Macro and micro of external finance premium and monetary policy transmission

Layer 1: Overview

This paper establishes basic facts about the external finance premium (EFP) faced by euro area firms borrowing from banks, and studies how monetary policy is transmitted to it. The EFP — the extra cost a firm pays for external funds versus the opportunity cost of holding cash — is a central object in financial-accelerator theory (Bernanke-Gertler, Kiyotaki-Moore), but its determinants below the country level have rarely been measured directly. The motivation is that euro area policy discussion treats country-level sovereign spreads as sufficient summary statistics for financial conditions, yet there is little micro evidence on whether country variation actually captures the bulk of loan-level variation.

Data and strategy: The authors use AnaCredit, a loan-level database of all euro area firm loans of at least €25,000, restricted to all new, unsecured loans (so they are not directly affected by Covid government guarantees) in the ten largest euro area economies (Austria, Belgium, Germany, Spain, Finland, France, Ireland, Italy, Netherlands, Portugal), which cover 93% of both the number and value of new euro area loans and 95% of euro area GDP. The sample spans January 2019 to December 2023 and contains about 36 million loans (35,919,600 in the contract tables). Loans are matched to Orbis (firm controls), ECB IBSI and supervisory data (bank balance sheets and capital), CSDB (bank bond yields) and iMIR (aggregate loan rates). The EFP is the loan spread over a maturity-matched OIS rate. They sequentially decompose it via weighted least squares (loan-size weighted) into country-time, then bank-time, then firm-time fixed effects, with contract-level effects as a residual — so each fixed effect is a value-weighted index at that level. Sequence runs aggregate-to-granular so any covariance is attributed to higher aggregation levels, making covariate explanatory power a lower bound.

Decomposition findings: Country-time effects capture 48.5% of the variance; bank-time 23.8%; firm-time 16.3% (bringing country+bank+firm to 88.6%); residual contract-level variation is 11.4%. Banking relationships are highly local — 96% of bank-firm pairs are in the same country (84% value-weighted). At the country level, the relevant covariate is the euro-area average sovereign spread, not the country-specific one: local spreads explain 48% of country-level variation while the EA average explains nearly 80%, and local spreads add no power beyond the EA average — pointing to a common (global) risk factor. The EFP is roughly 2.6 times larger than the sovereign spread. The EFP is countercyclical (higher with lower GDP and higher unemployment). Bank-level: weaker banks (less capitalized, less liquid, more exposed to risky assets, higher funding costs, larger) charge higher EFPs; the 95-5 quantile range of Tier 1 capital implies almost 100 bps higher EFP. Firm-level: smaller, younger, more leveraged, less profitable firms pay more — the 5-95 leverage range implies 90 bps higher EFP, the probability-of-default range about 20 bps, and old (50yr) vs young (5yr) about 30 bps. Crucially, bank-, firm- and contract-level variation remains largely unexplained (R-squared on bank regressions ~0.01-0.05; firm ~0.11-0.18; contract ~0.0001-0.0003).

Monetary policy transmission: Using Jorda local projections on high-frequency identified ECB surprises (Altavilla et al. 2019: Target, Forward Guidance, QE factors from OIS changes around announcements), a null EFP response means exact pass-through. A one-SD Target surprise (8 bps) raises the EFP about 10 bps (peaking 3-5 months); a one-SD QE surprise (€500 bn) lowers the EFP about 20 bps, split roughly equally across bank and firm levels. Effects are asymmetric: policy-rate tightening (not easing) and QE (not QT) are amplified through the EFP. Tightening amplification is mostly at the bank level (bank lending channel, driven by weaker banks); QE additionally narrows the EFP at the firm level (firm balance-sheet channel, helping fragile firms). QT, while fully passed through to tighten lending, leaves the EFP unchanged — attributed to QT’s slower, more predictable, “loud-bang-less” implementation versus QE’s large-envelope announcements (a difference-in-difference on QE envelope months shows a significant EFP decline after envelope announcements). Implication: as the ECB shrinks its balance sheet (lowering liquidity), rate hikes become more likely to generate financial amplification via the EFP, since less-liquid banks respond more to rate hikes. The QT result is caveated by limited sample evidence.

Layer 2: Deep Dive

What is the empirical strategy for decomposing the EFP, and why does the order of fixed-effect extraction matter?

The EFP (loan spread over maturity-matched OIS) is decomposed sequentially via weighted least squares (each observation weighted by loan size) into country-time, then bank-time, then firm-time fixed effects, with contract-level effects as the residual (Equations 1-3). Each fixed effect is effectively a value-weighted index of spreads at that level. The sequence MUST run from aggregate to granular: starting with loan-level effects would soak up all variance. Because aggregate effects are estimated first, any covariance (e.g., a particular firm type clustering at a particular bank, or a country with a strong/weak banking system) is attributed to the higher aggregation level. This means variance attributed to higher levels may be slightly overstated relative to joint estimation, but covariate explanatory power can be read as a lower bound. The authors avoid simultaneous estimation for two reasons: it is computationally infeasible to estimate ~10 million fixed effects jointly and retrieve their values (which are the dependent variables in the second stage), and the sequential method makes clear exactly where covariances land. A check absorbing firm/bank effects via differencing while explicitly estimating country-time effects yields a 98% correlation between sequential and jointly estimated country-time fixed effects.

What is the variance decomposition result, and what is its headline interpretation?

Country-time effects capture 48.5% of loan-level variance, bank-time 23.8%, firm-time 16.3% (country+bank+firm = 88.6%), and residual contract-level variation 11.4%. The headline: country-level variation — the usual focus of euro area policy — is the single largest component but only about half the story. Policymakers and researchers must look at more disaggregated (bank and firm) data to understand financial conditions. The ‘proverbial glass is half full.’

Why is the euro-area average sovereign spread, not the country-specific spread, the relevant covariate at the country level?

Regressing country-time EFP fixed effects on sovereign spreads: country-specific spreads explain 48% of country-level variation, while the EA-average spread explains nearly 80%; adding local spreads on top of the EA average yields no additional explanatory power (the local-spread coefficient is insignificant). This is consistent with variance along the time (t) dimension being much larger than across countries (c), suggesting a common factor — likely global risk aversion — drives country-level EFP variation. The EFP is roughly 2.6 times larger than the sovereign spread (specification 2). Heterogeneity: aggregate (EA) spreads matter most for large firms (multi-country operators) and short-maturity loans; for small firms and long-maturity loans the country-specific spread becomes relevant (verified with a Patton-Timmermann monotonicity test).

What evidence supports the bank lending channel at the bank level?

Bank-time EFP is regressed on bank balance-sheet and funding-cost variables. Higher EFP is associated with weaker banks: less capitalized, more exposed to risky assets, less liquid, and with higher funding costs. The 95-5 quantile range of Tier 1 capital implies almost 100 bps higher EFP. These covariates (except the interbank rate, which is common across banks and captures time variation) are bank-specific, so they reflect the bank’s own balance sheet rather than its average borrower — the essence of the bank lending channel. Larger banks also charge higher rates, which the authors suggest may reflect market power. Caveat: R-squared values are very low (~0.01-0.05), so most bank-level loan-rate behavior remains unexplained.

What evidence supports the firm balance-sheet channel at the firm level?

Firm-time EFP (net of country and bank effects) is regressed on firm fundamentals. Smaller, younger, more leveraged, and less profitable firms pay higher EFPs — a clear balance-sheet/financial-accelerator mechanism. Magnitudes from specification (4): the 5-95 leverage range implies 90 bps higher EFP; the probability-of-default distribution implies about 20 bps; old (50yr) versus young (5yr) firms differ by about 30 bps. This is notable because the sequential extraction attributes all bank-firm covariance to banks, yet firm-level drivers still appear. Caveats: covariates explain only about a fifth of firm-time variation, and part of the fit comes from including probability of default (itself a financial price).

What is found at the contract level?

After controlling for country, bank, and firm effects, residual contract-level variation arises only for firms borrowing multiple times in the same month at different rates. Regressing on loan size and maturity, both are statistically significant but collectively explain a negligible share (R-squared ~0.0001-0.0003). The authors call this a ’nothing to see here’ result and conjecture that unobserved contract characteristics — likely loan covenants — drive it; because these would correlate with size and maturity, there is omitted-variable bias, so they do not interpret the coefficients. Notably these are unsecured loans, so covenants are not about explicit collateral.

What is the identification strategy for monetary policy transmission, and what are its limits?

The authors estimate Jorda (2005) local projections of cumulative changes in the bank-time and firm-time EFP (h = 0..5 months) on high-frequency identified ECB monetary policy surprises from Altavilla, Brugnolini, Gurkaynak, Motto and Ragusa (2019) — rotated factors from OIS changes in a narrow window around announcements, interpretable as Target, Forward Guidance, and QE surprises (the QE sign is flipped so larger = larger easing). A null EFP response indicates exact pass-through of the policy rate to the loan rate, not ineffectiveness. Limits: at the country level, the analysis acknowledges it does not condition on exogenous variance, so causal claims at the country/macro covariate level are ’not strongly grounded’; the paper frames the country-level work as comovement/fact-finding. The local-projection monetary-policy results are stated as causal. Forward-guidance surprises are too small in this sample (the ECB deliberately withheld guidance) to generate identifying variation, so FG results are relegated to the appendix.

What are the main asymmetries in monetary policy transmission to the EFP?

Two sign/instrument asymmetries: (1) Policy-rate tightening (but not easing) is amplified via the EFP, mostly at the bank level, driven by weaker (less capitalized, less liquid, higher-NPL) banks. The weaker amplification from rate cuts is linked to limited policy space near the effective lower bound, which binds for cuts but not hikes. (2) QE (but not QT) is amplified via the EFP, reducing it at both bank and firm levels, with the firm-level reduction indicating a firm balance-sheet channel that helps fragile firms. Magnitudes: a one-SD Target surprise (8 bps) raises EFP ~10 bps (peak 3-5 months); a one-SD QE surprise (€500 bn) lowers EFP ~20 bps, split roughly equally bank/firm. QT is fully passed through to tighten lending but leaves the EFP unchanged.

Why does QT leave the EFP unchanged while QE moves it, and how is this tested?

The authors consider three channels: (i) QE’s signalling channel (signalling an accommodative stance near zero rates) has no QT equivalent; (ii) QE is announced in financial distress while QT occurs in calmer periods — but these concern ‘periods’ not ‘surprises,’ and in the event-study framework many QT surprises actually fall within the QE period as smaller-than-expected QE, so policy-cycle explanations don’t apply directly; (iii) the operationally relevant channel: QE arrives via large ’envelope’ announcements generating sizeable stock and flow effects (‘a loud bang’), whereas QT is implemented slowly, predictably, and designed to be ‘as unsurprising and gentle as possible,’ muting both effects. They test the third channel with a difference-in-difference comparing EFP changes around the five/six QE envelope announcement months (APP/PEPP announcements/recalibrations: September 2019, and March, April, June, December 2020) versus all other months. Both bank- and firm-level panels show no pre-trend divergence but a significant EFP decline after the envelope announcement, beyond the risk-free curve. Caveat: QT results rest on limited accumulated evidence and need reassessment; deviations from gradual balance-sheet normalization could have significant effects.

How is the bank/firm channel split corroborated via cross-sectional interactions?

Equation (9) adds interactions of monetary policy surprises with bank/firm fragility characteristics (reporting h=3). Consistent with the bank lending channel, transmission of rate-tightening and QE-easing surprises is amplified for banks with weaker regulatory positions, less liquid assets, and higher funding costs. Consistent with the firm balance-sheet channel, the EFP is reduced more strongly for fragile firms (by size, age, leverage, profitability). Two implications: QE narrowed not just sovereign spreads but also the EFP on loans to more fragile firms; and because less-liquid banks respond more to rate hikes and QT lowers system liquidity, QT and rate hikes interact — as the ECB shrinks its balance sheet, rate increases are more likely to generate financial amplification via the EFP.

What robustness checks are run on the country-level results?

Three main checks (appendix): (A1) excluding 2020 (the Covid year) entirely leaves results unchanged, so country results are not Covid-driven; (A2) restricting to loans where bank country equals firm country strengthens the result, so the irrelevance of local spreads is not driven by bank-vs-firm country matching; (A3) a long macro sample built directly from aggregate iMIR data spanning April 2005 to December 2023 yields similar results, addressing the short-T concern and validating the bottom-up micro construction. Results are also robust to using 2-year or 10-year sovereign spreads, and main results hold under OLS rather than WLS (though equal-weighting overweights small loans — the smallest 90% of loans are just 1.3% of the market). Westerlund-style cointegration tests address potential non-stationarity/spurious regression.

How does this paper relate to and differ from prior work?

It builds on the financial-accelerator literature (Bernanke-Gertler 1989; Kiyotaki-Moore 1997; Bernanke-Gertler-Gilchrist 1999) resting on a failure of Modigliani-Miller due to information asymmetries. Unlike the applied EFP literature that proxies the premium with bond spreads (Gilchrist-Zakrajsek 2012; Gilchrist-Mojon 2018) — relevant only to firms able to issue bonds, a significant limitation in the bank-intermediated euro area — this paper measures the EFP directly from bank loan rates. Unlike standard microdata work that saturates regressions with fixed effects (Khwaja-Mian 2008; Amiti-Weinstein 2018; Degryse et al. 2019) to separate supply from demand and then discards those fixed effects, this paper makes the fixed effects themselves the objects of study. On asymmetry, it adds to the literature on asymmetric monetary policy over the cycle (Keynes 1936; Cover 1992; Tenreyro-Thwaites 2016) and to the scant literature comparing instrument effectiveness during easing vs tightening (Wei 2022; Crawley et al. 2022), and complements Todorov (2020) showing QE shrinks risk premia for less creditworthy bond-market borrowers.

What are the policy implications and their scope conditions?

(1) Country-level sovereign spreads are inadequate summary statistics for euro area financial conditions — they capture only half the EFP variance — so monitoring must extend to bank and firm levels. (2) QE is effective at the micro level, narrowing the EFP especially for fragile banks and firms; it is a ‘fine substitute’ for interest-rate policy. (3) QT’s gentle, predictable implementation has so far avoided EFP amplification, but this is contingent on that specific implementation modality — a fast or surprising QT (a tightening-direction ’envelope’) could have significant effects on firm and household lending conditions. (4) Interest-rate and balance-sheet policies are complementary: as the balance sheet shrinks and liquidity falls, rate hikes become more amplifying via the EFP. Scope conditions: country-level/macro comovements are not conditioned on exogenous variance so are not strong causal claims; sovereign spreads are asset prices, not fundamentals; QT conclusions rest on a limited sample and need reassessment; the policy result reflects the specific ECB communication and operational modalities observed in 2019-2023.

What significant caveats and unexplained findings does the paper itself flag?

The paper is explicitly framed as a ‘fact-finding effort’ rather than a complete causal narrative. Most bank-, firm-, and essentially all contract-level variation remains unexplained by an extensive list of covariates (low R-squared). The finding that larger banks charge more (market power) is presented as an interpretation worth studying, not established. Country-level comovements are not causal. The QT/EFP-unchanged result rests on limited evidence. Contract-level drivers (likely loan covenants) suffer omitted-variable bias and are left uninterpreted. The authors repeatedly invite future work on causal mechanisms and sub-country determinants.

Key Concepts