Countercyclical Return Expectations: Evidence from the Livingston Survey

The paper builds three measures of the annualized expected excess return, all based on the median across survey participants of the expected gross capital gain from t+1 to t+2, and they differ only in how they correct for expected dividends. Because participants were not asked to report the current level of the stock index until 1992, the authors follow Dokko and Edelstein (1989) and compute the forward expected return from the 6-month-ahead and 12-month-ahead S&P forecasts, dividing one by the other. The first measure, denoted with superscript “0”, makes no dividend correction and simply subtracts the annualized 3-month Treasury bill rate. The second, “div”, adds the current dividend–price ratio scaled by the sample average of realized dividend growth, following Nagel and Xu (2023). The third, “cor”, instead scales the dividend–price ratio by the survey’s own expected gross growth in tax-adjusted corporate profits; this series is unavailable before June 1971, so this third measure runs on a restricted sample. The median rather than the mean is used “to ensure that these expectations are not sensitive to a few extreme observations.” The three measures are highly correlated — 0.97 between the first two, 0.95 between the first and third, and 0.999 between the second and third — so the dividend corrections have “a relatively small impact.”

Across 13 state variables scaled to be procyclical, 11 give slope coefficients that are negative and statistically significant, which the paper reads as “significant and sizable countercyclical variation in return expectations.” The regression is a simple OLS of expected excess returns on one standardized state variable at a time, using non-overlapping biannual observations with Newey-West (1987) standard errors at one lag. Panel A of Table 1 covers business-cycle indicators: the NBER recession dummy gives slopes of −1.88, −2.29 and −2.36 across the three return measures (R² = 13.3%, 21.5%, 23.4%), the PMI −1.59 to −1.89 (R² = 9.6% to 15.5%), ADS −2.16 to −2.44 (R² = 21.5% to 26.8%), and CFNAI −2.22 to −2.59 (R² = 23.3% to 31.1%), all with p-values of 0.00. Panel B covers macroeconomic predictors: Cochrane’s (1991) investment–capital ratio, the Cooper–Priestley (2009) output gap, and the cyclical consumption measure of Atanasov, Møller and Priestley (2020) all give significantly negative slopes. Panel C uses the Fama–French (1989) financial predictors; the credit spread and the term spread give countercyclical slopes, while the price–dividend ratio gives mixed results. Panel D uses the Jurado, Ludvigson and Ng (2015) macro and financial uncertainty indices, both of which have “a significant countercyclical effect on all three measures of return expectations.” The two exceptions are the consumption–wealth ratio and the price–dividend ratio.

No: once the paper controls for structural breaks or estimates in first differences, both exceptions turn negative, and the paper notes these two variables “display the highest persistence among the considered state variables.” Using the break-adjusted price–dividend ratio of Lettau and van Nieuwerburgh (2008) — whose semiannual persistence falls from 0.95 to 0.82 and whose correlation with the output gap rises from 0.04 to 0.34 — all three slope estimates become negative, significant at the 5% level for the “div” and “cor” measures. Using the Markov-switching consumption–wealth ratio of Bianchi, Lettau and Ludvigson (2022), all three slopes are again negative, significant at the 10% and 5% levels for the first two measures; the third is insignificant with a p-value of 0.15, which the authors attribute to “lower statistical power for this return measure, which is unavailable before June 1971.” Estimating the regressions in first differences gives negative slopes for all 13 state variables and all three return measures, with every slope except the consumption–wealth ratio statistically significant — including the price–dividend ratio. The authors observe that the R² values in the first-difference regressions “are very similar” to those in levels, which they take as “yet another indication that the relation in the level regressions is not spurious.”

The paper describes them as “approximately rational”: under the Adam, Marcet and Beutel (2017) test, the null of rational expectations generally cannot be rejected at the 5% level, with cyclical consumption the single exception. The test compares the covariance of survey expectations with a state variable against the covariance of realized excess returns with the same variable; under rational expectations, the two slope coefficients should be equal. The estimated slopes for realized excess returns are “mostly negative and therefore also consistent with countercyclical return expectations,” except for the two uncertainty indices, which have slightly positive estimates. The realized-return slopes tend to be larger in absolute value than the survey slopes, “showing that survey expectations tend to be less cyclical than the expectations inferred from realized excess returns, as emphasized by Nagel and Xu (2023),” but the realized-return confidence intervals are wide. For cyclical consumption the p-value is 0.02 for the first two return measures, though not for the third.

Nagel and Xu (2023) reject rational expectations for four of their nine state variables; the paper attributes the difference to two methodological choices rather than to a different reading of the data. First, Nagel and Xu implement the test using all available data, so some of their realized-return estimates use data back to 1926 while the Livingston survey starts in 1952; this paper uses “the same time span when estimating β from realized returns to ensure that any differences between these estimates reflect deviations from rational expectations rather than structural breaks.” Second, Nagel and Xu use the forward return measure from 1952 to 1991 as a proxy for the annual spot return expectation available from 1992, and pair it with realized spot returns based on overlapping observations; this paper uses the forward measure throughout “to ensure consistency between the survey expectations and realized returns.”

Belief revisions respond to contemporaneous news in the direction implied by countercyclical expectations, and generally do not respond to lagged news — a pattern the paper describes as “consistent with the rational expectations hypothesis.” News in each state variable is measured as the residual from a first-order autoregression, with the NBER recession indicator omitted because it “by construction contains a small amount of news.” Regressing the belief update from t to t+1 on contemporaneous news gives negative slopes on all coefficients, with nine of the twelve significant at the 10% level: “negative news leads to an upward revision in expected future excess returns.” Regressing the belief update on lagged news gives coefficients that are all insignificant except for the term spread.

Livingston return expectations are positively and significantly correlated with the expected excess returns implied by all three benchmark models, which the paper reads as showing they “align closely” with a rational investor’s. The correlation with the one-year expected excess equity return of Martin’s (2017) log-utility investor (available 1996–2019) is 0.47, p-value 0.00. The correlation with the six-month forward expected excess return in the Campbell–Cochrane (1999) habit model — computed from historical consumption data for surplus consumption, using the original calibration except for the mean and standard deviation of consumption growth, which are adjusted to the 1952-onward sample — is 0.35, p-value 0.00. The correlation with the Bansal–Yaron (2004) long-run risk model, using the estimated version of Schorfheide, Song and Yaron (2018) with data through 2014, is 0.44, p-value 0.00. The paper states these correlations suggest the survey is consistent with a rational investor facing habit-induced changes in risk aversion and with one requiring compensation for time-varying economic uncertainty.

The alternative surveys go the other way: Gallup, AAII and Investor Intelligence imply “highly procyclical” return expectations, and the paper confirms rather than overturns Greenwood and Shleifer’s finding for those surveys. Following Greenwood and Shleifer, the authors compute bull–bear spreads for Gallup (1996–2019), AAII (from 1987) and Investor Intelligence (from 1963), plus an Expectations Index formed as their first principal component, all converted to semiannual frequency. The three alternative surveys are highly positively correlated with each other, while Livingston expectations are uncorrelated with AAII, Investor Intelligence and the Expectations Index, and “highly negatively correlated with the Gallup survey.” Regressing the alternative surveys on the same state variables reverses the sign of the slope coefficient “in almost all cases,” with the reversal usually statistically significant. All the alternative surveys are also negatively correlated with the three model-implied expected excess returns (Investor Intelligence is uncorrelated with the habit-model measure).

Controlling for the realized S&P excess return over the past year does not remove the countercyclical pattern in the Livingston survey, and the estimated loading on past returns there is negative. For the Expectations Index, the coefficient on past realized returns is significantly positive across all state variables, and many state variables lose 5% significance once past returns are controlled for — all four business-cycle indicators and both uncertainty indices become insignificant — which the paper takes as confirming Greenwood and Shleifer’s (2014) finding of extrapolation among individual investors. For the Livingston survey over 1952–2019, by contrast, all the state-variable coefficients remain negative and significantly different from zero, and all 14 estimated coefficients on past realized returns are negative, 11 of them significant at 5% — “a similar finding is reported in Nagel and Xu (2023) for the Livingston survey.” The sign of the state-variable coefficients is robust to using the shorter 1987–2019 sample, “although the shorter sample reduces the statistical power.”

It proposes that the economy contains at least two types of investors who form expectations differently, rather than concluding that survey expectations are acyclical. Livingston participants are “economic experts from the industry, government, and academia” who “should therefore have a solid understanding of return and risk”; Gallup and AAII participants are individual investors who “do not necessarily have economic training” and “might be more inclined to invest according to their own previous performance or the performance of their peers,” making them more subject to herding or positive-feedback trading. Investor Intelligence newsletter authors are grouped with the individual investors, on the grounds that publishing a newsletter requires no particular educational background, that subscription revenue gives an incentive to propose high-return strategies with limited attention to risk, and that Jaffe and Mahoney (1999) find newsletters propose securities that performed well in the past, while Graham (1999) supports herding among them. The paper relates this to Barberis et al. (2015), in which some investors extrapolate past price changes while others hold fully rational beliefs. It explicitly acknowledges the alternative that “noise and measurement issues in general may contaminate the results from the various surveys,” and cites Boons, Ottonello and Valkanov (2024) as evidence favouring the first explanation.

Cash flow expectations, measured from survey forecasts of tax-adjusted corporate profits, also display countercyclical variation, but the paper states plainly that “the statistical evidence is not as strong as for return expectations.” Nominal cash flow expectations give negative slopes for every state variable except the consumption–wealth ratio, with only 7 of 13 slopes significant at the 5% level; for real cash flow expectations — deflated using the survey’s own median CPI expectations via a first-order approximation — all slopes are negative and 6 of 13 are significant at 5%. The sample runs 1971–2019 at semiannual frequency. The paper reads the timing from Figure 4: cash flow expectations “often attain a local minimum just before recessions, and then increase during the course of recessions such that cash flow expectations attain a local peak around the end of, or just after, recessions,” clearly visible for the 2001 and 2007–09 recessions, and coinciding with the output gap’s own local peaks and troughs. This late peak is offered as an explanation for why the countercyclical evidence is weaker for forward-looking state variables such as the term spread and financial uncertainty.

Adding realized cash flow growth over the past year leaves the state-variable slopes “more or less unaffected,” and the Adam, Marcet and Beutel (2017) test generally fails to reject rationality at the 5% level for both nominal and real cash flow expectations. The coefficient on past realized cash flows is generally insignificant at 5% across state variables — the PMI is the exception for one measure — so the paper concludes participants “do not display extrapolative cash flow expectations.” In the rationality test, slope coefficients for realized cash flows are generally negative, again indicating countercyclical dynamics, but are “estimated somewhat imprecisely,” with only 3 of 13 significant for nominal and 4 of 13 for real realized cash flows. The exceptions where rationality is rejected are the investment–capital ratio and the term spread.

Both benchmark models fail on the cash flow dimension: the paper concludes “these rational asset pricing models are not fully consistent with survey expectations.” The Campbell–Cochrane (1999) habit model assumes cash flow growth is white noise around a non-zero mean, which “implies constant and hence acyclical cash flow expectations in the habit model.” In the Bansal–Yaron (2004) long-run risk model, expected cash flow growth is linear in the long-run risk state, and regressing that state on the same set of business-cycle variables gives generally positive slopes — i.e. procyclical cash flow expectations — with the consumption–wealth ratio the exception, and this is robust to controlling for past realized cash flows. So the models match the countercyclical return expectations in the survey but not the countercyclical cash flow expectations.

Return and cash flow expectations “appear to share a common countercyclical component,” which the paper presents as a new result, with the implication that both the discount rate channel and the cash flow channel matter for equity prices. Both cash flow and return expectations have an annualized volatility of roughly 6% across the various measures. The authors link this to de la O and Myers (2021), and state the conclusion explicitly rests on two implicit assumptions: “that the dynamics of the considered short-term expectations (12 months ahead) are similar to more long-term expectations, and that the applied expectations are valid proxies for the beliefs among investors.”

Several, stated in the paper’s own words. The Livingston survey “gives a relatively clean measure of expected excess returns, although it is not free from potential measurement errors.” Glaser, Iliewa and Weber (2019) provide evidence that “asking for forecasts of prices rather than returns may result in lower expectations,” which applies directly since Livingston asks for index levels. The dividend correction in the “div” measure “may understate the variation in expected excess returns if expected dividends display considerable variation,” and the de la O and Myers (2021) data that would address this are unavailable before 2003. The “cor” measure exists only from June 1971. The alternative surveys have shorter samples that “cover only a few business cycles.” The paper also notes it “cannot rule out that certain investment advisors include considerations based on past returns” without their own expectations being extrapolative, which it flags as “a potential pitfall associated with using investment advice to measure return expectations.”

Structural-break adjustments, first-difference regressions, alternative return definitions, and a set of Online Appendix checks. From June 1992 the survey also asks for the S&P level at the end of the current month, which allows 6-month and 12-month spot return expectations to be computed. Over 1992–2019, 12 of 13 state variables give negative slopes for the two spot measures, 10 of them significant at the 10% level, and the 6-month spot slopes are numerically larger than the 12-month spot slopes, which in turn exceed the benchmark forward-measure slopes — so the evidence of countercyclicality is “even stronger” on the spot measures. The paper relates this to Gandhi, Gormsen and Lazarus (2023). The Online Appendix additionally reports robustness across subsamples, to using the mean instead of the median as the summary statistic for survey expectations, and to accounting for dispersion in return expectations within the survey.