Capital Flows in Risky Times: Risk-on/Risk-off and Emerging Market Tail Risk

That research on cross-border capital flows has “largely focused on the first moment of the relevant distributions,” so it cannot speak to how global risk reshapes the extremes. The authors position the paper as building on Gelos et al. (2019) but going further: “we focus on the full distributions of emerging market capital flows and returns; most important, we characterize the manner in which extreme capital flow and returns realizations are tied to global risk appetite” (§1, pp. 2-3). The complaint is put sharply: “The weight placed on means and variances as sufficient summary statistics precludes the data from speaking to the underlying distributional granularity of global risk – a challenge we overcome by turning to heterogeneous effects across quantiles” (§1, p. 4). They also frame the distinction that organises the empirical work: a risk-off shock “can make the whole emerging market capital flow or return distribution wider by pulling out both of the tails,” or it “could simply fatten the left tail,” and which of these happens matters for how downside risk should be assessed (§1, pp. 3-4).

As the z-score of the first principal component of daily changes in a set of standardised asset-market series, normalised so that a positive change means risk-off, and scaled by their historical standard deviations before the principal component is taken. The construction follows Datta et al. (2017) and spans four groups (§2.1, pp. 9-10). Credit risk: the change in the ICE BofA BBB Corporate Index option-adjusted spread for the United States and for the euro area, plus Moody’s BAA corporate bond yield relative to ten-year Treasuries. Advanced-economy equity: the additive inverse of daily total returns on the S&P 500, STOXX 50 and MSCI Advanced Economies Index, plus changes in the VIX and VSTOXX. Funding liquidity: the daily average change in the G-spread on 2-, 5- and 10-year Treasuries, plus the TED spread, the 3-month LIBOR-OIS spread, and the bid-ask spread on 3-month Treasuries. Currencies and gold: growth in the trade-weighted dollar index against advanced foreign economies and the change in the gold price. Four sub-indices, one per group, are constructed the same way, with the groupings “chosen to maximize the total explained variation of the components.”

Four reasons, one of them an identification argument. First, aggregating sources lets the analysis “abstract from any one source of risk-off behavior,” since elevating a single asset price risks a relationship driven by one set of market participants: “the group of market participants who take out, for example, S&P 500 options to hedge against U.S. equity volatility have characteristics that may not extend equally to all other risk assets” (§2.1, p. 10). Second, the multivariate measure permits decomposition into drivers. Third, it admits non-US sources of risk-off movements. Fourth – the identification argument – “because we measure the relationship of our index to assets in emerging markets, we can appeal to the small open economy character of the recipient markets to strengthen the plausibility of the index’s exogeneity to local market fundamentals” (§2.1, pp. 10-11). The paper also cites a body of work arguing the VIX has become a less reliable proxy since 2008 – Avdjiev et al. (2017) attributing the decline to the shifting composition of flows, Cerutti et al. (2019) finding the VIX-flow correlation limited to crisis times – and reports that “the parameter values using our composite risk-on risk-off measure are remarkably stable (if not stronger) over time given that the measure captures alternative sources of risk” (§1, pp. 7-8).

The risk aversion series of Bekaert, Engstrom and Xu (2020), used because the statistical index “may nevertheless confound information about variation in risk appetite with variation in physical risk.” That work builds on habit-based asset-pricing models to separately identify variation in risk and in risk aversion from a wide set of macro and financial data, and the paper takes its daily risk-aversion series as the second RORO measure while including the model’s quantity-of-risk series as a control (§2.1, p. 11; §3.3, p. 16). The authors are symmetric about the trade-off: the structural measure “has the advantage of measuring risk aversion using a structural model specifically designed to separate the price of risk (or risk aversion) from the quantity of risk. However, inference about this separation may, of course, be contaminated by any model mis-specification” (§1, p. 3). Running both is the point: “we consider both approaches to gauge the extent to which we can draw relatively robust conclusions.”

Weekly EPFR country flows and daily total returns for 22 emerging markets. Flows come from EPFR Global’s Country Flows dataset, which combines fund-level flow data with fund managers’ country weightings; because the country flows are sums of fund-level reallocations they “come cleansed of valuation effects and therefore represent real quantities” (§2.2, pp. 11-12). The underlying universe is more than 14,000 equity funds and more than 7,000 bond funds with more than $8 trillion under management. Returns are country-level: the J.P. Morgan EMBI for dollar bonds, the Bloomberg Barclays Local Bond Index for local-currency bonds (available only from 2008), and MSCI local-currency and dollar indices for equities. The country set is Argentina, Brazil, Chile, Colombia, the Czech Republic, Egypt, Hungary, India, Indonesia, Malaysia, Mexico, Pakistan, Peru, the Philippines, Poland, Qatar, Russia, South Africa, Taiwan, Thailand, Turkey and the United Arab Emirates; China is excluded “due to its unique characteristics, including its size relative to other emerging market economies and measurement issues” (§2.2, fn. 8). The sample begins 7 January 2004; the paper gives its end date as 9 April 2020 in the introduction and 15 April 2020 in the data section.

Global “push” and country-specific “pull” controls, all lagged except emerging-market news surprises, plus country and time fixed effects. Push controls are advanced-market returns (from Kenneth French’s data), the advanced-economy monetary stance measured by the shadow rate, and advanced-economy real GDP growth, each a USD real-GDP-weighted average of the United States, the UK, the euro area and Japan (§2.2.1, pp. 12-13). Pull controls are local policy rates, average real GDP growth over the previous eight quarters, and the broad real effective exchange rate. The Citigroup Economic Surprise Index for emerging markets captures local macroeconomic news in the intervening week or day. The paper is candid that the controls are not innocent: “Both sets of controls affect capital flows and returns, but also likely react directly to changes in risk sentiment. In fact, our advanced economy push variables not only react to risk-on/risk-off shocks but likely also drive them.” A lagged dependent variable is included to absorb autocorrelation introduced by scaling over lagged positions, and time fixed effects are meant to absorb slow-moving cycles and structural change in the ETF and mutual-fund market.

Risk-off shocks lower flows and returns across the distribution, and in nearly every case the fifth-percentile coefficient exceeds the median in absolute value, which exceeds the ninety-fifth. The paper writes the pattern as an inequality on the quantile coefficients and states the interpretation directly: “in nearly every case we consider, the ‘worst’ realizations (in the left tail) change more than the median realization, and the ‘best’ (right tail) realizations change less than the median, lengthening the tails of the distribution” (§3.1, p. 14). Its own summary of the implication is that “the focus in the literature on measures of central tendency is incomplete” (§1, p. 5).

Bond flows shift left with tails pulled apart; equity flows shift left with tails modestly pulled in – and the paper flags the equity case as an important exception to its general finding. For bonds, “risk-off shocks (however measured) increase the worst portfolio outflow realizations more than they decrease median flows, and therefore risk-off shocks significantly fatten the left tail of the portfolio flow distribution. The net effect on bond flows from a risk-off event is that the entire distribution moves to the left” (§1, p. 5). For equities, the largest outflow realisations change less than the median while the largest inflow realisations change more, so “the net result is a leftward shift in the distribution, with a modest shortening of the tails. This is an important exception to the more general findings” (§3.2, p. 15). The practical consequence the paper draws for bonds is that “the lengthening in the left tail causes ’large’ outflow realizations in the unconditional distribution to appear more common in the post-shock distribution.”

From sensitivity to risk aversion specifically, not to physical risk – the structural decomposition separates the two and shows equity flows respond to a physical-risk shock the way bond flows do. Under the Bekaert-Engstrom-Xu decomposition, “in the face of a physical risk shock, the distribution of equity flows reacts in step with the distribution of bond flows in that we observe a leftward shift, with tails lengthening relative to the median. In contrast, a risk-off shock (the change in risk aversion) causes the equity flow distribution to behave as it does with our first RORO measure, with the range of the distribution shrinking as it shifts left” (§3.3, p. 16). The authors read this as explaining the statistical measure’s behaviour: the tail compression “emanates from a sensitivity to changing risk aversion, which causes a convergence of realizations relative to the unconditional distribution.” They also note that for both asset classes “variation in the quantity of risk has a larger impact across the distribution and puts more weight in the left tail compared to risk aversion itself.”

Every return series shifts left with lengthened tails, but equity returns react more than fivefold what fixed income does, and within each class dollar indices react more than local-currency indices. “Across all return types, a risk-off shock shifts the distribution to the left and lengthens the tails, worsening the most negative return realizations more than the median” (§3.2, p. 15). On magnitudes: “a risk-off shock impacts the total return on the equity indices at a rate more than five times the impact on fixed income returns.” Within fixed income, dollar returns decrease “three to six times the rate of the local currency index” – with the honest qualification that “the impact on the local currency index is statistically insignificant,” so the comparison rests on the dollar index reacting significantly (§3.2, fn. 14). For equities, “MSCI USD total returns decrease 28 - 32% more than the local currency equity returns in the face of a risk-off shock.” Under the structural decomposition the two components pull in opposite directions on the shape: “changes in the level of risk dramatically lengthen the tails of the return distribution, while changes in risk aversion shrink the tails,” with risk aversion mattering more for equity returns and the quantity of risk more – “albeit less systematically” – for fixed income (§3.3, pp. 16-17).

Corporate spreads supply most of the leftward shift, funding liquidity supplies the bond funds’ tail-lengthening, and the advanced-economy equity factor can move upper-quantile equity-fund realisations to the right once credit risk is in the model. In the nested regression including all four sub-indices, “much of the baseline’s ‘shifting’ impetus emanates in large part from corporate spreads for both equity and bond funds,” while “the nested model also reveals funding liquidity as a prime source of the bond funds’ ’tails-out’ behavior relative to equity funds” (§3.4, p. 17). Once credit risk enters alongside the advanced-economy equity and volatility factor, the latter “evinces less impact on equity funds and even moves median-plus realizations to the right, reflecting the risk-off behavior of rotating out of equities and into bonds.” For returns, advanced-economy returns and volatility together with corporate spreads drive most of the shift, more so for equities, while “asset classes and currencies all share increased distributional dispersion emanating primarily from credit risk as measured by corporate spreads”; in equity returns the tails-out push from credit risk “counteracts the ’tails-in’ impetus from advanced economy equity return shocks” (§3.4, pp. 17-18). Currency risk is a shift factor, smaller than returns or spreads but contributing more for dollar-denominated returns, while “across asset classes and currency denominations, the impact of funding liquidity barely registers… and almost always statistically insignificant.”

Yes, and the paper treats this as one of its contributions. “While advanced economy equity returns and volatility along with corporate spreads proxying for credit risk were the most significant risk factors during the global financial crisis, movement in corporate spreads and stoppages in funding liquidity predominantly explain capital flows and returns in the aftermath of Covid-19 shock” (§1, pp. 4-5). The corporate-spreads factor in the COVID era is described as “an order of magnitude” larger than in the global financial crisis. On the structural decomposition: “the risk-aversion factor was more prominent in the global financial crisis relative to the quantity of risk, while the opposite prevails during the Covid crisis.” This is the paper’s answer to the literature reporting a weakening VIX-flow relationship: the constituent sources “assert their importance or come to the forefront at different points in time.”

Some, concentrated in institutional money-market funds, and weaker under the structural measure than under the statistical index. Repeating the exercise with the weekly growth rate of assets in US government money-market funds (from Investment Company Institute data) in place of emerging-market flows, the paper finds risk-off RORO shocks drive inflows across the distribution, “particularly… for the right tail, where larger inflows becomes more likely in the face of a risk-off shock” (§3.5, pp. 18-19). Under the Bekaert-Engstrom-Xu decomposition the evidence is thinner: a physical-risk shock has “some positive effect… (although this effect is not consistent across the distribution),” while risk-aversion shocks “drive the left tail of the distribution toward the median, but we do not observe statistically significant impacts elsewhere.” Splitting by investor type, “the largest effects… are associated with institutional money market fund flows. Retail flows are considerably less sensitive.” The authors’ own summary is appropriately hedged: “we do detect across our various specifications some reaction to risk-off shocks in the allocation to safe assets in a manner that complements what we observe for risky assets.”

Applying the fitted quantile coefficients to observed 2020 flows, a shock at the peak of the COVID episode moves bond-flow quantiles down by roughly $160-$260 million a week depending on the quantile, and pulls the bond tails apart while pulling the equity tails in. The exercise adds the estimated quantile coefficient times the shock times average assets under management to the observed 2020 quantile of weekly country flows (§4, pp. 19-21, Table 5). For bonds, a one-unit shock turns a pre-shock 2020 median weekly inflow of $3.7 million into a median reallocation that is an outflow of $14.09 million; it increases fifth-quantile outflows by $22 million a week against $17.8 million at the median, and reduces ninety-fifth-quantile inflows by $14.8 million. At 3.1 units the corresponding figures are $68.03 million, $55.2 million and $43.5 million. The index reached 11.56 standard deviations at its COVID peak, which implies falls of $256.7 million, $205.9 million and $162.14 million at the fifth, fiftieth and ninety-fifth quantiles – “a shock of this size pulls the tail realizations apart by $92 million.” For equities the ordering reverses: a one-unit shock raises fifth-quantile outflows by $16.28 million against $18.8 million at the median and $21.4 million at the ninety-fifth; at the peak shock the three quantiles fall by $188.2 million, $217.7 million and $247.3 million, so the shock “pulls the tail realizations in by $59 million,” and “under the peak shock, even the ‘best’ realizations manifest as equity fund outflows.” A caution for the reader: the introduction reports a different set of COVID magnitudes for what appears to be the same exercise – $45 million (bonds) and $6.8 million (equities) of tail expansion/compression for a one-standard-deviation shock, and $170.2 million and $25.4 million at the largest COVID observation (§1, p. 6) – which do not reconcile with the Section 4 figures above. This record reports the Section 4 derivation and flags the discrepancy rather than choosing between them.

It converts the quantile coefficients into a statement about probability: after a risk-off shock, what used to be a tail event becomes an ordinary one. Fitting kernel densities to predicted values for a three-unit shock – chosen because three units “represents the threshold of the 10th percentile among risk-off shocks in 2020” – the fitted fixed-income flow distribution has longer tails and more skew toward outflows than the unconditional one (skewness moving from -0.61 to -1.01), and “in the face of a 3-unit shock, what would be a tail event in the unconditional distribution looks more like a 10th quantile shock, and therefore more probable. The post-shock median now falls in the bottom 25% of pre-shock realizations” (§4, p. 21). The equity flow distribution’s skewness is roughly unchanged, but its post-shock median also falls in the bottom quarter of pre-shock realisations. Returns move more: “what the unconditional equity distribution would label a tail outcome lay near the median in the post-shock distribution,” while for EMBI returns the pre-shock tail event “still falls within the interquartile range of the unconditional distribution.” Decomposing by component, risk aversion and risk contribute roughly equally to the extra left-tail mass in bond flows, risk is the greater force pushing equity outcomes past the fifth quantile, and risk aversion matters more for worsening return distributions, especially bond returns.

Whether the effects differ across recipient countries, and how much the growth of passive vehicles constrains the answer. The authors flag country-level heterogeneity as the natural next question, citing Gelos et al. (2019) on recipient-country policy and fundamentals, and framing the underlying issue as “the extent to which fund managers view emerging markets as a single asset class or whether country fundamentals matter for fund allocations” (§5, p. 22). But they identify a structural obstacle: passive index funds and ETFs, “representing about half of the emerging market fund space towards the end of the sample,” have very little manager discretion, so their reallocation “may induce elevated correlations among countries and minimize the effect of cross-country heterogeneity,” whereas actively managed funds have considerable discretion. The stated plan is to examine the vehicles themselves rather than flows alone.