Spillovers at the extremes: The macroprudential stance and vulnerability to the global financial cycle

The literature on macroprudential leakages and the literature on extreme events, because a policy could be invisible on average and decisive in the tails (Section 1, pp. 1-2). On the first: macroprudential policy has “some success in accomplishing specific domestic goals (such as moderating credit growth or foreign currency-denominated borrowing)” but is “less effective by other metrics (such as stabilizing cross-border capital flows),” and generates “spillovers and leakages that shift risks outside the regulated banking sector – particularly to corporate bond markets and the broader ‘shadow’ financial system (Ahnert et al., 2021; Avdjiev et al., 2020; Forbes, 2019).” The consequence the paper wants to test is stated as a conditional: “If the magnitude of these spillovers and leakages is large enough, and the corresponding risk exposure shifts to financial intermediaries that are more vulnerable to shocks, macroprudential regulations could undermine, rather than mitigate, financial sector vulnerabilities during certain periods. There is little systematic analysis, however, of what these spillovers and leakages imply over different phases of the global financial cycle.” On the second: work on extreme capital flow and return events (Bergant et al. 2020; Chari et al. 2020; Eguren-Martin et al. 2020; Gelos et al. 2019) suggests “some policies may have minimal impact during stable periods but are highly effective at mitigating vulnerabilities during extreme events,” so “assessing the effectiveness of macroprudential policies may require focusing on the distribution of outcomes – and not just on average effects or during stable periods.”

Marginal effects across the risk distribution, high-frequency data, intensity-based stance measures, and a policy-shocks identification – and the authors say the last depends on the others (Section 1, pp. 1-2). First, “we analyze the marginal effects of policy choices at different points in the risk distribution to test how relationships change across the global financial cycle. This methodology highlights how the standard practice of focusing on averages across the cycle can overlook highly consequential relationships that prevail during ’extreme’ periods.” Second, weekly data, because macroprudential policy targets “sharp (and often short-lived) events … that are challenging to identify using standard data available at a quarterly (or annual) frequency.” Third, new measures “to capture the intensity of existing regulations,” which “improves most work focusing on dummy variables of recent policy changes.” Fourth, the policy-shocks approach for reverse causality, and the interdependence is explicit: “This methodology can successfully identify and estimate the exogenous component of the macroprudential stance because of the other innovations in the paper: the higher frequency of the data and the more accurate measure of the intensity of the macroprudential stance.”

Only an illustrative partial-equilibrium example in an appendix, and it is framed as motivation rather than proof (Section 2.2, pp. 31-33). “A full model of these various channels is beyond the scope of this paper, but Appendix 1 sketches a simple partial-equilibrium example to motivate the key results found in the empirical analysis.” The example “shows how macroprudential regulations as tighter FX capital requirements or stricter CCyB regulations, have minimal effects on banks, firms, and investors in modest risk-off states but can have significant effects in extreme risk-off states. During these extreme risk-off states, macroprudential regulations can achieve the desired outcome of meaningfully increasing bank resilience (by allowing banks to avoid bankruptcy) but simultaneously make bond investors less resilient as firms shift their borrowing from banks to bond markets.” It also generates the tool-specific prediction the empirics later confirm: “certain types of targeted regulations (such as on FX exposures) can lead bond investors to incur more significant losses if they create incentives for greater reductions in bank lending and more risk-shifting onto investor balance sheets.” The authors also acknowledge why they stop short of a full model: the literature “highlights … the complexity of modeling these channels – especially given the multiple potential direct effects, leakages, and spillovers from different types of macroprudential regulations.”

By cumulating policy changes since 2000 and, where possible, using the actual levels at which regulations are set – because the usual dummy variables record only that a rule changed, not how tight it is (Section 3.1, pp. 34-47). Two sources are combined. BIS and ESRB data on the countercyclical capital buffer give “a quantitative measure of the stringency of the regulation that is comparable across countries” through late 2020, but cover only one tool. The IMF’s iMaPP database (Alam et al. 2019) covers 134 countries monthly from 1990 to 2018 across 17 instrument types, but “tracks when the tools are tightened or loosened using dummy variables for each measure,” which “have the drawback … of only capturing when a regulation was changed, with no information on the overall intensity of the regulation or magnitude of the change. The sole exception is for loan-to-value (LTV) ratios.” Aggregating iMaPP changes from 2000 – chosen because “the use of these tools was limited, so each country can be assumed to start from a similar, neutral stance” – gives stances “rang[ing] from -7 to 72 across 72 countries, with a higher value indicating a tighter stance and a panel median of 0 and a mean of 2.3.” China is tightest at 72, then South Korea (41), Russia and Hong Kong (both 40); Iceland loosest at -7, then India and Argentina (-6). By 2018Q4 “the mean stance was 15 tightenings and mean CCyB was 0.21%.”

No, not on its own – the paper names the specific way it misleads and builds its preferred index around the two tools that can be measured in levels (Section 3.1, pp. 42-47). The problem is measurement, not data quality: “some of these cross-country differences, especially near the end of the sample, reflect different approaches toward adjusting macroprudential policy rather than fundamentally different intensities of their stances. For example, China tends to make frequent but small adjustments to its macroprudential tools, which aggregate to many net tightenings and what appears to be a very tight macroprudential stance by this measure. In contrast, other countries (such as the UK) tend to adjust macroprudential policy less frequently but in larger increments, which results in what seems to be a significantly weaker stance” (citing Forbes 2021). The response is the preferred Broad Intensity Index: an equally weighted combination, each component scaled by its standard deviation, of the CCyB, the LTV ratio expressed as 100 minus LTV so that higher means tighter, and the cumulated FX macroprudential stance. Its advantages are “incorporating the two best intensity measures of macroprudential policy comparable across countries (the CCyB and LTV ratio),” capturing 2020 COVID adjustments through the CCyB, and covering “three of the most widely used tools that target different risk exposures: countercyclical risk in banks, the housing sector, and foreign currency.” Its disadvantage is stated too: “the measure does not incorporate other tools that could be an important part of the macroprudential framework in certain countries.” The Narrow Intensity Index (the first principal component of the CCyB and LTV) “let[s] the data speak” but drops FX regulations, “which are an important part of the macroprudential toolkit for many emerging markets.” Even the LTV ratio comes with a caveat: “Different countries can use different definitions and have different coverage for their LTV ratios, so that they are not directly comparable across countries – albeit still a better measure of relative intensities than dummy variables.”

Because a broader measure captures more of the risk-responsive asset prices, and the choice turns out to matter for equities (Section 3.2, pp. 47-49; Section 5.5, pp. 113-114). The RORO index is the first principal component of standardized daily changes in credit risk (ICE BofA BBB option-adjusted spreads for the US and euro area; Moody’s BAA yield over 10-year Treasuries), advanced-economy equity risk aversion (the additive inverse of daily returns on the S&P 500, STOXX 50 and MSCI Advanced Economies index, plus the VIX and VSTOXX), funding liquidity (average daily changes in the G-spread on 2-, 5- and 10-year Treasuries, the TED spread, the 3-month LIBOR-OIS spread and the bid-ask spread on 3-month Treasuries), and finally the trade-weighted dollar against other advanced economies and the spot gold price. Components are signed so positive means risk-off. “The distribution is skewed with long tails toward risk-off, indicating that large risk-off events occur more frequently than large risk-on events.” Replacing RORO with the VIX leaves bond results robust but kills significance for equities, and the authors attribute the difference to coverage: “the RORO measure includes a broader set of risk-responsive asset prices than the VIX, such as center-country equity returns, corporate spreads, gold prices, other option-implied volatilities, and several different spreads intended to capture liquidity risk.”

Weekly EPFR fund-level allocations – high frequency and free of valuation effects, but not cross-border flows and not all investors (Section 3.3, pp. 49-51). EPFR Global’s Country Flows dataset combines fund flow data with country weightings for “more than 14,000 equity funds and over 7,000 bond funds, with more than $8 trillion of capital under management,” and “because the country flows comprise the sum of fund-level aggregate re-allocations, they come cleansed of valuation effects and represent real quantities.” The limitations are stated plainly: “this dataset does not focus on cross-border capital flows (as it includes domestically domiciled funds) and does not include all portfolio investors (such as sovereign wealth funds and hedge funds),” though “the flows have significant predictive content for lower frequency, aggregate data on international portfolio flows (Koepke and Paetzold, 2020).” Flows are scaled by the previous week’s holdings with the lagged scaled variable included as a control, “ensur[ing] that larger countries with larger capital flows do not mechanically drive the analysis,” and are winsorized at the 0.5 and 99.5 percent levels “to prevent several large outliers (that appear to be errors) from driving the results.”

A first stage regresses the stance on eighteen candidate drivers, keeps a parsimonious significant subset, and uses the residual – and the first-stage F-statistics are around 100 (Section 3.4 and Section 4.1, pp. 51-67; Section 4.2, p. 68). The reverse-causality problem is concrete: “a sharp increase in portfolio flows could raise concerns about domestic financial stability risks, leading policymakers to tighten macroprudential regulations. The resulting positive correlation between portfolio flows and the macroprudential stance would be amplified during large risk shocks when policymakers are likely to pay closer attention to large moves in portfolio flows.” Most papers lag the policy variable instead, “but this approach is unlikely to address endogeneity concerns fully (Forbes, 2021).” The eighteen first-stage variables fall into Crisis (recent domestic crisis, banking-sector distance-to-default z-score, counts and intensity of crises abroad), Credit (cross-border borrowing ratio, domestic credit growth, property price growth), Growth (real exchange rate appreciation, forecast and realized GDP growth, lagged CPI inflation), and other macro/institutional characteristics (Chinn-Ito financial openness, FX volatility, ICRG institutional quality, the policy rate, the differential versus the fed funds rate, and a fixed exchange rate dummy), with country fixed effects; a backward-and-forward inclusion procedure at a 10 percent threshold narrows the set. First-stage findings: a tighter stance corresponds to “a long time since a domestic banking crisis, fewer countries in crisis …, more cross-border borrowing, faster domestic credit growth, slower domestic growth, lower domestic policy rates, and a larger interest rate differential with the global rate.” On why the fit is good where earlier work struggled, the paper credits its three other design choices: estimating the stance rather than the timing of changes, since “attempting to predict the timing of policy changes can be challenging as many hard-to-measure factors can affect the precise timing of policy adjustments (including political events, institutional structure, pre-set meeting dates, and so on)”; measures that capture intensity; and higher-frequency financial data.

Weekly flows on the policy shock, RORO, their interaction, push and pull controls, country and year fixed effects, with 10,000 bootstrap replications over both stages clustered by country (Section 4.1, pp. 61-67; Section 3.4, pp. 55-58). Push variables are the advanced-economy monetary stance (a shadow short rate) and industrial production growth, both GDP-weighted across the US, Japan, UK and euro area, with year fixed effects added “to control for other slow-moving aspects of the business cycle and for changes to the mutual fund and ETF industries over time.” Pull variables are the bilateral dollar exchange rate, the interest differential with the US, domestic real GDP growth, FX volatility, Chinn-Ito financial openness and ICRG institutional quality, all lagged except FX volatility. The sample excludes the United States, Japan and Switzerland, “as the relationships between risk shocks and capital flows that are the focus of this paper would likely differ for these safe-haven countries,” leaving 65 countries. One design feature the paper defends explicitly is mixing frequencies – the first stage uses lower-frequency country characteristics while the second uses weekly flows and risk – on the grounds that “analyzing joint processes at a common low frequency would ignore relevant information that may be available at mixed frequencies and lead to model misspecification” (fn. 22, p. 16, citing Ghysels 2016).

Risk matters a lot, the stance alone matters not at all, and the interaction matters modestly at the mean (Section 4.2, pp. 67-82). Risk-off shocks are “associated with sizable and statistically significant declines in portfolio bond flows across all macroprudential measures,” while “tighter macroprudential policy (ignoring the interaction with risk) is not significantly correlated with bond flows” – both consistent with prior work, as are the push and pull coefficients, with global variables “more consistently significant.” Quantitatively, an increase of one unit in RORO – which “is equivalent to one standard deviation and close to the 90th percentile of the distribution for the full the sample” – “corresponds to an 0.09%-0.10% decline in weekly bond flows, equivalent to -$2.3 to -$2.4 billion (based on AUM at the start of 2020),” while the unconditional macroprudential effect is about one-tenth of that and insignificant. The interaction term is “negative and usually significant (for three of the four macroprudential measures, including the two preferred Intensity indices),” but small in magnitude: a one-unit tighter stance combined with a one-unit risk increase “correlates to a further decline in bond inflows of about $151-$543 million (0.01% to 0.02%).” The authors note this cuts against an intuitive prior: “macroprudential regulation in place at the time of a high-risk event could aggravate the impact of the shock, contrary to what we might have expected if tighter regulation moderates the build-up of risks during booms and moderates the unwinding of risk exposures during risk-off episodes.”

An order of magnitude, and enough to roughly double the unconditional effect of an extreme risk shock (Section 4.2, pp. 76-82). Evaluated across the RORO distribution, “the smaller estimates near the mean obscure larger and usually significant relationships in the tails,” with positive marginal effects at the risk-on end and negative ones at the risk-off end – amplification at both ends – “highly significant across our preferred indices of the macroprudential stance and only insignificant using the time-relative measure.” For the Broad Intensity Index, a one-unit tightening “corresponds to bond flows statistically indistinguishable from zero when risk is at the median level, but a decline in flows of -$636mn, -$1,529mn, and -$2,076mn when risk is at the 95th, 99th and 99.5th percentiles of the distribution, respectively,” set against an unconditional risk effect of about -$2 billion. Risk-on effects are “meaningful but smaller at the extremes,” at +$631 million, +$969 million and +$1,215 million at the 5th, 1st and 0.5th percentiles. Put in event terms, a 99th-percentile shock (a RORO of 3.49, seen “during the Global Financial Crisis in 2008 and 2009, the Euro crisis in 2011, and during the COVID pandemic”) produces $7.9 billion of bond outflows for a country with the stance index at zero, against $11.5 to $15.8 billion one unit higher – “amplif[ying] the impact of risk-off shocks on bond outflows by about 22%-87% (based on all four measures of the macroprudential stance, or by 45%-67% for our preferred two intensity indices).” (Readers should note the Introduction quotes a wider range, “30%-96%,” for the corresponding comparison; the Section 4.2 figures are the ones tied to the reported table.)

Yes, and the policy-shocks approach turns out to be the more conservative choice in terms of power rather than a source of the result (Section 4.3, pp. 82-84). Replacing the constructed policy shock with a simply lagged measure of the stance leaves “coefficient signs and significance patterns … qualitatively similar,” and “the estimated interactions between the macroprudential stance and risk at different points in the risk distribution continue to suggest that a tighter stance amplifies the impact of risk shocks on bond flows.” The difference: “the policy shocks approach has greater power and usually delivers larger coefficient estimates at both the mean and the margins of the risk distribution.” So the identification strategy sharpens rather than manufactures the finding.

LTV ratios, FX measures and bank-credit-supply measures; not the countercyclical capital buffer, and not demand-side measures (Section 4.4, pp. 84-93). Repeating the analysis for five granular measures – the LTV ratio and CCyB (both expressible in comparable magnitudes) plus cumulated FX, Demand and Supply measures – “adjustments in LTV ratios, FX Measures, and Supply Measures correspond to the results for the aggregate macroprudential measures; they amplify the impact of risk shocks, particularly for extreme ‘risk-off’ shocks. The CCyB and Demand Measures appear to work in the same direction, but the effects are not significant, including at the extremes.” The CCyB null is given a reason rather than left as noise: it “is a policy focused on moderating the impact of the financial cycle on banks. It adjusts bank capital buffers across the cycle, such that buffers should be higher during risk-on periods and lower during risk-off periods,” so “it is not surprising that the CCyB does not significantly amplify the impact of risk shocks” – consistent with Buch et al. (2019), which “finds that macroprudential regulations focused on general capital requirements tend to have smaller spillover effects than those focused on specific sectors.” A decomposition supports this: splitting Supply Measures, the interaction “is not significant for Capital Measures (which are more cyclically focused and include the CCyB), but is significant for the other subcomponents.”

It points to documented risk shifting, borrowed from other papers rather than observed here (Section 4.4, pp. 89-91). For FX measures, “Ahnert et al. (2021) document that tighter FX regulations on banks reduce bank lending and borrowing in FX but then cause companies to shift to other sources of cheaper FX credit, especially through issuing bonds sold to non-bank investors. Their underlying model shows that this shift away from bank loans occurs in riskier firms less well hedged against currency risk – a shift which would make bond flows more sensitive to global financial conditions.” For LTV ratios, “Sveriges Riksbank (2012) provides a concrete example …: When Sweden increased LTV limits on secured lending, making it harder for borrowers to purchase homes with mortgages secured by property, there was an increase in unsecured loans. These unsecured loans, which are then often packaged and sold to bond investors, are likely to be more sensitive to risk shocks than those backed by assets.” This is the honest boundary of the paper’s contribution: it establishes that portfolio flows in tightly regulated countries are more risk-sensitive, and attributes that to risk shifting on the strength of other work’s micro evidence.

Yes, and it is the paper’s cleanest internal evidence that measuring intensity matters (Section 4.4, pp. 91-92; fn. 2, p. 4). Demand Measures “primarily consist of changes in LTV and DSTI ratios,” yet the LTV ratio’s interaction is significant and Demand Measures’ is not. “The difference in results reflects that the LTV measure is a precise magnitude measuring the intensity of the LTV ratio, while Demand Measures is the sum of dummy variables for any past changes in these housing-related ratios.” The direct test confirms it: recomputing an LTV statistic from dummy variables for past changes, “in which case the interaction term becomes insignificant (including at the extremes).” The authors draw the general lesson: “Although summing dummy variables of past policy changes may create a better measure of the policy stance than simply focusing on whether a policy was changed recently, it does not fully capture that policy’s intensity. This lack of precision introduces noise, making it more difficult to estimate any relationship between the macroprudential stance and risk.”

Same pattern for equities with smaller coefficients but larger dollar amounts; no significant difference between emerging markets and advanced economies; stronger for dollar-denominated flows (Sections 5.1-5.3, pp. 95-107). For equities, the interaction is “negative and … usually significant when the macroprudential stance is measured using one of the indices,” with coefficients “usually smaller in magnitude than those for bond flows” – which the authors explain as expected, “given that many macroprudential measures explicitly target excessive leverage in the economy and would therefore be more likely to affect debt than equity investments.” But the dollar amounts are bigger because equity portfolios in the data are larger: a one-unit tightening gives equity outflows of $740 million, $1,859 million and $2,545 million at the 95th, 99th and 99.5th percentiles against an unconditional effect of -$3.8 billion, and inflows of +$845 million to +$1,577 million at the risk-on extremes; at a 99th-percentile shock the amplification is “about 20%-127% (or 20%-48% for our two intensity indices).” On country groups, the triple interaction of risk, stance and an emerging-market dummy is “statistically insignificant for both types of capital flows” and the marginal effects “are not statistically significant at any points in the risk distribution,” so “the effect of macroprudential regulations during periods of extreme risk-on or risk-off sentiment does not differ between emerging markets and advanced economies in a statistically significant manner.” On currency, “dollar-denominated flows (for both equities and bonds) respond more strongly to risk-on/risk-off shocks,” and amplification is significant “for all dollar-denominated flows and non-dollar bond flows,” while “the relationships follow the same patterns for non-dollar equity flows but are insignificant and smaller.”

No – they behave similarly but do not displace the macroprudential result (Section 5.4, pp. 106-109). Substituting a capital-controls measure from Fernandez et al. (2015, updated to 2017) for the macroprudential stance gives “similar” results: “capital controls have no independent, significant impact on bond or equity flows (ignoring the interaction with risk), but on average, appear to magnify the impact of risk shocks on portfolio bond and equity flows,” though “the interaction effects for extreme risk shocks … are less often significant.” Adding lagged capital controls as an extra control to the baseline leaves “the central results … little [different] from the baseline,” with capital controls themselves insignificant. The authors flag three limitations of the capital-controls data that weaken this comparison: it “does not capture the intensity of the capital controls,” it ends in 2017, and it is annual.

The COVID window, the pre- and post-crisis split, the safe-haven exclusion, and the risk measure (Section 5.5, pp. 109-114). Dropping everything from 15 February 2020 – or just March 2020, “the month of the sharpest risk-off move” – leaves the key results robust, so “the sharp movements during the pandemic do not drive the key estimates.” Splitting at the global financial crisis gives interaction terms that are “insignificant for the earlier window (but the key results are unchanged for the post-GFC window),” and the authors decline to interpret that: “We are cautious in drawing any firm conclusions from the comparison with the post-GFC window, however, as there is little variation in the macroprudential stance in the earlier window, plus this earlier window is significantly shorter. It only includes an extended ‘risk-on’ period plus the sharp ‘risk-off’ episode around the global financial crisis.” Adding back the United States, Japan and Switzerland raises the sample by 5 to 7 percent and leaves signs and significance “largely unchanged.” Switching from RORO to the VIX is the one substantive failure, killing the equity interactions (see Q6).

The signs carry over but significance mostly does not – and bank flows go the other way (Section 6, pp. 114-125). Redoing the analysis on quarterly IMF-based cross-border flows from Forbes and Warnock (2021) – 59 countries, 1980Q1-2020Q3, covering portfolio debt, portfolio equity, FDI and bank flows – the authors are candid that this test has little power: “Regressions predicting quarterly movements in capital flows often have low explanatory power, and coefficient estimates are often insignificant,” particularly post-2008, when “the relationship between global risk measures and capital flows (including extreme capital flow episodes) appears to have broken down.” For portfolio debt and equity the patterns “generally agree with the results from the analysis using the higher-frequency portfolio data. … None of these effects are significant, however.” The result the authors single out anyway is the sign flip: “A tighter macroprudential stance appears to amplify movements in international portfolio flows, especially at the extremes and for risk-off shocks (as found above for portfolio investment flows). For bank flows, however, a tighter stance appears to work in the opposite direction, especially at the extremes of the distribution. This result that a tighter macroprudential stance dampens (instead of amplifies) the impact of risk shocks on bank flows is not surprising as most macroprudential regulations apply to banks.” The reading is a two-sided one: “even if macroprudential regulations improve the resilience of bank flows to the global financial cycle, they may simultaneously shift risks to bond and equity markets and increase the sensitivity of these flows.” Note this is the paper’s weakest evidence and the piece that most directly closes the risk-shifting loop – the two halves of the story rest on data of different quality.

That tool choice and the distribution of outcomes belong in any cost-benefit assessment – not that macroprudential policy should be loosened (Section 1, p. 2; Section 7, pp. 125-127). The refusal is stated twice. Early: “The findings do not imply that macroprudential regulations should be diluted or rolled back, as they may still provide significant benefits by improving the resilience of the domestic banking system. Instead, it is vital to consider the precise macroprudential tools, spillovers, leakages, and corresponding vulnerabilities when designing an optimal policy package.” And in the conclusion: “It is important to highlight that we do not suggest that macroprudential policies render the broader economy less resilient or more sensitive to risk shocks – as the increased resilience of banks may outweigh the greater sensitivity of non-bank financial intermediation. Our results do, however, suggest that the broader spillovers and interaction effects deserve attention in any discussion of the costs, benefits, and effectiveness of macroprudential regulation.” The methodological message is equally explicit: research should “carefully examin[e] the impact of different tools (and not just the overall regulatory stance) and incorporat[e] the intensity of various policies (and not just focusing on recent changes or measures based on dummy variables).”