Uncertainty Shocks and the Cross-Border Funding of Banks: Unmasking Heterogeneity
What this paper finds — and why it matters
Layer 1: Overview
Research question and motivation: How does country-specific uncertainty explain variation in the cross-border funding of banks? Studying this link is practically relevant given rising reliance on international borrowing under financial globalization and the role of international banking in transmitting the Global Financial Crisis (GFC). The few prior studies on uncertainty and cross-border bank funding (Cerutti et al. 2017; Choi and Furceri 2019) focus on a single uncertainty measure and aggregate flows. Bénétrix and Curran’s innovation is to decompose both the funding source (banks vs. non-banks) and the type of uncertainty measure, “unmasking” heterogeneity that aggregate panel studies hide.
Data and setup: International bank funding is measured as cross-border liabilities (loans plus debt securities) of banking systems reporting to the BIS Locational Banking Statistics (LBS), decomposed into liabilities vis-a-vis banks and non-banks (non-bank flows derived as the difference between all-sector and bank liabilities). The core sample is 24 reporter countries (excluding small states/financial centers driven by global shocks, e.g. Russia/China omitted for short coverage), quarterly 2003Q1–2018Q4. The crisis period is defined as 2008Q3–2012Q2 (start = TED spread record/Lehman; end = Draghi’s “whatever it takes”), with pre-crisis 2003Q1–2008Q2 and post-crisis 2012Q3–2018Q4 sub-samples. A newly compiled uncertainty dataset spans three classes: volatility-based (implied volatility at 1-month and 3-month maturities from Bloomberg OVM; realized volatility from national equity indices), news-based (EPU and the World Uncertainty Index WUI from policyuncertainty.com), and forecast-based (forecast dispersion = standard deviation of GDP-growth forecasts across forecasters, from Bloomberg ECFC). Coverage: 24/24 countries for realized vol, implied vol, and WUI; 16/24 for EPU; 15/24 for forecast dispersion.
Empirical strategy: Two parts. First, descriptive dynamics of banking and uncertainty series (moments, persistence via AR(1)). Second, dynamic panel regressions with country fixed effects and Pesaran-Smith mean-group (MG) estimators, plus country-by-country regressions, of log cross-border liabilities on log uncertainty and a lagged dependent variable (so beta is an elasticity); standard errors clustered by source country. Multivariate models add lagged conditioning factors (real GDP growth, stock-market growth, policy rates, credit growth, exchange-rate growth, inflation, external debt/GDP). A GFC dummy and uncertainty-GFC interaction capture the time dimension.
Main findings with magnitudes: Uncertainty is associated with less cross-border borrowing; effects are sizable but heterogeneous. A 1% rise in 3-month implied volatility can contract funding by up to 4.1%; across implied/realized volatility (same sample) elasticities run 1.5%–4.1% depending on measure, sector, and estimator. Volatility-based measures show the largest elasticities, then news-based. Contractions are largest for non-bank funding and smallest for aggregate (suggesting bank/non-bank substitution that mutes the aggregate). Economically, a one-standard-deviation uncertainty shock typically cuts aggregate funding by between $573 billion and $889 billion (the bounds correspond to 1-month vs. 3-month implied volatility; average aggregate funding is $820B, average non-bank funding $223B). Country regressions give similar but more often insignificant results. Over time: volatility-based uncertainty matters only during the GFC (interaction term strongly negative), while news-based uncertainty (EPU, WUI) is the only measure whose first two moments rose since the GFC and is the only one that dampens funding outside the crisis, particularly for European countries (EU15/euro area). Mechanisms discussed but not tested: deleveraging/precautionary saving, liquidity management, demand vs. supply channels (weaker supply channel for advanced “safe” countries).
Layer 2: Deep Dive
What is the identification strategy and what are the main threats to it?
The paper is explicitly descriptive/documentary, not structural (‘The goal of this paper is to document empirical evidence, not to model mechanisms’). Identification comes from dynamic panel fixed-effects and mean-group regressions of log cross-border liabilities on log uncertainty with a lagged dependent variable, plus country-by-country regressions. The main threat is reverse causality (uncertainty and bank flows co-determined). The authors mitigate this following Bruno and Shin (2015b) by re-estimating with uncertainty lagged one period (similar results, in the online appendix) and by lagging conditioning factors one quarter. They argue the lagged dependent variable absorbs much variation, leaving less for uncertainty and ameliorating omitted-variable bias, but they do not claim causal estimates. They do not use instruments; the multilateral (vs-the-rest-of-the-world) data is used to avoid purely idiosyncratic counterparty shocks.
What heterogeneity is documented?
Four dimensions. (1) Funding sector: non-bank funding grows faster and is more volatile than bank funding, which is more volatile than aggregate; non-bank funding grew faster than bank funding in 75% of countries over the full period (54% pre-crisis, 75% during, 75% post-crisis). Uncertainty contractions are largest for non-banks, smallest for aggregate. (2) Uncertainty measure: volatility-based show the largest elasticities, then news-based; forecast dispersion is weakest/often insignificant. (3) Country: riskier countries (emerging markets like Brazil/Turkey; peripheral euro members Italy/Portugal/Spain) show significance for bank flows, while safe havens (Germany, USA) show significance for non-bank flows; some countries (Singapore, Norway, Switzerland) are largely unaffected; Finland and Japan show positive (wrong-signed) responses. (4) Time: volatility-based uncertainty matters only during the GFC; news-based matters outside it, especially for Europe.
What are the candidate mechanisms and are they tested?
Mechanisms are discussed but explicitly left for future research. Deleveraging/precautionary saving: under higher uncertainty banks shrink balance sheets and borrow less abroad. Liquidity management: uncertainty creates liquidity concerns, so banks may borrow more or less depending on term horizons. Rebalancing: volatility-based uncertainty (tracking equity risk) may drive borrowing from a risk-management/rebalancing perspective, while news-based uncertainty may operate through liquidity. Demand vs supply: higher uncertainty can cut a country’s banks’ demand for funds or foreign supply of funds; advanced/safe-haven countries are argued to face a weaker supply channel because the rest of the world keeps trusting them, consistent with safe havens reducing non-bank funding demand while aggregate is little changed (a shift between bank and non-bank funding).
Why does volatility-based uncertainty produce the strongest results even though it is narrower than news-based?
A priori the broader news-based measures might be expected to matter more, but the authors find volatility-based the strongest. They reason that cross-border banking decisions place greater weight on financial-system conditions, which volatility-based uncertainty (tracking the stock market) captures directly; banks holding securities may need to rebalance, diversify, or recapitalize via international borrowing/lending in response to equity risk.
What robustness checks are run?
(1) Bivariate vs multivariate: adding conditioning factors (GDP, stock market, inflation, policy rate, exchange rate, credit, external debt) leaves the negative uncertainty relation; multivariate panel elasticities narrow to roughly -2.2% to +0.5% vs bivariate -4.1% to +0.3%, MG largely unchanged. (2) Balanced 13-country fixed sample (panels C/D of Table 1) to compare measures on identical samples; similar negative, heterogeneous results. (3) One-period lag of uncertainty to address reverse causality (similar). (4) Crisis dummy plus interaction and separate pre/post-crisis estimation. (5) Alternative forecast-based measures (forecast-error dispersion, mean absolute forecast error) gave similar results. (6) An earlier version purged realized/implied volatility of the VIX to get idiosyncratic volatility (similar). (7) Persistence robust to including a constant; AR(1)/half-life analysis.
How does this paper relate to and differ from Choi and Furceri (2019)?
It is closest in spirit to Choi and Furceri (2019), who find a negative relation between banking flows and uncertainty using realized volatility and EPU on bilateral, aggregate flows (assets and liabilities). Bénétrix and Curran instead decompose flows into bank vs non-bank sub-components and use a broad set of uncertainty measures (implied volatility at two maturities, realized volatility, EPU, WUI, forecast dispersion), arguing this avoids the limitations of relying only on backward-looking realized volatility or cross-country-incomparable EPU. The nuanced result that news-based uncertainty matters outside the GFC (because only it rose since the crisis) departs from existing panel studies like Choi and Furceri. From Cerutti et al. (2017) they take the relevant takeaway that cross-border flows decline when the US VIX rises.
What are the dynamic/descriptive findings on the data?
Cross-border funding grew over two decades, especially pre-GFC; non-bank funding dominates growth during/after the crisis and is the most volatile, aggregate the least (e.g., Singapore and Finland std devs of 4.1 and 21). Cross-country average growth of non-bank liabilities is 2.2% vs 1.3% for bank liabilities. 64% of countries show positive autocorrelation in aggregate liabilities for the full period, while ~60% show negative autocorrelation for the two sub-components; pre-crisis ~80% show negative aggregate autocorrelation. Means/medians of flows are u-shaped (positive-negative-positive across pre/during/post), std devs n-shaped. For uncertainty, volatility-based moments peak during the crisis; only news-based (EPU, WUI) rose during and since the crisis. Uncertainty shocks are short-lived (half-lives about one quarter); ordering from least to most persistent: forecast-based, WUI, EPU, 1-month implied vol, realized vol, 3-month implied vol.
What are notable country-specific results?
3-month implied volatility elasticities range -14.1% to 11.5% (non-negative ones all insignificant); 1-month range -11.4 to 10.3; realized volatility -18.7 to 14 (with some significant positive estimates: Japan +4.7 overall, Finland +13.6 and +13.9 for overall/bank). EPU ranges -11.2 to 20.9 (positive significant for Japan in aggregate/bank, Brazil non-banks); WUI tighter, -4 to 2.7 (max contraction 4% for Austria bank funding; India positive). Forecast dispersion -30.7 to 4.4 (or -8.2 to 4.4 excluding Brazil); significant negative for UK (all sectors) and Brazil/Italy/UK (non-banks). France, Portugal, Ireland show robust negative responses; Portugal is significantly negative for all measures and sectors.
What are the policy implications and their scope conditions?
Policymakers should note that uncertainty mattered most during the GFC and European Sovereign Debt Crisis, and that news-based uncertainty has a distinct, sizable dampening effect on cross-border flows since the Great Recession, particularly for European nations (EU15/euro area), because only news-based uncertainty rose post-crisis. A single uncertainty measure does not fit all, since banking systems differ in structure, ownership, cross-border activity, size, and local-economy exposure. Scope conditions: results are associations not causal effects; effects are concentrated in the crisis window for volatility measures; non-European and emerging markets show no significant news-based effect outside the crisis; the sample is 24 countries, 2003Q1–2018Q4, multilateral liabilities only.
What are the main caveats and limitations the authors acknowledge?
Data limitations prevent regression analysis on intragroup, financial, and non-financial flow sub-components (explored only preliminarily). Non-bank liabilities are derived as a residual (all sectors minus non-banks) because bank-counterparty data are partly missing, though the authors argue the impact is minimal. Uncertainty coverage is unbalanced across measures (EPU 16, forecast dispersion 15 of 24 countries). Implied volatility (OVM) and forecast (ECFC) series could not be automated and required manual snapshots. The AR(1) persistence choice may miss nonlinearities/structural breaks and gives an upper bound on persistence. Country-level coefficients are often statistically insignificant given the strong lagged dependent variable. Mechanisms/channels are not tested and left for future work.