Capital flow waves: Surges, stops, flight, and retrenchment

Previous work on extreme capital-flow episodes relied on proxies for net capital flows, which cannot distinguish a withdrawal by foreign investors from an inflow created by domestic investors bringing money home. The authors note that “almost all previous work on capital flow episodes relied on proxies for net capital flows, which cannot differentiate between changes in foreign and domestic behaviour” (§1, p. 2). They argue the older focus was defensible for its time: “in the early and mid-1990s net capital inflows roughly mirrored gross inflows, so the capital outflows of domestic investors could often (but not always) be ignored” (§1, p. 2). What changed is that gross flows grew larger and more volatile while net flows became more stable, so net flows can no longer be read as driven solely by foreigners. The paper also widens the object of study: rather than examining one episode type in isolation, it treats surges, stops, flight and retrenchment together “in order to understand the full cycle of international capital flows” (§1, p. 1).

An episode is a run of consecutive quarters in which the year-over-year change in four-quarter gross flows stays more than one standard deviation from a rolling historical mean, provided it crosses two standard deviations at some point and lasts more than one quarter. The three criteria are stated together in §2.3 (p. 7): the change must exceed two standard deviations above or below the historical average in at least one quarter of the episode; the episode runs over all consecutive quarters for which the change exceeds one standard deviation; and the episode must be longer than one quarter. The historical mean and standard deviation are computed over the trailing five years (20 quarters), “which means that episodes are always defined relative to the recent past,” and a country must have at least four years of data to compute a historic average (§2.3, fn. 9). Summing over four quarters both matches the older literature’s annual focus and removes seasonality. Flight and retrenchment use gross outflows with the balance-of-payments sign convention in which domestic acquisition of foreign assets is negative, so flight (more money sent abroad) is gross outflows falling below the band and retrenchment is gross outflows rising above it (§2.3, p. 8). The authors are candid that operationalising “marked slowdown” requires “several ad hoc decisions” – a slowdown relative to what, and how sharp (§2.2, p. 6) – and that “several methodologies can be used to identify capital flow episodes; each has advantages and disadvantages” (§2.3, p. 6).

Quarterly gross inflows and outflows for 58 emerging and developed economies, drawn from the IMF’s International Financial Statistics (accessed through Haver Analytics in January 2012), running from 1980 at the earliest through 2009, with unbalanced start dates. Thirty countries have data from 1980, 37 from 1990, 52 from 1995, and all 58 by 2000 (§2.3, pp. 8-9). Baseline gross inflows are direct investment plus portfolio plus other inflows; gross outflows are the analogous sum with reserve accumulation omitted, so the baseline is private outflows (§2.3, p. 9). The 2007 cross-section captures $10.8 trillion of gross capital inflows, “capturing 97% of global capital inflows recorded by the IMF” against reported worldwide financial-account liabilities of $11.2 trillion (§2.3, p. 9, and fn. 12). China is absent because it “only recently began to publish quarterly capital flow data” (fn. 10). The regression sample narrows to 54 countries because of explanatory-variable availability (fn. 7), and the regression period is 1985-2009 (§3.2, p. 16). One further caveat the authors raise themselves: the inflow/outflow split is determined by the residency of the asset, so tax-haven booking can misattribute flows, and “to our knowledge, no residency-based system can get around this issue” (fn. 6).

From 1980 through 2009 the method identifies 167 surge, 221 stop, 196 flight and 214 retrenchment episodes, with an average length of roughly one year each (§2.4, p. 9). Stops are slightly more common than surges, but surges last longest at an average 4.5 quarters and retrenchments shortest at 3.9 quarters. Because the two-standard-deviation cutoff is what it is, “a country’s gross flows will be in an episode about one-third of the time,” but the cross-country dispersion is large: Argentina was in a surge only 12 percent of the time between 1985 and 2009 but in a stop in one quarter out of four, while India was almost twice as likely to be in a surge as in a stop (§2.4, p. 10). The Brazil illustration in §2.3 (p. 8) shows how the dating maps onto recognisable events: stops in 1993Q1-1993Q3 (hyperinflation), 1995Q1-1995Q2 (the Mexican peso crisis), 1999Q1-1999Q2 (the Brazilian devaluation) and 2008Q2-2009Q3 (the crisis); surges in 1990Q2-1991Q1, 1994Q1-1994Q3, 1995Q4-1996Q2 and 2006Q3-2007Q4.

At the height of the crisis the net-flow method identifies 13 surges and 22 stops; the gross-flow method identifies 1 surge and 47 stops over the same two quarters. Comparing 2008Q4-2009Q1, the authors report that net flows find thirteen surges where gross flows find only one – Bolivia, and only because a surge that began in 2007 was ending in 2008Q4 – and that net flows find “less than half as many” stops, 22 against 47 (§2.5, pp. 10-11). The reason is mechanical and, once stated, obvious: “each country defined as having a surge episode based on the net flows data – but not using the gross data – had a retrenchment episode.” When domestic investors sold foreign holdings and repatriated the proceeds, net inflows rose, and the older method read that as foreign money arriving. Chile is the worked case: during the crisis its gross inflows dropped sharply while its gross outflows also fell sharply, so the net measure calls it a surge and the gross measure calls it a stop combined with a retrenchment (§2.5, p. 11). Retrenchment by locals also masked genuine stops: “even though foreign capital inflows suddenly stopped, retrenchment meant that net capital flows did not fall enough to qualify as a ‘sudden stop’ episode based on the older methodology.”

Episode incidence moves in waves that sweep across the sample at once, and the pairwise correlations line up with global rather than country-specific drivers. Figure 3 shows years in which no country has a stop or a retrenchment and others in which a majority of the sample does; the crisis produced “an unprecedented number of countries” in stops and retrenchment, with the incidence of sudden stops spiking to 78% of the sample in 2008Q4 (§2.6, pp. 11-12). The authors are careful that this co-movement is not universal across crises: after the collapse of LTCM in 1998Q4 the incidences of stops and retrenchment were elevated but much lower, at 35% and 19%, and by 1999Q3 retrenchment had fallen rapidly to 6% while stops fell more slowly to 15%. The correlation table reinforces the reading: stops with retrenchments at 0.47 and surges with flight at 0.37 are “consistent with an important role for global factors,” while the correlations one would expect if country-specific factors dominated – stops with flight, surges with retrenchment – are in fact negative (§2.6, p. 12). Against this the authors note genuine cross-country heterogeneity: during late 2008 and early 2009 Poland had a retrenchment episode while Russia did not, and a diverse list including Argentina, Australia, Brazil, Greece, India, Indonesia, New Zealand, Norway, Portugal, the Slovak Republic, South Africa and Turkey also had none, “suggest[ing] that even in the presence of substantial global shocks and possibly regional contagion, domestic characteristics may also be important” (§2.6, p. 13).

A complementary log-log model of the probability of each episode type on one-quarter-lagged global, contagion and domestic variables, estimated as a system allowing cross-episode error correlation, with standard errors clustered by country. The cloglog is chosen because episodes are rare – “83 percent of the sample is zeros” – which makes the cumulative distribution function asymmetric, so the estimation assumes the extreme-value cdf F(z) = 1 - exp[-exp(z)] (§3.2, p. 16). Each episode type is estimated separately, but “we use a seemingly unrelated estimation technique that allows for cross-episode correlation in the error terms… without assuming a structural model specifying a relationship between episodes.” The global variables are the VXO for risk, year-over-year growth in a global money supply (US, euro-area and Japanese M2 plus UK M4, in dollars) for liquidity, an average long-term government bond rate across the US, core euro area and Japan for global interest rates, and quarterly global real activity growth. Contagion enters through the regional dummy, the export-weighted trade-linkage measure and the BIS-based financial-linkage measure. The domestic variables are stock-market capitalisation to GDP for financial depth, the Chinn-Ito KAOPEN index (sign reversed so higher means more controls) for capital controls, a growth shock measured as the deviation of actual from trend growth, public debt to GDP, and GDP per capita (§3.2.1-3.2.3, pp. 16-18).

Global risk is the only variable consistently associated with all four episode types; contagion is strongly associated with stops and retrenchment; domestic variables are mostly weak and often not robust. Higher global risk is positively correlated with stops and retrenchment and negatively correlated with surges and flight, though for flight only at the 10% level in the baseline (§3.3, p. 19). Among the other global factors, “strong global growth is associated with a higher probability of surges and a lower probability of stops and retrenchment,” and higher global interest rates are usually correlated with retrenchment. Countries are more likely to have a stop or retrenchment if their major trading or financial partners just had the same episode, and more likely to have stops and flight if their neighbours did. On the domestic side the baseline finds stops more likely after a negative growth shock, surges more likely after a positive one, retrenchments more likely in richer countries, and flight more likely where capital controls are greater and debt lower – but the authors immediately caution that “many of these results are not robust across alternative specifications.” They also flag the null results as substantive: there is no evidence capital controls reduce the likelihood of a surge or stop, and “global liquidity and global interest rates are not significantly related to most extreme capital flow episodes,” the exception being retrenchment and high global rates. Flight is the hardest episode to explain: in a simple logit the pseudo-R-squared is only 0.04 for flight, against 0.07 for surges, 0.13 for retrenchments and 0.15 for stops (§3.3, fn. 23).

The central result disappears: global risk is not significantly related to either surges or stops when the episodes are built from net flows, and some coefficients flip to counterintuitive signs. Re-running the same specification on net-flow-defined episodes gives “starkly different results” (§3.4, p. 20): global risk, significant for every gross-flow episode type, is insignificant for net-flow surges and stops, and the net-flow estimates suggest countries are more likely to have stops if they have more stringent capital controls and lower debt ratios. The authors give the mechanism: lower global risk raises both foreign inflows and domestic outflows, and because the two large movements partly cancel, the change in the aggregated net flow is small (§3.4, p. 20). This is the paper’s sharpest methodological point – the older measurement was not merely noisier but could mask the relationship entirely.

Both appear to matter, but the measure that most cleanly isolates risk aversion is significant only for stops and surges, not for the full cycle. The paper re-estimates with four risk proxies: the VXO (baseline), the VIX, the Credit Suisse First Boston Risk Appetite Index, and the Variance Risk Premium (§3.5, pp. 21-22). The VXO and VIX “capture both economic uncertainty as well as risk aversion”; the RAI aims at risk aversion alone but, following Misina (2003), may not control for overall risk unless strict theoretical conditions hold; the VRP “is based on a less rigid set of assumptions and therefore is a more accurate measure of risk aversion independent of expectations of future volatility.” The broad measures are positively correlated with stops and retrenchment and negatively with surges. The VRP is positively and significantly related to stops and negatively to surges, which the authors read as showing “that risk aversion (and not just increased economic uncertainty) is an important factor associated with stop and surge episodes.” Throughout, the coefficient on risk remains highly significant “except in regressions for flight episodes, which continue to be more idiosyncratic.” The VXO and VIX are 99% correlated, and the VXO is used as baseline to maximise sample length since the VIX begins only in 1990 (fn. 17).

On this evidence, no: across seven different measures of controls or integration, the coefficient on controls remains insignificant for surges and stops, the episodes controls are typically aimed at. Beyond the baseline Chinn-Ito index the paper tries a de facto integration measure (foreign assets plus liabilities over GDP), Schindler’s (2009) broad capital-account restriction index for 1995-2005, that source’s separate inflow- and outflow-specific indices, and two Ostry et al. (2011) indices covering financial-sector controls and foreign-exchange regulations (§3.5, pp. 22-23). Correlations among these measures are low, “in part because they measure different aspects of controls.” The single exception significant at 5% runs the other way and is not really a control measure: countries more integrated through foreign assets and liabilities are less likely to experience stops. FX regulations are negatively associated with surges only at the 10% level, and that result “is not robust to several sensitivity tests.” The authors state the upshot plainly: “There is no evidence that reduced integration with global financial markets, including through the use of capital controls, is associated with a reduced vulnerability to episodes caused by foreigners” (§3.6, p. 25).

Global risk survives everywhere except sometimes flight; the contagion results survive; the domestic results largely do not. The tests drop the crisis window 2008Q3-2009Q2, add country fixed effects, substitute probit and logit for cloglog, estimate equations in isolation, add controls for youth and old dependency ratios, an exchange-rate peg dummy, sovereign credit ratings, terms of trade and reserves to GDP, swap in alternative measures for global interest rates, global liquidity, financial-system size and strength, and the growth shock, and redefine episodes with an HP filter and 30% deviations from trend, with a three-standard-deviation cutoff, excluding monetary-authority transactions from 2008Q3, including reserves in outflows, and including errors and omissions (§3.6, pp. 23-25). Focusing on 5% significance, “global risk is significantly related to surge, stop, and retrenchment episodes in each of the robustness tests, and is often (although not always) significantly related to flight episodes.” Trade and financial linkages remain important for stops and retrenchment, with geography adding for stops. The domestic variables “continue to show less consistent patterns,” and for several of them “the significance of each of these results, and often the sign, however, fluctuate across specifications.”

They read the primacy of global risk as support for the recent class of models built on global shocks, and as evidence against models that put the weight on liquidity or interest rates in a major economy. “Our finding that the primary factor associated with capital flow episodes is changes in global risk supports this recent theoretical focus on global factors, especially risk,” they write, while “the results, however, do not support the widespread presumption that changes in global liquidity or interest rates in a major economy, such as the United States, are important factors driving surges in capital flows (independent of any effect on global risk and growth)” (§1, p. 3). On productivity-shock models they are deliberately split: an emphasis on domestic productivity shocks “might be relevant in explaining gross capital inflows, but appears to be less applicable in explaining the volatility in domestic residents’ international investment; a country’s economic growth is associated with surges and stops but not episodes driven by domestics,” which the conclusion describes as “mixed support” for the real-business-cycle literature (§1, p. 3; §4, p. 26). They also note that a lax monetary policy has been found to lower risk aversion with a lag of about five months, “so it is possible that the risk measure may be capturing a lagged effect of monetary policy” (§4, fn. 34) – a caveat that limits how far the global-risk finding can be read as independent of major-economy policy.

That governments should concentrate on withstanding capital-flow volatility rather than trying to suppress it – stated as an implication of correlations, with an explicit disclaimer about causation. The authors write that “while we do not assess causation in this paper, our results suggest that most domestic factors only have a limited correlation with capital flow volatility,” and that “we find no evidence that capital controls insulate an economy against capital flow waves” (§1, p. 4). Because most of what is associated with episodes – global risk, global growth, contagion – is “outside the control of policymakers in most countries,” they conclude that “governments may wish to focus more on strengthening their country’s ability to withstand capital flow volatility rather than to attempt to directly reduce this volatility,” and that the role for global factors and contagion “suggests an important role for global institutions and cross-country cooperation” (§4, pp. 26-27). The costs motivating the exercise are drawn from prior work rather than estimated here: surges have been found correlated with real-estate booms, banking crises, debt defaults, inflation and currency crises, and sudden stops with currency depreciations, slower growth and higher interest rates (§1, p. 4).

That the disaggregation is not a refinement but a precondition – the older aggregate data could not recover the dynamics or the correlates of capital-flow waves. “By differentiating gross inflows from gross outflows, our analysis shows that many episodes previously identified as ‘surges’ of foreign investment are actually driven by the retrenchment of domestic residents. Similarly, the earlier methodology missed periods of sudden stops in foreign capital inflows when these stops occurred simultaneously with an increase in global risk and retrenchment by domestic investors” (§1, p. 3). The closing section adds the behavioural reason the split matters going forward: domestic and foreign agents can respond differently to the same shock, “due to factors such as different exposure to the domestic exchange rate or different degrees of access to liquidity,” so their responses may offset each other and stabilise net flows, magnify each other and destabilise them, or only one group may respond at all (§4, p. 27).