Global Factors in Noncore Bank Funding and Exchange Rate Flexibility
What this paper finds — and why it matters
Layer 1: Overview
Research question and motivation: The paper asks how far global factors drive the foreign-borrowing component of advanced-economy banks’ non-core funding, and whether exchange rate flexibility (and macroprudential policy) can insulate national banking systems from those global factors. This speaks to the long-running “trilemma vs. dilemma” debate (Rey 2015 vs. Mundell 1963; Miranda-Agrippino and Rey 2020) over whether a flexible exchange rate buys monetary/financial autonomy under open capital accounts. Non-core funding (funding other than deposits — repos, debt securities, foreign borrowing) matters because, per Shin and Shin (2011), Hahm et al. (2013) and Jorda et al. (2017), it is an elastic, crisis-predictive funding source closely tied to credit booms and leverage.
Data and method: A balanced quarterly panel of 31 advanced (high-income) economies, 2004:Q1-2022:Q1, >2,000 country-quarter observations (most specifications drop Iceland as an outlier, leaving 30 countries, 72 periods, 2,160 obs). The non-core ratio is foreign liabilities (IFS line 26c) over deposits (lines 24+25); mean 78%, SD ~94%. The loan-to-deposit ratio (mean 122%, SD ~58%) is a robustness outcome; the two are correlated at ρ=0.92. Sample is ~53% fixed exchange rate (Ilzetzki et al. 2019 coarse classification, monetary union counts as fixed); average Chinn-Ito index 0.95, so capital accounts are essentially fully open. Identification combines the Pesaran (2006) Common Correlated Effects (CCE) estimator with the Mean Group (MG) estimator in a three-step procedure: (1) CCE-MG with observed global factors plus cross-section averages to absorb unobserved factors; (2) extract principal components (number set by Ahn-Horenstein 2013 criterion) from the composite residual; (3) re-estimate with PCs, allowing PC loadings to differ by exchange rate regime.
Main findings with magnitudes: (1) The non-core ratio is highly persistent (lagged dependent variable significant at 1% throughout; coefficient 0.659 in the baseline MG-PC specification) and overwhelmingly driven by global factors; the number of common factors in the non-core ratio is estimated at 3, and the three PCs explain ~80% of the explained variance (PC1 0.795, PC2 0.585, PC3 0.138 — note these sum to >1 and are reported as the lower panel of Table 3). (2) Standard two-way fixed effects leave strong residual cross-sectional dependence (CD test rejects), so are likely biased; the CCE step drives the residual CD statistic to a non-rejection 0.797 (p=0.425) with zero residual factors. (3) Central result: global factors raise non-core ratios more for fixers than floaters — the PC1 loading is 0.984 for fixers vs. 0.302 for floaters; PC2 is significant for fixers, PC3 for floaters; a test on the summed PC loadings (statistic 7.12) confirms larger loadings for fixers. So flexible exchange rates partially insulate. (4) Insulation is stronger away from crises: in the no-crisis 2010-2019 sample the fixer-floater gap in PC1 widens and PC3 (a crisis factor) turns insignificant. (5) Among domestic variables, only the lagged dependent variable, a more appreciated real exchange rate, and higher money/GDP significantly raise non-core ratios; country-specific factors play a minor role overall.
Mechanisms and implications: Relating PCs to observables, PC1 loads most on world macroprudential stringency (tighter regulation lowers non-core ratios), PC2 on the US shadow rate (positive in-sample, reflecting QE/QT dynamics), PC3 on financial-crisis dummies. VIX, oil prices and the US real exchange rate carry expected signs but smaller effects. Using BIS Locational Banking Statistics (23 of 30 countries), the global-factor effect works mainly through interbank borrowing (cross-border liabilities to banks), a flighty source; currency denomination matters little. Tighter macroprudential policy provides complementary insulation, especially for fixers against PC2 and PC3 (which together explain ~21% of non-core variation): for fixers the PC2/PC3 loadings of ~1.47/1.55 under loose regulation fall to essentially zero under tight regulation; for floaters macroprudential tightness adds no insulation. Policy upshot: the Mundellian trilemma is broadly supported for bank funding — flexible exchange rates and tighter macroprudential rules each dampen transmission of the global financial cycle to bank balance sheets, though not against crisis shocks.
Layer 2: Deep Dive
What is the identification strategy and what are the main threats to it?
The authors estimate a dynamic interactive-fixed-effects panel where the non-core ratio depends on its lag, country-specific variables, observed global factors, and unobserved common factors with country-specific (heterogeneous) loadings. Identification proceeds in three steps: (1) a CCE-MG regression (Pesaran 2006; Chudik-Pesaran) that includes observed global factors directly and approximates unobserved factors via cross-section averages of the dependent and independent variables, identifying the country-specific slopes off the variation in regressors orthogonal to common factors; (2) extraction of principal components from the composite residual u-hat that encapsulates the entire factor structure (number of PCs = 3, the estimated number of common factors in the non-core ratio); (3) re-estimation with the PCs, with loadings split by exchange rate regime. The main threat is that omitted/unobserved common factors correlated with the regressors cause strong cross-sectional dependence and biased, inconsistent estimates — exactly what they show afflicts two-way fixed effects (CD test rejects weak dependence; 2 residual factors remain). They verify the CCE step removes this: residual CD statistic 0.797 (p=0.425) and zero estimated residual factors, so the composite captures the full factor structure. They use one-quarter lags of all observables to limit endogeneity, and the rank condition is met with six cross-section averages exceeding the number of factors.
What are the main mechanisms and how are they distinguished empirically?
After establishing the PCs statistically, the authors give them economic content by regressing each standardized PC on observed global factors (Table 6). PC1 loads most strongly on world macroprudential stringency (coefficient -2.957 on the non-core ratio direction, i.e., tighter global regulation lowers non-core ratios), R2=0.971. PC2 is driven by the US shadow rate (coefficient 1.171, positive), R2=0.921. PC3 is driven by financial-crisis dummies — adding a US banking crisis dummy (2007:Q4-2011:Q4) raises the PC3 regression R2 and the crisis dummy (coefficient 2.050) dominates the macroprudential variable. The positive PC2-US-rate relation seems to contradict the GFC literature (lower US rates usually raise cross-border flows), but they explain it via QE: lower shadow rates from bond purchases flatten the yield curve and push banks to fund via long-term bond issuance rather than short-term interbank borrowing; since their non-core measure is dominated by interbank borrowing, lower shadow rates reduce it. They show the sign flips to the conventional negative when using the loan-to-deposit ratio (Appendix Table 11) or a pre-2007 (pre-QE) sample (correlation -15.7%).
What heterogeneity is documented?
Two main dimensions. (1) Exchange rate regime: PC loadings are larger for fixers than floaters — PC1 loading 0.984 (fixers) vs. 0.302 (floaters); PC2 significant for fixers, PC3 for floaters; the summed-loading difference test statistic is 7.12 (p in the test reported as 0.011 for PCF1>PCF0). (2) Macroprudential stance: countries that tightened macroprudential policy more than the median country are less affected by PC2 and PC3. The insulation from tight macroprudential policy is concentrated in fixers — for fixers the PC2 (PC3) loading of ~1.47 (1.55) under loose regulation falls to essentially zero under tight regulation; for floaters, macroprudential tightness gives no additional insulation. Beyond this, country-specific slopes are confirmed necessary by slope-heterogeneity tests (the delta tests reject homogeneity).
What robustness checks are run?
Five (Table 4): (1) dropping the United States (since observed global factors are US-dominated) — results hold, PC1+PC3 affect floaters, PC1+PC2 affect fixers. (2) Including Iceland — results similar but less precise and some residual cross-sectional dependence reappears. (3) Dropping COVID (sample ends 2019:Q4) — virtually unchanged, slightly lower significance. (4) A pure no-crisis sample 2010:Q1-2019:Q4 — PC1 and PC2 still larger for fixers, the fixer-floater PC1 gap widens (insulation stronger outside crises), and PC3 turns insignificant for both groups (consistent with PC3 being a crisis factor). (5) Loan-to-deposit ratio as alternative outcome — PC1 and PC2 significant for floaters, PC1 only for fixers; the apparent lack of flexible-rate insulation to PC1 here is driven by the crisis episodes, and disappears when GFC/COVID are dropped. The three-step CCE diagnostics (first-stage CD non-rejection, zero residual factors) hold across columns.
How does this paper relate to and differ from closely related prior work?
It extends the global-financial-cycle literature (Rey 2015; Miranda-Agrippino and Rey 2020; Bruno and Shin 2015; Obstfeld et al. 2019) and the non-core-funding literature (Shin and Shin 2011; Hahm et al. 2013) by focusing specifically on the non-core-to-core funding ratio of advanced-economy banking systems rather than capital flows or interest rates. Relative to Amiti et al. (2017) — who find global factors explain cross-border flows mainly in expansions — and Cerutti et al. (2019) — who find the global component explains less than a quarter of capital-flow variation — this paper finds global factors overwhelmingly dominate the non-core ratio. Methodologically it differs by combining Pesaran’s CCE estimator with PC extraction and MG estimation to identify and economically label the global factors, rather than relying on two-way fixed effects, which it shows are biased here by uneliminated cross-sectional dependence. It sides with the trilemma camp (exchange rate flexibility insulates, at least partially) against the strong ‘dilemma’ view.
What are the policy implications and their scope conditions?
Flexible exchange rates partially insulate bank non-core funding from the global financial cycle, and tighter macroprudential regulation provides complementary insulation — supporting the Mundellian trilemma for bank balance sheets. Scope conditions: (1) insulation works against regulatory/financial/real drivers (PC1, PC2) but NOT against financial-crisis shocks (PC3), which hit fixers and floaters similarly; (2) insulation is stronger away from global crises; (3) macroprudential insulation operates mainly for fixed-rate countries; (4) the global financial cycle cannot be summarized by a single observable (VIX or otherwise) — it is best captured by composite principal components, so policymakers should monitor a bundle of real, monetary and financial indicators. The authors explicitly caution the currency-denomination-doesn’t-matter result and the broader findings are advanced-economy-specific and may not extend to emerging markets with larger currency mismatches and more volatile exchange rates.
Through which liability channel does the global-factor effect operate?
Using BIS Locational Banking Statistics (23 of 30 countries) in fixed-effects regressions of cross-border liability components on the three PCs (Table 7), all three PCs are positively correlated with total cross-border liabilities. The effect materializes through both domestic- and foreign-currency liabilities (currency denomination matters little — sample correlations 80% foreign-currency, 82% domestic-currency) and, crucially, through cross-border liabilities vis-a-vis other banks (interbank borrowing, correlation 89% with the non-core ratio). Liabilities to nonbank financials (correlation 80%) and other sectors (correlation 18%) are hardly, or even negatively, related to the PCs. Interbank funding is emphasized as a particularly flighty source.
Why use the CCE/MG estimator instead of two-way fixed effects, and what is the cost?
Two-way fixed effects assume additive country and time effects and cannot absorb unobserved common factors that load heterogeneously across countries or are correlated with regressors; in this data they leave strong residual cross-sectional dependence (CD test rejects; two residual factors), implying biased and inconsistent slopes. The CCE estimator approximates unobserved factors by cross-section averages without needing to know the exact number of factors, and the MG estimator allows country-specific slopes (confirmed necessary by slope-heterogeneity tests). The pooled CCE estimator failed to remove residual cross-country correlation in every specification and was inferior to MG. A cost is that the PCs span observed and unobserved factors and lack a clean one-to-one economic meaning, which the authors address by separately regressing PCs on observables (Section 5.1).
What does the descriptive evidence show before the regressions?
The non-core ratio and loan-to-deposit ratio co-move strongly (ρ=0.92). The non-core ratio is generally higher for fixed-rate countries, shows long-term trend shifts and co-movement across regime groups, rose before the GFC to a global peak of 70% in 2008, then fell to about 30% by 2022, with short-term fixer-floater divergence only in 2015-2020. The benchmark non-core ratio correlates 88% with the overall BIS cross-border liability variable.