International Spillovers and Local Credit Cycles

The gap is that the literature has studied the aggregate sources and effects of the Global Financial Cycle but not how it reaches an economy’s banks and firms; the contributions are micro estimates of the cycle’s effect on borrowing costs and credit, and causal evidence on the channels. “While the majority of research has focused on understanding the aggregate impacts and sources of the GFC, little is known on how it transmits to countries’ economies and financial sectors” (Section 1). The first contribution is framed against the limitation of aggregate work: because the micro regressions “control for time-varying bank characteristics, the estimates are more reliable than similar macro estimates using country-level data, which are unable to separate the banking sectors’ financial conditions from macroeconomic business cycle conditions.” The second is the channel evidence: “the causal identification of such a supply-side channel of the international spillovers of global financial conditions is not possible using macro data, as the econometrician will not be able to control for demand-driven borrowing using only aggregate data.” The four motivating aggregate correlations with log(VIX) are stated up front: 0.52 with the aggregate bank lending rate, minus 0.51 with the median bank’s non-core liabilities ratio, 0.61 with Turkey’s UIP deviation, and 0.01 with the aggregate collateral-to-loans ratio – the last foreshadowing the paper’s negative result.

Transaction-level reports from every bank, with a richer variable list than most registers – and two specifically Turkish complications that must be handled. Banks report each outstanding loan separately as long as its amount exceeds a bank-specific reporting cutoff, with smaller loans aggregated at branch level; the data give “the (i) interest rate; (ii) maturity date as well as extended maturity dates if relevant; (iii) collateral provided; (iv) credit limit (only beginning in 2007); (v) currency of the loan; (vi) detailed industry codes… (vii) bank-determined risk measures of the loans” (Section 3.1). The paper uses cash loans in outstanding principal rather than credit limits, since limit data are unavailable for the full period. The important adjustment concerns foreign currency loans, which come in two reporting types: those indexed to the exchange rate, reported at initial lira value, and those issued in foreign currency, whose lira value is revalued each period. The latter “naturally creates a valuation effect, which we need to correct for,” because otherwise “a depreciation of the TRY against the USD would appear to increase total loans outstanding for all existing FX loans issued in dollars… but this expansion would solely have been due to a currency depreciation, rather than issues of new loans”; the correction uses official end-of-period exchange rates. Data are then deflated to 2003 prices and winsorised at 1 percent. The aggregated quarterly sample tracks the economy: “the two series track each other very closely, with a correlation of 0.86” between the paper’s aggregated loan growth and economy-wide credit growth.

Because banks are the dominant intermediary of capital inflows and, unlike Mexico, the banking system is dominated by domestic rather than foreign banks – which is what lets the paper isolate a domestic-bank transmission channel. On bank dominance: “in 2014, banks held 86% of the country’s financial assets and roughly 90% of total financial liabilities. The past decade has witnessed a doubling of bank deposits and assets, while loans have increased five-fold” (Section 3.2), with the sector’s assets averaging about 78 percent of GDP, loans 38 percent and deposits 46 percent over the sample. On firms’ inability to bypass banks: “Firms’ direct external borrowing is very limited in Turkey and hence banks are the key intermediary of capital flows… the external corporate bond issuance is negligible as a percentage of GDP, whereas banks’ external borrowing is as high as 40 percent of GDP” (Section 3.3). On the contrast: “domestic banks dominate overall banking activity relative to foreign banks. This is an important fact to note and makes Turkey different than other large emerging markets (e.g., Mexico), where foreign banks play a larger role in the banking sector.” The sample covers 45 banks over 2003-13 – 28 commercial, 14 investment and development, 5 branches of foreign banks – with 35 to 43 active in the corporate loan market in any period, and begins after the post-2001-crisis wave in which 25 banks were taken over by the deposit insurance fund.

Wholesale rather than deposit funding, largely foreign-currency, and – crucially – co-moving with capital inflows for domestic banks but not for foreign ones. The construction is explicit: “Non-core liabilities equals Payables to money market + Payables to securities + Payables to banks + Funds from Repo + Securities issued (net),” with total liabilities defined as deposits plus non-core funding, and banks split by a time-invariant dummy comparing a bank’s average ratio to the sample median (Section 3.2.1). The justification for the measure is the time-series relation rather than the cross-section: “What is important for our purposes is the time-series relation between non-core liabilities and foreign financing (capital flows)… domestic banks’ non-core liabilities move together with capital flows (both total and banking sector), whereas foreign banks’ do not.” Two features of the split matter for interpretation. The groups are not a proxy for nationality: “the low non-core sample [has] 63% of the banks being domestic, while the high non-core sample is composed of 64% of domestic banks” (fn. 17). And the groups differ in ways that motivate later robustness checks: “the low non-core banks are larger on average, as well as having larger leverage ratios than the high non-core bank sample,” while high non-core banks have somewhat larger liquidity and capital ratios and similar returns on assets.

An interest-rate elasticity of 0.019 and a loan elasticity of minus 0.067, the latter accounting for 43 percent of cyclical aggregate corporate credit growth. The specification regresses the log of one plus the weighted-average nominal rate, or log loans, on lagged log(VIX) with firm-by-bank fixed effects, a linear trend, an FX dummy, the lagged domestic overnight rate, lagged GDP growth, inflation and exchange rate change, and lagged bank characteristics – log assets, capital ratio, liquidity ratio, non-core ratio and return on assets – estimated by weighted least squares with weights equal to the time-series mean of log bank assets and standard errors double-clustered on firm and quarter (Section 4.1). The interest-rate coefficient of 0.019 “implies a 1 percentage point fall in the average borrowing rate resulting from a fall in log(VIX) equal to its interquartile range over the sample period.” The aggregation step is spelled out in an appendix: applying the estimated coefficient to observed changes in log(VIX) gives a predicted aggregate loan growth series, which is divided by the linearly detrended series of actual aggregate credit growth and averaged, yielding 43 percent (fn. 22). The paper notes it deliberately includes the domestic policy rate directly rather than relying on UIP, “as we later document a UIP violation,” and that the bank controls include the capital ratio – the inverse of leverage – precisely because that variable “has been highlighted as responding to global financial conditions and wealth effects arising from exchange rate and asset price changes.”

From the sign of the interest rate coefficient: a demand story predicts the opposite sign, a supply story predicts the sign observed. The argument is set out carefully. “If movements in VIX were in fact picking up local demand factors, we would expect loan interest rates to be negatively correlated with VIX. For example, imagine that VIX falls as global conditions improve and Turkish firms react by beginning to demand more credit… we would then expect there to be an upward pressure on lending rates given increased demand, holding banks’ supply constant. This would generate a negative correlation between rates and VIX, which is counterfactual to our regression results. If, on the other hand, a fall in VIX is associated with an improvement in global financial conditions and allows Turkish banks to access foreign capital more cheaply, these banks can then offer loans at a lower interest rate in the domestic credit market. This in turn implies a positive correlation between VIX and lending rates as we find” (Section 4.1). The key logical point is that the volume coefficient alone cannot discriminate: “under both a demand or supply scenario we would expect the same signed coefficient between VIX and loan volume.” The authors nonetheless treat this as suggestive rather than conclusive – “the estimation of (1) may still suffer from omitted demand variable, especially at the firm-level and thus be biased” – which is why the difference-in-differences design follows. Two supporting checks are reported: regressing Turkish capital inflows on log(VIX), macro controls and consumer confidence over 39 observations gives a log(VIX) coefficient of minus 1.493 significant at 1 percent while consumer confidence, the natural demand variable, is positive but insignificant (fn. 24); and splitting exporters from non-exporters yields a smaller interest-rate elasticity and larger volume elasticity for exporters, consistent with some demand content in VIX, though “the coefficients remain strongly significant and are the same sign” and exporters are “less than ten percent of our sample of loans” (fn. 25).

That domestic banks, not foreign ones, carry the cycle into Turkish credit – which the paper flags as a direct contrast with the Mexican evidence. “While the interest rate coefficients in panel A are positive and highly significant for both the domestic and foreign bank sub-samples, the coefficient estimated in the domestic bank sample is almost twice as large as that of the foreign bank sample” – the working paper prints these as 0.22 against 0.12, figures whose scale sits oddly with the 0.019 all-bank estimate in the same panel and which are most plausibly read as 0.022 and 0.012 (Section 4.1). The volume result is starker: “the coefficient on log(VIX) is negative and strongly significant for domestic banks, while it’s actually slightly positive and statistically insignificant for foreign banks. These results point to the crucial role that domestic banks playing in transmitting the GFC to the Turkish credit market, a result that differs drastically from the literature that focuses on the role of foreign banks in transmitting foreign monetary policy (e.g., Morais et al.).” The paper positions its own mechanism one layer below the global-bank literature: “We drill down one more layer and argue that domestic banks, which obtain their dollar funding from global banks, in turn, provide more funding to domestic firms at a lower cost; a mechanism of pass-through from a lower cost of international funding for domestic banks to a lower cost of borrowing for domestic firms” (Section 2).

Large enough that high-non-core banks account for most of the aggregate effect, and a selection check suggests the bias runs against the finding. With firm-by-quarter fixed effects absorbing demand, the non-core-by-VIX interaction is 0.013, “which is almost as large as the estimated elasticity of 0.019 between the interest rate and VIX in the macro regression,” so “the relative differential in changes in interest rates for high non-core banks given movements in the GFC is economically large, and high non-core banks are responsible for a significant part of the aggregate effect” (Section 4.2). Dropping the firm-by-quarter effects to recover a level term gives 0.015 on VIX and 0.015 on the interaction, so “the estimated interest rate-VIX elasticity for high non-core banks is double (0.015 + 0.015 = 0.03) that of low non-core banks (0.015).” Two robustness results support the reading. Splitting banks by non-core quartiles rather than a median dummy, “the Q3 and Q4 interactions are significant, with the Q4 interaction coefficient being significantly larger (in absolute value) for both the interest rate and loan regressions” (fn. 27). And the selection check on changing firm-bank relationships runs the reassuring way: firms on average hold more relationships with low-non-core banks, and the slightly negative correlation with log(VIX) implies that “during boom times (i.e., low VIX), firms tend to form more relationships with high non-core banks. If anything, this could bias our results towards zero, as more firms borrowing from high non-core banks would also imply a greater chance of low-quality firms asking for loans, and thus high non-core banks would tend to charge higher, not lower interest rates during boom periods.”

A 7 percentage point average price advantage for foreign currency loans that narrows in booms – but it vanishes once the comparison is confined to the same firm borrowing from the same bank in two currencies in the same quarter, which the paper calls the ideal test. “The average price differential between FX and local currency loans remains at 7 percentage point when evaluated at the sample mean of log(VIX). More interestingly, during high-VIX episodes, this differential gets larger, where FX loans are 8 percentage points cheaper during high VIX episodes, whereas they are only 6 percentage points cheaper during low VIX episodes based on the interquartile range of log(VIX). This result implies that local currency borrowing becomes relatively cheaper during low VIX episodes” (Section 4.3). Progressively tightening the fixed effects erodes the average differential: with firm-by-quarter effects its significance drops to 10 percent, and with bank-by-firm-by-quarter effects “the coefficients become insignificant. The cyclical effect captured by the variable FX×log(VIX) becomes borderline significant.” The authors treat this as informative rather than as a failure: “Although this sample is restrictive, it is the ideal sample to test for UIP at the micro level. In theory, it is possible that a given firm can default on its FX loan obligations and not on its Turkish lira loans upon a depreciation of the Turkish lira. In the data, it seems to be the case that a given bank does not price this risk differentially at the loan level for a given firm over time. This result shows the critical role of firm risk and bank heterogeneity that are both controlled in this [specification].” The introduction states the finding in its final form: the differential “comoves with VIX, but the magnitude and significance of this effect are statistically indistinguishable from zero once we control for time-varying firm credit risk and bank heterogeneity in access to foreign currency funding,” and this “is not straightforward to rationalize with existing models that assume an exogenous and static country-level UIP violation.”

Because two bank incentives offset within firm-bank pairs, and because the majority of firms borrow in only one currency and so contribute nothing to the within-pair comparison. The offsetting incentives are stated directly: “during the boom phase of the GFC, Turkish banks are able to more cheaply fund themselves in dollars/euros and therefore would have an incentive to increase lending in FX. At the same time, as we show above, there is a falling risk premium on the Turkish lira, which implies that banks can offer TRY loans with better terms. Thus, banks also have an incentive to increase lending in domestic currency… our results show that these two effects offset each other so that the loan composition does not change at the firm-bank level over the GFC” (Section 4.3). The aggregate pattern is nonetheless real – the detrended FX share of loans co-moves with log(VIX), and the FX share declines in levels over a sample period dominated by the boom phase – and is reconciled by composition: “there are also firms who borrow only in local currency in the aggregate data. These firms make up 63% of the sample. In the difference-in-differences regressions that use the FX dummy to identify the differential pricing and amounts, these firms will not provide any information to identify a differential effect over the cycle since they always borrow only in one currency.” Interacting the FX dummy with the non-core dummy shows where the deviation comes from: “high non-core banks play a key role in the differential pricing of FX and Turkish lira loans. These banks price FX loans higher during low VIX periods and lower during high VIX periods, driving the UIP deviations at the firm and loan levels,” with the differential again disappearing within firm-bank pairs. On volumes the paper corrects an apparent result it judges spurious: the raw finding that non-core banks supply less FX credit in booms “captures the credit demand effect, since during booms when firms’ demand for credit increases, there will be more demand for local currency credit, creating a spurious negative correlation with the supply of FX credit. Once these effects are controlled for, we find that high non-core banks supply more FX credit during normal times (on average) as they fund themselves in FX in the international capital markets.”

Maturity lengthening by high-non-core banks in booms, robustness of the non-core effect to bank leverage and size, and a clean null on the exchange-rate net-worth channel. On maturity, using new issuances so that maturity at origination is observed, “high non-core banks issue less short-term loans on average than low non-core banks… and that this differential becomes larger during boom phases of the GFC,” which the authors read as “high non-core banks take more ‘risk’ by providing more long-term loans during the boom phase of GFC as such loans entail higher default risk” (Section 4.4). Pricing does not discriminate by maturity because “high non-core banks offer lower rates for loans of any maturity.” In a horse race against the leverage and size interactions motivated by Coimbra and Rey (2017), “the non-core interaction remains strongly significant while the leverage ratio interaction is insignificant,” and though larger banks do cut rates more and lend more in low-VIX periods, the non-core coefficients survive throughout; with all three included, the leverage-VIX interaction becomes significant for interest rates but remains insignificant for volumes. The exchange-rate channel of Bruno and Shin is tested with a triple interaction of bank leverage, a firm’s FX loan share, and either the lira-dollar log change or top-quartile depreciation and bottom-quartile appreciation dummies, with both firm-by-quarter and bank-by-quarter effects: “Looking across all specifications we never see a significant coefficient. Therefore, the exchange rate risk-taking channel cannot explain the reduction in borrowing costs and increased lending we have seen in our above regressions.” Two caveats accompany that null: the lender side of the channel is shut off by regulation, since “there is no currency mismatch on Turkish banks’ balance sheets” (fn. 30), and the firm FX share is a proxy built from the currency composition of loans because “the firm-level balance sheet data are not broken down by currency.”

That posted collateral prices loans but does not ration them, and that the pricing relation stops moving with the cycle once the same firm-bank pair is compared – a test made possible by observing collateral directly rather than proxying with net worth. The design point is stated up front: “rather than proxying for financial constraints by a firm’s net worth as is common in the literature, we can use the actual collateral posted for a loan at its issuance,” estimated at the individual new-loan level with firm-by-bank-by-month effects so that “we solely identify from changes in the amount of new loans and their interest rates for a given firm-bank pair. Hence, in our most stringent specification, we do not allow firms to switch banks and vice versa for banks” (Section 4.5). On pricing: “the collateral ratio coefficient is significant and has a negative sign in all columns… although there are theories that predict a negative relation between posted collateral and the loan rate, our paper is the first to provide evidence on this relationship,” and “once we focus on the same firm borrowing from the same bank over time, the relationship between the collateral posted and a loan’s interest rate does not respond to the GFC.” On quantities: “We find no significant relationship between the collateral-to-loan ratio and loan volumes in all columns. Furthermore, we also find no relationship between collateral ratio and credit volumes during high and low VIX episodes.” The horse race settles which margin matters: regressing loan volumes on both the rate and collateral, “the coefficient on the interest rate is negative and strongly significant… Meanwhile, the coefficient on the collateral ratio is never significant once we control for time-varying firm-level fixed effects.” Two measurement choices bound the exercise: the collateral-to-loan ratio is winsorised at 200 percent “to match an upper-bound used by the banks,” since ratios above one arise from liquidation and legal costs, property collateral requirements and collateral posted against whole credit lines; and zero-collateral loans are dropped because they “are generally given to very large multinationals or for loan amounts that are very small compared to firm value,” which “decreases the sample by approximately thirty percent.”

Because a shock to borrowing capacity moves demand and supply together, so without those controls the estimate would not isolate the supply side. The authors cite the theoretical reason directly: “Including these controls is important since as Guerrieri and Lorenzoni (2017) show in a heterogeneous agent environment, shocks to agents’ borrowing capacity affect both borrowers’ demand for loans and lenders’ supply of loans, and we want to solely focus on the supply side” (Section 1). The same care appears in how they interpret the null: finding no relation between posted collateral and loan amounts “does not mean collateral is not an important variable for loans,” since the pricing relation is strong and novel. Their reading of the combined evidence is that the two margins are separable: “firms can borrow at lower rates on average, while their ‘hard’ collateral constraints do not change much over the boom part of the cycle. Thus, collateral-constrained firms are still allowed only to borrow some fraction of their capital stock, and this amount may not change if the value of the capital stock does not change much when capital flows into the banking sector as opposed to the corporate sector” (Section 2) – a channel they note Fostel and Geanakoplos (2015) rationalise theoretically.

For policy, that capital controls on foreign currency borrowing are insufficient; for theory, that a new class of models is needed in which external shocks move risk premia rather than collateral constraints. The policy conclusion follows from the combination of the non-core and UIP results: “it is not enough to limit the foreign currency borrowing for agents in an economy by imposing a capital control. Lower borrowing costs also fuel local currency borrowing if domestic banks can fund themselves cheaply in the international financial markets. Furthermore, lower borrowing costs can drive a credit boom in spite of the collateral constraints staying intact during capital inflow bonanzas” (Section 5). The theoretical claim is put as a specific displacement: “The traditional macro-finance model perspective, in an international setting, is one where external shocks affect the collateral constraint and propagate to the economy. Our evidence shows that external shocks affect risk premia, and which then propagate to the economy by-passing collateral constraints. This finding requires another class of models that highlight the role of risk premia – and how they may vary at the micro level.” The paper also positions the non-core channel as a distinct object for open-economy modelling: it is “akin in spirit to the risk-taking channel highlighted in work such as Bruno and Shin (2015b) or Coimbra and Rey (2017),” but “the non-core channel is a function of the composition of only the liabilities side of the balance sheet, rather than a net position,” with relaxation working “via cost of funding instead of volume of funding related to balance sheet strength.”

By including identified monetary policy shocks alongside it and reporting that the VIX coefficients survive while the shock coefficients survive only for volumes. “We have also run regressions including both the VIX and measures of monetary policy shocks (Gertler and Karadi, 2015)… The regressions show that the coefficients on log(VIX) remain strongly significant. Further, the coefficients on the monetary policy shocks, FF4 or MP1, in the interest rate regressions are either barely significant or not significant at all. Meanwhile, the coefficients for these variables in loan volume regressions are highly significant” (fn. 23). The paper treats VIX as a proxy for the cycle rather than as a structural shock throughout, describing it in the introduction as having “a strong common component that comoves with VIX, which is related to US monetary policy and to changes in risk aversion and uncertainty,” and grounding the choice on the related literature that uses the same proxy. Further robustness reported includes splitting by maturity, restricting to loans earmarked for domestic activity, examining sensitivity across sectors – where “the coefficients on log(VIX) do not vary very much” – and alternative sample cuts, all described as robust (fn. 25 and 28).