors
Predicted effect
Cue Merton Life-Cycle Questionnaire Estimate
Panel A. Outside the benchmark (invariance)
Question language, local vs. English 0 0 0 −1.3
Local instrument names (Version C) 0 0 0 −0.7
U.K. rather than U.S. country name 0 0 0 +0.7
Twenty-year rather than one-year horizon 0 0 \(+\) +39.9
Other country names (Version B) 0 0 0 −11.0 to +1.4
Panel B. Inside the benchmark (responsiveness)
Stated risk tolerance \(+\) \(+\) \(+\) +46.3
Unhedged currency risk across countries 1 1 0 \(\hat{\kappa}=0.13\)
Age 55 rather than 40 0 \(-\) \(-\) −11.1
Stated generous pension 0 \(+\) 0 −1.0 / +2.6
Pension replacement rate \(\times\) age 55 0 \(+\) 0 −0.02
Pension in stocks / guaranteed annuity 0 / 0 \(-\) / \(+\) \(-\) / 0 −16.9 / +4.8
Stake is all / a third of liquid wealth 0 / 0 \(-\) / \(+\) \(-\) / 0 −23.3 / +2.5
Currency-hedged fund \(+\) \(+\) 0 +0.4
Real mortgage rate across countries 1 1 – \(\hat{\kappa}^{Q2}=0.23\)
Notes: Estimates are effects on the recommended equity share in percentage points, except the two calibration slopes. The three prediction columns give the effect each advisor of Section 6 predicts, where \(+\) raises the share, \(-\) lowers it, 0 leaves it unchanged, and 1 is a calibration slope of one. The Merton advisor applies equation (1) to the stake as a standalone portfolio, the Life-Cycle advisor applies equation (2) to the household’s balance sheet, and the Questionnaire advisor follows a retail risk questionnaire. All entries are pre-registered (Appendix Table A7) except the last four rows of Panel B, which are post-filing exploratory estimates.
() Prediction from the capacity-for-loss screen of Section 6.3, which I added to the Questionnaire advisor after seeing the probe results.

Table 2 collects the evidence of Sections 5 to 6 against the two properties of Section 4 and against the predictions of the three advisors. The models pass most invariance tests. The language of the question, the local names of the instruments, and the U.K. country name each move the advice by about a point or less. The exception is the investment horizon, which moves the advice by forty points although it lies outside the benchmark. However, the models fail most responsiveness tests. Stated risk tolerance moves the advice strongly and in the right direction, but currency risk, pension generosity, stated safe wealth, and the currency hedge barely move it, and the mortgage advice follows the real mortgage rate in direction but not in magnitude. The advice is thus close to invariant where theory asks for invariance and insufficiently responsive where it asks for responsiveness.

The three advisor columns of the table summarize the comparison of this section. The Merton advisor fails on the horizon, the age gradient, and the two exposure cues. The Life-Cycle advisor fails on the horizon, the stated pension, the pension-by-age interaction, and the safe side of the balance-sheet probes. Both academic advisors also fail on currency risk and the currency hedge. The Questionnaire advisor fits the most entries. It predicts the horizon response, the age gradient, the absent pension response, and the compression across countries. None of the three advisors predicts the negative country-name contrasts.