3.1. Do NVIX and Its Components Have Predictive Power for Future Returns?
I start by testing the hypothesis that time variation in news-based uncertainty is an important driver of variation in expected returns on US equity. Asset pricing theory allowing for time-variation in risk premia predicts that times of relatively high risks are followed by above average excess returns on the market portfolio (e.g.,
Merton 1973 or
Gabaix 2012). Following
Manela and Moreira (
2017), I perform monthly return predictability regressions based on news-implied (NVIX) and option market-implied (VXO) volatilities
2. I use the older VXO as my option-implied risk measure because it grants me more data, but the results do not change if I use the more recent VIX. The dependent variables are monthly log excess returns on the market index. My explanatory variables are normalized to have unit standard deviation over the entire sample. I always present
p-values that are based on
Newey and West (
1987) standard errors with the same number of lags as the forecasting window, as well as
p-values based on (wild) bootstrapped standard errors that further account for the fact that the regressor is estimated in a first stage. News-based uncertainty measures are estimated based on machine learning methods, where tools for statistical inference are underdeveloped.
Manela and Moreira (
2017) develop a
Murphy and Topel (
2002)-type methodology combined with bootstrapping techniques for computing standard errors. The regression equations and my results are presented in
Table 2, Panels A and B.
The findings of
Manela and Moreira (
2017) suggest that NVIX embeds priced information. Positive NVIX predicts positive returns at medium and longer horizons (from 6 months onwards). In contrast, the regression results presented in
Table 2 suggest that the ability of NVIX and VXO to predict returns at the short horizon (1 month) is statistically weak. This is in line with the findings of
Adrian et al. (
2019) for the VIX based on a similar research design. As a robustness check, I also perform quantile regressions. Quantile regressions are useful in the presence of outliers in the data, e.g., due to market crashes, because the median (0.5 quantile) is much less affected by extreme values than the mean in OLS. This is particularly relevant for my long sample starting in 1926, which includes various market crashes. Panels C and D in
Table 2 present the results. I conduct quantile regressions for the 0.25, 0.50, and 0.75 quantiles, where, loosely speaking, the 0.50 (0.25, 0.75) quantile refers to the market return during normal (bad, good) months, respectively. Overall, the results that I obtain for the mean in the standard OLS framework does hold for the median, the 0.50 quantile, in the quantile regressions. However, the present results suggest that NVIX significantly predicts positive return at the 1-month horizon, but VXO does not. Examining the other quantiles reveals some interesting features of the data. While NVIX significantly predicts returns during the good months, the relationship is insignificant during bad months. I obtained similar results for the VXO.
Manela and Moreira (
2017) decompose NVIX into five meaningful word categories (Government, Financial Intermediation, Natural Disasters, Stock Markets, and War) to gain insights into the origins of uncertainty fluctuations. The separate uncertainty measures allow me to analyze the origins of the potentially priced information of NVIX. The regression equations and my results are presented in
Table 3. I find that Government (policy- or tax-related uncertainty) and War (concerns related to violence or military interventions) embed priced information at the short horizon with (partly) strong predictive power. Surprisingly, Intermediation and Stock Market-related uncertainty are not reliably related to future returns. This is in line with
Manela and Moreira (
2017), who find that these concerns comove with realized volatility, but have no predictive power for future returns. Again, the quantile regressions, presented in Panel B, confirm these findings. Interestingly, for policy-related concerns, the relationship is homogeneously positive over all quantiles. In the following section, I further analyze the predictability results during recession episodes like the Great Recession, where market volatility is typically high. In Panel C, I conduct predictability regressions of the market returns on each of the news-based uncertainty measures but also add an interaction term that includes an NBER recession dummy. In particular, I run the regression,
which gives the relative beta of the predictive relationship of news-based uncertainty measures, and expected returns are conditional on recessions compared to the unconditional estimate, where
Drec,t is the recession dummy. For policy-related uncertainty, I find that
is insignificantly different than 0 and
is significantly larger than 0, which suggests that, for example, the positive ‘policy-related-uncertainty-expected-return’ relationship holds during recessions. Hence, in line with previous research, these results suggest that, during recession periods, investors tend to be more risk-averse, which allows us to observe a significantly positive ‘policy-related-uncertainty-expected-return’ relationship in the data. During non-recessions, investors are typically less risk-averse, which oftentimes leads to an insignificant relationship. Interestingly, for stock market-related uncertainty, the relationship with expected returns is typically significantly positive (
), and turns significantly negative during recessions (
). I further analyze this dynamic relationship later in the paper.
3.2. Is the Effect Economically Significant?
The logic behind uncertainty-managed portfolios is similar to the concept of volatility timing. Market volatility timing relies on a negative relationship between market volatility and expected return. The findings of
Moreira and Muir (
2017) suggest that volatility-scaled portfolios produce positive alphas and increase Sharpe ratios. I replicate their analysis and, in the first step, compute the monthly realized market volatility on the last day of each month from daily S&P500 index returns in the previous 21 days. I use the realized market volatility in month t to scale the monthly excess market returns in period t+1. Regarding the scaling,
Barroso and Santa-Clara (
2015) argue that having a certain target seems like a natural choice, as constant volatility is desirable to avoid the time-varying probability of having large losses in the portfolio. Hence, given that realized volatility negatively predicts returns, the scaled portfolio weight in the original market factor at time t+1 is given by
. The choice of the target is arbitrary, but, in line with other studies, I use an optimizer to find a certain target that produces scaled portfolio returns with the same ex-post volatility compared to the (unscaled) market factor.
In
Figure 2, I compare the volatility-managed market factor strategy with the original (unscaled) market factor returns and an alternative variance-managed strategy. I also include the NBER recession periods. In line with
Moreira and Muir (
2017), results suggest that the risk-managed portfolios outperform the unscaled market factor portfolio. This is a surprising feature of the data, because volatility-timing generates higher average returns, because changes in market volatility are not offset by proportional changes in expected returns. Nonetheless, studies indicate that market volatility can accurately forecast momentum returns. For instance, after adjusting for market condition and business cycle factors,
Wang and Xu (
2015) discover that market volatility has the strong ability to predict momentum profits. Loser stocks perform noticeably poorer than winner stocks during periods of low market volatility, which leads to huge momentum profits. In fact, I discover that a portfolio of winning stocks and the volatility-managed portfolio exhibit a large positive correlation (83%, significant at the 1% level). A time-series market volatility-based trading strategy might therefore take advantage of the well-known momentum anomaly recorded in the cross section, given the established inverse relationship between market volatility and momentum.
In
Table 4, Panel A, I present the equity risk-adjusted performance of the volatility-scaled and variance-scaled strategies. I report the Fama-French three-factor plus momentum alphas that are typically used in the literature.
Frazzini and Pedersen (
2014) also demonstrate that a strategy that shorts high-beta stocks and longs low-beta stocks (betting-against-beta, or BAB) can generate significant alphas in comparison to the Fama-French three-factor model, which incorporates a momentum factor, and the CAPM. The market volatility-based strategy is conceptually distinct from the volatility-managed market strategy because the high risk-adjusted return of the BAB factor reflects the fact that variations in average returns across time periods are not explained by variations in CAPM betas in the cross-section. However, given its relationship to market volatility, the BAB method may be able to identify similar events in the data, as momentum does. As a result, I additionally account for the
Novy-Marx and Velikov (
2021) version of the original BAB factor (EWBAB) and their updated version (VWBAB). The findings indicate that momentum is, in fact, the primary driver of the positive alpha for the volatility-managed market component. According to
Moreira and Muir (
2017), the market volatility- (variance-) managed market factor produces a significant Fama-French 3 factor (FF3) monthly alpha of 0.30% (0.41%). However, the FF3 plus momentum alpha, with a positive and statistically significant loading on momentum, decreases to an insignificant 0.11% (0.16%). Additionally, I discover that adding the VWBAB factor as a control has an effect on alphas as well. Therefore, the low-beta anomaly reported in the cross-section is not entirely different from the time-series volatility-managed portfolios.
Given the strong predictive power of policy-related uncertainty for the mean-variance trade-off, I propose a strategy, where the portfolio is ‘uncertainty-managed’. I use the policy-related uncertainty (Government) in period t-1 to scale the monthly market factor returns in period t+1 in order to achieve a given target in terms of Government. The scaled portfolio weight in the original market factor of the enhanced uncertainty-managed trading strategy at time t+1 is given by
. Hence, I invest more (less) when past uncertainties have been above (below) the target. Importantly, since our uncertainty measure positively corelates with the attractiveness of the mean-variance trade-off, the weights are defined as the inverse of the weights that
Moreira and Muir (
2017) use in their analysis. Again, I choose a target that produces scaled portfolio returns with the same ex-post volatility compared to the (unscaled) market factor
3.
In
Figure 3, I compare the performance of the ‘policy-related-uncertainty-managed’ strategy (GOV) with the unmanaged benchmark. I also plot the NBER recession periods. Alternatively, I show the performance of a strategy, where the market returns are scaled using squared weights (GOV
2), similar to the variance-managed portfolio. Interestingly, the GOV
2 strategy substantially outperforms the GOV strategy and the unmanaged benchmark. In particular, during NBER recession periods the scaled market factors appear to outperform the unscaled benchmark. In order to understand the magnitude of this effect, I run the following regression
, which gives the relative beta of the scaled factor conditional on recessions compared to the unconditional estimate, where
Drec,t is a recession dummy. I find that
is significantly smaller than 0, which suggests that the GOV
2-managed strategy takes less risk during recessions and thus has lower betas during recessions. The non-recession market beta of the uncertainty-managed market factor is 0.91 (t-stat 21.26) but the recession beta coefficient is −0.27 (t-stat 4.03), making the beta of the GOV
2-managed portfolio conditional on a recession equal to 0.64. By taking less risk, the average monthly return of the GOV-managed portfolio increases to 0.26% during recessions, from -0.55% for the unscaled market factor and −0.13% for the volatility-managed portfolio. During non-recession periods, the average monthly returns are 1.06%, 0.94%, and 1.02%, respectively. Hence, while the managed equity portfolio takes more risk when policy-related uncertainty is high, it takes less risk during recessions. This counterintuitive result can be explained by the observed negative correlation between market volatility and policy-related uncertainty. For example, while stock market-related concerns are twice as high, government-related concerns (policy- or tax-related uncertainty) are significantly lower during recessions.
In
Table 4, Panel B, I present the equity risk-adjusted performance of both policy-related-uncertainty-scaled strategies. The managed equity portfolios that take more risk when news-based uncertainty is high generate an annualized equity risk-adjusted alpha of up to 5.33% (Gov
2) with an appraisal ratio of 0.46. While cross-sectional anomalies like momentum and BAB explain the abnormal returns of the volatility-managed strategy, equity risk-adjusted alphas do not substantially change once I use the factors as controls. Thus, the time-series policy-related uncertainty-managed portfolios are completely distinct from the momentum and low-beta anomaly documented in the cross-section. Admittedly, this type of analysis is interesting from an economic point of view, but suffers from a look-ahead bias, given that the current version of the government uncertainty measure that is used for the strategy is constructed using the information for the full sample. However, the strategy is feasible, given that it, in principle, relies on past observations.
In the following, I analyze the time-series properties of the GOV
2 strategy. In
Figure 4, I show the rolling window CAPM alphas of the GOV
2 strategy, market volatility, and the standard deviation of the policy-related uncertainty measure. The figure presents the annualized CAPM alpha of the dynamic trading strategy for rolling 15-year windows. Each point in the graph corresponds to using the previous 15 years of data. Apparently, the CAPM-alpha of our strategy is not necessarily different during subsamples when volatility is relatively high. The subsample CAPM-alpha typically fluctuates around 5%, except during the time period 1980–2000. This should not be surprising as our results rely on a large degree of variation in policy-related uncertainty about work. Variations in policy-related uncertainty were relatively low over this period; hence, my strategy is not very different from a buy-and-hold strategy and alphas are low.
The findings presented in
Table 3 Panel A suggests that the ‘stock market-related-uncertainty-expected-return’ relationship, like the market volatility-expected return relationship, is on average, negative
4. As discussed earlier,
Moreira and Muir (
2017) propose a related market volatility-based trading strategy that takes less risk when market volatility is high. Similarly, I analyze a stock market uncertainty-based trading strategy that takes less risk (decreases the weights) when stock market uncertainty is high. Hence, the scaled portfolio weight in the original market factor at time t+1 is given by
, where SMU refers to my measure of stock market uncertainty. As a result, I invest more (less) when past uncertainties have been below (above) the target. Again, I choose a target that produces scaled portfolio returns with the same ex-post volatility compared to the (unscaled) market factor. When analyzing the ‘stock market-related-uncertainty-expected-return’ relationship in
Table 3 Panel C, the results suggest that it is typically positive, but turns negative during recessions. Hence, in the following, I design a dynamic to take into account this particular feature of the data. I analyze a dynamic stock market uncertainty-based trading strategy that typically takes more risk when stock market uncertainty is high, but takes less risk when stock market uncertainty is high. Hence, the scaled portfolio weight in the original market factor at time t+1 is given by
during ‘normal’ times, but changes to
during recessions. As a result, during ‘normal’ times, I invest more (less) when past uncertainties have been above (below) the particular ‘normal target’ and during recessions, I invest more (less) when past uncertainties have been below (above) the particular ‘recession target’. Again, I choose targets that produce scaled portfolio returns with the same ex-post volatility compared to the (unscaled) market factor. While both trading strategies generate a significantly positive market risk-adjusted alpha, the dynamic stock market uncertainty-based trading strategy clearly outperforms the static one with a CAPM-alpha (
p-value) of 4.6% (0.00) and 3% (0.04), respectively. Unsurprisingly, similar to the market-volatility-timing, both strategies do not ‘survive’ the further risk-adjustments. Hence, these results suggest that the positive risk-return still holds during ‘normal’ times, but becomes significantly weaker during crisis periods. Importantly, the market volatility-based trading strategy proposed
Moreira and Muir (
2017), which takes less risk when market volatility is high, seems to only pick up a feature of the data during crisis periods. In light of the controversies surrounding the divergent governmental policy approaches taken by Republicans and Democrats, I further examine the variations in the uncertainty-managed portfolio contingent on which political party holds the majority in the Senate. In general, I believe that, when Republicans hold a majority, there is less uncertainty around governmental policies. At the 1% level, the average government measure throughout these times is 0.75, and, during times when Democrats hold a majority in the Senate, it is 0.98. According to the earlier research, one may anticipate that the government plan would perform especially well at times when the Republicans hold a majority in the Senate because of the decreased level of uncertainty. In order to understand the magnitude of this effect, I run the following regression,
which gives the relative beta of the scaled factor conditional on the Republicans having the majority in senate (
) compared to the unconditional estimate. I find that the “Republicans market beta” is significantly lower compared to the “Democrats market beta”, which suggests that the government-managed strategy takes less risk during periods when the Republicans have the majority in senate. The Democrats market beta of the managed market factor is 1.05 (t-stat 19.46) while the Republicans market beta coefficient is −0.30 (t-stat −9.59), making the beta of the managed portfolio conditional on a Republicans period equal to 0.75. By taking more risk in a high policy-related uncertain market, the monthly CAPM alpha decreases to 0.23% during Democrats periods, from 0.33% during Republicans periods.