## The Riskiness of Credit Origins and Downside Risks to Economic Activity

## Source details

**Canonical URL:** [The Riskiness of Credit Origins and Downside Risks to Economic Activity](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024072-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024072-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024072-print-pdf.pdf.json)

---

### Introduction — core findings and scope
- Empirical motivation:
  - Large aggregate credit expansions tend to be followed by adverse macroeconomic outcomes and financial crises.
  - Credit expansions in environments of easy financial conditions and buoyant credit sentiment are especially associated with adverse outcomes.
- Research gap:
  - Cross-country studies focus on aggregate credit volume and price but largely ignore composition of credit origination and lender heterogeneity.
- Main contribution:
  - Construct an aggregate measure, the Riskiness of Credit Origins (RCO), capturing the extent to which credit growth concentrates in relatively riskier banks.
  - Sample: 3,071 banks across 42 countries over the 1990–2019 period.
  - Key empirical result: an increase in RCO predicts downside risks to GDP growth even after controlling for aggregate credit growth and financial conditions.
  - Magnitude: a one-standard-deviation increase in RCO shifts the left tail of the average cumulative two-year-ahead GDP growth distribution by about 30 basis point in the baseline specification.
- Robustness:
  - Results robust to additional controls (including an aggregate measure of banking sector riskiness), alternative bank-level riskiness metrics, restricted samples, and alternative quantile regression methods.

### Measurement of RCO and bank riskiness
- RCO construction (four-step Greenwood and Hanson (2013) inspired approach, country-year basis):
  1. Sort banks into deciles by a bank-level riskiness indicator; assign decile position (higher decile = higher riskiness).
  2. Sort banks into top and bottom lenders by annual loan growth relative to the country median (≥ median = top; < median = bottom).
  3. Compute the average lagged riskiness decile among top and bottom lenders.
  4. RCO = average lagged riskiness decile of top lenders − average lagged riskiness decile of bottom lenders; normalize by subtracting country-specific mean so units correspond to deciles.
- Formal expression:
  - RCOc,t = (1/Nc,t^Top) Σ RISK_DECILE_{i,c,t−1} for i ∈ Topc,t − (1/Nc,t^Bottom) Σ RISK_DECILE_{i,c,t−1} for i ∈ Bottomc,t
  - The raw measure is normalized by subtracting its country-specific mean; a value of 1 implies top issuers’ average riskiness is one decile above bottom issuers.
- Baseline bank riskiness metric:
  - Opposite of a bank’s z-score, where z-score = (return on average assets + leverage ratio) / historical (three-year) standard deviation of returns on average assets. Higher z-score = safer bank; opposite = riskier bank.
- Alternative (synthetic EDF) bank riskiness measure:
  - Use balance-sheet fundamentals inspired by CAMEL/CAEL: (i) capital adequacy (principal component of total equity/total assets and z-score); (ii) loan loss provisions/total assets; (iii) return on average assets; (iv) cost-to-income ratio; (v) liquidity (principal component of loans/assets, loans/deposits, liquid assets/total assets, liquid assets/deposits).
  - Procedure: run OLS of these fundamentals on log EDF for subsample with EDF data, include country-year fixed effects, use estimated coefficients (excluding fixed effects) to project bank riskiness for all banks with fundamentals data, then center to zero mean within country (referred to as a “synthetic EDF”).

### Data sources and sample construction
- Banks’ financial statement data sourced from Fitch Connect; two vintages merged (data downloads in June 2018 and March 2021).
- Historical consolidated financial statements merged with historical ownership data (Orbis Historical and IMF (2015)) and mergers and acquisitions data (Orbis M&A); acquiring bank kept and year of acquisition flagged (those bank-years excluded from econometric analysis); subsidiaries dropped when parent is in same country and sample.
- Sample inclusion criteria:
  - Banks with total assets above 0.5 percent of the total assets of their country’s largest bank during at least one year and with at least five years of data.
  - For each country, only years with at least ten banks meeting the above criteria are kept.
  - Only countries with at least five years meeting those conditions are retained.
  - Only countries where a financial conditions index (FCI) could be built are kept.
- Resulting baseline sample:
  - 44,515 total bank-year observations.
  - 39,730 observations have required information to construct the z-score.
  - Observations come from 3,071 banks in 42 countries from 1990 to 2019.
- Loan-level dataset (Dealogic syndicated loans) final regression sample: 1,676 bank-year observations from 1993 to 2019, covering 99 banks from 17 countries; average LLS in regression sample is 22.6 percent.
- Financial Conditions Index (FCI) estimated for 1990–2019 at quarterly frequency using eight price-based indicators; annual average of quarterly FCIs used.

### Main empirical results — cyclicality and micro-level behavior
- RCO cyclicality (country-level OLS):
  - Greater credit expansions are associated with statistically significant higher RCO contemporaneously and one year ahead; lead relationship stronger.
  - Quantitative example: An increase in the credit-to-GDP ratio by one standard deviation (4.9 percentage points) is associated with an increase in RCO by 0.2 standard deviations (a value of 0.26) one year ahead (column (2) example).
  - FCI is positively related to contemporaneous and future RCO, but statistical significance is sensitive to inclusion of year fixed effects.
- Bank-level lending behavior (Table 3):
  - On average, a riskier bank expands its loan portfolio more slowly than a safer bank (Loan Growth coefficient on riskiness -0.348***, standard error 0.035).
  - The negative effect of bank risk on loan growth weakens when aggregate credit expansion is stronger (Interaction Riskiness X Δ(Credit/GDP) for Loan Growth: 0.047***, standard error 0.009).
  - Quantitative illustration: A one-standard-deviation increase in the credit-to-GDP ratio is associated with relatively faster annual lending growth by 1.1 percentage points for a bank at the second decile of the z-score distribution relative to a bank at the 8th decile.
  - Similar patterns for total assets and total debt growth; no analogous pattern for total equity growth.

### RCO and downside risks to GDP growth (growth-at-risk)
- Method: panel quantile regressions for percentiles τ ∈ {20, 50, 80} of 1- to 3-year-ahead average cumulative GDP growth; explanatory variables enter as three-year moving averages; estimation primarily via Canay (2011).
- Main findings (Table 4):
  - Greater RCO shifts the left tail (20th percentile) and, to a lesser extent, the median (50th percentile) of the one-year ahead growth distribution leftward.
  - The leftward shift remains significant at two- and three-year horizons and extends to the right tail (80th percentile) at the three-year horizon.
  - Quantitative finding: A one-standard-deviation increase in RCO moves the left tail of the two-year ahead average annual real GDP growth distribution by 31 basis points (RCO: -0.311***, standard error 0.091).
  - Stronger credit growth worsens future downside risks across quantiles; looser global financial conditions lead to leftward shifts especially in low quantiles.
- Robustness (Table 5):
  - Results robust across alternative estimation methods (including Machado and Santos-Silva), restricted samples, and synthetic EDF-based RCO; negative association between RCO and downside growth is particularly robust at lower quantiles and multi-year horizons.

### Channels: asset quality, sentiment, and resilience
- Asset-quality channel (Table 6):
  - Banks that grow their loan book relatively faster suffer deterioration in loan-loss provisions and NPL ratios down the road.
  - Deterioration stronger when the expanding bank is ex-ante riskier (interaction Riskiness X High Loan Growth positive and significant across horizons).
  - Examples (basis points):
    - Flow of Loan Loss Provisions: Riskiness X High Loan Growth h=2: 1.421*** (0.433).
    - Change in NPLs: Riskiness X High Loan Growth positive and highly significant for h=1 to h=3.
- Sentiment channel (Tables 8–10):
  - Higher RCO predicts future tightening in bank lending standards at horizons of 1 to 2 years ahead (e.g., Table 8 h=1: 0.142* (0.072); h=2: 0.115** (0.052)).
  - Higher RCO predicts future tightening in financial conditions at 1 and 2 years ahead; predictive power conditional on current financial conditions appears at the two-year horizon.
  - Including ΔFCI(t+1) in growth-at-risk regressions reduces RCO coefficient magnitude and can eliminate significance at horizons up to two years, consistent with a sentiment/intermediary channel.
- Resilience channel (Tables 11–13):
  - RCO predicts adverse banking-sector stock return performance, significantly for extreme left tail (e.g., one-decile increase in RCO predicts a shift in very left tail of one-year-ahead stock return distribution by about 480 basis points).
  - Riskier banks exhibit lower future loan growth, especially during systemic crises (Interaction Riskiness X Crisis h=1: -0.445*** (0.084)).
  - Controlling for average riskiness and leverage skewness reduces RCO coefficient magnitude by about 20 percent at some horizons, suggesting a resilience component to RCO’s predictive power.

### Syndicated loans and leveraged loan share (LLS)
- Leveraged Loan Share (LLS) defined as share of a bank’s syndicated lending origination classified as leveraged or highly leveraged in a year; computable only for banks active in syndicated loans market.
- Local projections show:
  - Riskier banks have relatively lower LLS when portfolio growth is low.
  - The relationship reverses when portfolio growth is high; stronger risk-taking occurs up to 2 years ahead.
- Caveat: smaller sample size for syndicated-loan exercises; average LLS in loan-level regression sample is 22.6 percent.

### Key quantitative statistics and summary figures
- Sample and data counts:
  - 3,071 banks, 42 countries, coverage 1990–2019.
  - 44,515 total bank-year observations; 39,730 observations with required information to construct the z-score.
- Summary statistics (selected, preserved exactly as reported):
  - z-score (bank-level): Average z-score: 147; Median z-score: 59; 25th percentile z-score: 25; Less than 1 percent of observations have a z-score smaller than 1.
  - Synth EDF: By construction has a zero average; Standard deviation across all observations: 0.3.
  - Annual nominal loan growth (local currency): Overall average: 7.8 percent; Standard deviation: more than twice the mean.
  - LLP: Average LLP: 1 percent.
  - NPL: Average NPL: 4.4 percent; Average change in NPL Δ(NPL): 0.2 percent.
  - Country-level Δ(Credit/GDP): Average: 1.1 percent; Standard deviation: 4.9 percent.
  - FCI: mean 0; standard deviation 1.
  - RCO baseline SD: 1.3; IQR: 1.6. RCO synthetic EDF SD: 1.4; IQR: 1.8.
- Selected regression magnitudes (preserved):
  - Country-level OLS Δ(Credit/GDP) coefficients: h=0: 0.029* (0.016); h=1: 0.051*** (0.017).
  - RCO effect on growth-at-risk (Table 4): h=1 τ=20: RCO -0.186* (0.099); h=2 τ=20: RCO -0.311*** (0.091); h=3 τ=20: RCO -0.278*** (0.062).
  - Bank-level Loan Growth riskiness coefficient: -0.348*** (0.035); Interaction Riskiness X Δ(Credit/GDP): 0.047*** (0.009).
  - Asset-quality interactions (Table 6): Riskiness X High Loan Growth h=2 LLP: 1.421*** (0.433); NPL interactions highly significant across h=1 to h=3.
  - Banking-sector returns (Table 11): h=1 τ=10 RCO -4.788** (2.359).

### Interpretation, implications, and policy relevance
- Interpretation:
  - RCO is procyclical with respect to aggregate credit expansions and less strongly related to financial conditions.
  - Large credit expansions not only increase aggregate leverage but also shift composition of lending toward relatively riskier banks, adding a compositional channel to macrofinancial vulnerability.
  - Asset-quality deterioration following faster lending by riskier banks is a key mechanism linking RCO to downside growth risks; sentiment and resilience channels also contribute.
- Policy implications:
  - Importance of accounting for lender heterogeneity (in addition to borrower heterogeneity) in empirical and theoretical models of the credit cycle.
  - Findings are relevant for prudential authorities and calibration of macroprudential capital buffers, linking micro- and macro-prudential approaches.
- Suggested directions for future research:
  - Use more granular datasets to further analyze mechanisms through which RCO matters, including lender–borrower matching over the credit cycle.

*Source: wpiea2024072-print-pdf - References (IMF Working Paper content unit).*

### References .............................................................................................................

### wpiea2024072-print-pdf - References .............................................................................................................

### Acronyms/Glossary
- BLS: Bank Lending Standards
- CA: Current Account
- EDF: Expected Default Frequency
- FCI: Financial Conditions Index
- GDP: Gross Domestic Product
- GFC: Global Financial Crisis
- GUO: Global Ultimate Owner
- IFS: International Financial Statistics
- LLP: Loan Loss Provisions
- OLS: Ordinary Least Squares
- NPL: Nonperforming Loans
- RCA: Riskiness of Credit Allocation
- RCO: Riskiness of Credit Origins

### Introduction — core findings and scope
- Empirical motivation:
  - Large aggregate credit expansions tend to be followed by adverse macroeconomic outcomes and financial crises (cites include Jorda et al. 2011; Schularick and Taylor 2012; Mian et al. 2018).
  - Credit expansions in environments of easy financial conditions and buoyant credit sentiment are especially associated with adverse outcomes (cites include Krishnamurthy and Muir 2017; López-Salido et al. 2017; Kirti 2021; Adrian et al. 2022; Greenwood et al. 2022).
- Research gap addressed:
  - Existing cross-country empirical studies focus on aggregate measures of volume and price of credit and largely ignore the role of composition of credit origination and lender heterogeneity.
- Main contribution:
  - Construct an aggregate measure, the Riskiness of Credit Origins (RCO), capturing the extent to which credit growth concentrates in relatively riskier banks.
  - Sample: 3071 banks across 42 countries over the 1990–2019 period.
  - Key empirical result: an increase in RCO predicts downside risks to GDP growth even after controlling for aggregate credit growth and financial conditions.
  - Magnitude: a one-standard-deviation increase in RCO shifts the left tail of the average cumulative two-year-ahead GDP growth distribution by about 30 basis point in the baseline specification.
- Robustness:
  - Results robust to additional controls (including an aggregate measure of banking sector riskiness), alternative bank-level riskiness metrics, restricted samples, and alternative quantile regression methods.

### Mechanisms and channels explored
- Credit-quality channel:
  - Bank-level finding: banks that expand credit relatively faster experience greater increases in loan loss provisions and nonperforming loan ratios later.
  - Amplification: the deterioration is even stronger when the expanding bank is ex-ante riskier (lower relative z-score).
  - Macro-level: inclusion of a variable capturing riskier allocation of credit affects RCO’s explanatory power for horizons up to two years.
- Sentiment channel:
  - RCO predicts future changes in aggregate bank lending standards and financial conditions at horizons up to two years, supporting a sentiment interpretation.
- Resilience (distributional) channel:
  - RCO may reflect concentration of loan portfolios in riskier banks, reducing aggregate resilience if those banks cut lending after shocks and borrowers face frictions in switching lenders.
  - Supporting evidence: bank riskiness predicts future bank-level lending activity following large negative shocks; RCO predicts leftward shifts of the extreme left tail of banking sector stock returns.

### Theoretical underpinnings and literature links
- Theoretical ambiguity:
  - Risk-shifting incentives under limited liability (Jensen and Meckling 1976) and lower capitalization reducing monitoring incentives (Holmstrom and Tirole 1997; Allen et al. 2011) suggest riskier banks take more risk.
  - Countervailing forces: threat of runs and regulatory/bondholder constraints can limit risk-taking (Jacklin and Battacharya 1988; Diamond and Rajan 2000; Ashcraft 2008; Dewatripont and Tirole 2012).
- Amplification during booms:
  - Weaker banks are predicted to relax lending standards more in good times due to debt-overhang and lower screening gains (Myers 1977; Ruckes 2004; Dell’Ariccia and Marquez 2006; Berger and Udell 2004).
  - Diagnostic expectations and neglect of crash risk lower the price of risk in expansion phases (Bordalo et al. 2018; Baron and Xiong 2017).
- Relation to existing empirical literatures:
  - Bank-level evidence on cyclicality of lending and risk-taking (Igan and Tamirisa 2008; Igan and Pinheiro 2011).
  - Risk-taking channel of monetary policy (Jimenez et al. 2014; Dell’Ariccia et al. 2017).
  - Banking crisis and growth-at-risk literatures (Gourinchas et al. 2001; Obstfeld 2012; Schularick and Taylor 2012; Dell’Ariccia et al. 2016; Jordà et al. 2021; Giglio et al. 2016; Adrian et al. 2019; Adrian et al. 2022).
  - Micro evidence on loan growth and future deterioration (Jimenez and Saurina 2006; Chavan and Gambacorta 2019; Foos et al. 2010; Fahlenbrach et al. 2018).
  - Distributional banking vulnerability and leverage skewness (Coimbra and Rey 2018, 2023; Korinek and Nowak 2017).

### Measurement of RCO and bank riskiness
- RCO construction (four-step Greenwood and Hanson (2013) inspired approach applied to banks, country-year basis):
  1. Sort banks into deciles by a bank-level riskiness indicator; assign decile position (higher decile = higher riskiness).
  2. Sort banks into top and bottom lenders by annual loan growth relative to the country median (≥ median = top; < median = bottom).
  3. Compute the average lagged riskiness decile among top and bottom lenders.
  4. RCO = average lagged riskiness decile of top lenders − average lagged riskiness decile of bottom lenders; normalize by subtracting country-specific mean so units correspond to deciles.
- Formal expression (verbatim structure described in text):
  - RCOc,t = (1/Nc,t^Top) Σ RISK_DECILE_{i,c,t−1} for i ∈ Topc,t − (1/Nc,t^Bottom) Σ RISK_DECILE_{i,c,t−1} for i ∈ Bottomc,t
  - The raw measure is normalized by subtracting its country-specific mean; a value of 1 implies top issuers’ average riskiness is one decile above bottom issuers.
- Baseline bank riskiness metric:
  - Opposite of a bank’s z-score, where z-score = (return on average assets + leverage ratio) / historical (three-year) standard deviation of returns on average assets.
  - Interpretation: higher z-score = safer bank; opposite (less negative) = riskier bank.
- Alternative (synthetic EDF) bank riskiness measure:
  - Use balance-sheet fundamentals inspired by CAMEL/CAEL: (i) capital adequacy (principal component of total equity/total assets and z-score); (ii) loan loss provisions/total assets; (iii) return on average assets; (iv) cost-to-income ratio; (v) liquidity (principal component of loans/assets, loans/deposits, liquid assets/total assets, liquid assets/deposits).
  - Procedure: run OLS of these fundamentals on log EDF for subsample with EDF data, include country-year fixed effects, use estimated coefficients (excluding fixed effects) to project bank riskiness for all banks with fundamentals data, then center to zero mean within country. This projected measure is referred to as a “synthetic EDF”.
- Implementation notes:
  - The paper focuses on within-country dynamics; normalization removes cross-country sectoral composition effects.
  - Using contemporaneous riskiness decile instead of lagged decile yields very similar results in country-level analysis.

*Source: wpiea2024072-print-pdf - References*

### Appendix 2 for a detailed explanation).

### wpiea2024072-print-pdf - Appendix 2 for a detailed explanation)

### Data sources and sample construction
- Banks’ financial statement data sourced from Fitch Connect; two vintages merged (data downloads in June 2018 and March 2021).
- Financial statements filtered and cleaned by bank specialization and market description; basic requirements imposed on key balance sheet ratios; duplicates removed.
- Historical consolidated financial statements merged with historical ownership data (sourced from Orbis Historical and IMF (2015)) and mergers and acquisitions data (sourced from Orbis M&A).
  - In cases of mergers and acquisitions, the acquiring bank is maintained in the sample and the year of acquisition is flagged; these bank-years are not included in the econometric analysis.
  - Subsidiaries are dropped when the parent bank is in the same country and in the sample to avoid double counting.
- Banks that ceased to exist earlier and were removed from the March 2021 download were reintroduced using the June 2018 download.
- Sample inclusion criteria:
  - Banks with total assets above 0.5 percent of the total assets of their country’s largest bank during at least one year and with at least five years of data.
  - For each country, only years with at least ten banks meeting the above criteria are kept.
  - Only countries with at least five years meeting those conditions are retained.
  - Only countries where a financial conditions index (FCI) could be built are kept.
- Resulting baseline sample:
  - 44,515 total bank-year observations.
  - 39,730 observations have the required information to construct the z-score.
  - Observations come from 3,071 banks in 42 countries from 1990 to 2019.
- A restricted sample is also constructed considering only country-years with at least 20 banks satisfying the criteria to check robustness.

### Macroeconomic and supplementary data
- Baseline credit series sourced from IMF’s International Financial Statistics (IFS); captures credit to the private sector from domestic money banks.
- Other macroeconomic series (nominal GDP, real GDP, current account) sourced from IFS.
- Financial Conditions Index (FCI):
  - Estimated for 1990–2019 at a quarterly frequency using eight price-based financial indicators:
    1. term spread
    2. corporate spread
    3. sovereign spread
    4. interbank spread
    5. first difference in real long-term rate
    6. equity returns
    7. equity volatility
    8. house price returns
  - Annual average of quarterly FCIs used as the measure of financial conditions for each year.
- Country-level data series on the riskiness of credit allocation (RCA) constructed based on Brandao-Marques et al. (2022) using Worldscope data.
- Banks’ EDF data sourced from Moody’s CreditEdge.

### Bank-level descriptive statistics (Panel A)
- z-score:
  - Average z-score: 147.
  - Median z-score: 59.
  - 25th percentile z-score: 25.
  - Less than 1 percent of observations have a z-score smaller than 1.
- EDF (synthetic EDF riskiness measure):
  - By construction has a zero average.
  - Standard deviation across all observations: 0.3 (very similar to its interquartile range).
- Annual nominal loan growth (measured in local currency terms):
  - Overall average: 7.8 percent.
  - Standard deviation: more than twice the mean (indicating substantial variation).
- Loan loss provisions (LLP) and nonperforming loans (NPL), both expressed as a share of lagged gross loans:
  - Average LLP: 1 percent.
  - Average NPL: 4.4 percent.
  - Average change in NPL (expressed as a percentage of lagged gross loans and labeled Δ(NPL)): 0.2 percent.

### Country-level descriptive statistics (Panel B)
- Change in the ratio of total credit to GDP across country-years:
  - Average: 1.1 percent.
  - Standard deviation: 4.9 percent.

*Source: wpiea2024072-print-pdf - Appendix 2 for a detailed explanation).*

### 4.4 percent, thus exhibiting significant variation. By construction, the FCI has a mean of zero and a standard

### The Riskiness of Credit Origins and Downside Risks to Economic Activity

### Summary statistics and measures
- The Financial Condition Index (FCI) is constructed to have a mean of zero and a standard deviation of 1.
- RCO measures have a zero mean; baseline (z-score) and alternative (synthetic EDF) summary statistics:
  - Baseline (z-score): standard deviation across country-years is 1.3; interquartile range is 1.6.
  - Alternative (synthetic EDF): standard deviation is 1.4; interquartile range is 1.8.

### Cyclicality of the Riskiness of Credit Origins (RCO)
- Cross-country and selected-country patterns (visual evidence described):
  - Global RCO increased before the late 1990s dot-com boom, declined after it, rose before the global financial crisis (GFC), plateaued during the crisis, and dropped abruptly during the euro area sovereign debt crisis.
  - Country-specific dynamics:
    - United States: RCO dropped significantly during the GFC and immediate aftermath and rebounded strongly afterward.
    - Germany: no post-GFC rise.
    - Ireland: post dot-com decline was brief; RCO rose almost non-stop from the late 1990s to 2007 GFC onset.
    - Spain: an additional RCO cycle with a spike in 2015 after recovery from an earlier financial crisis.

### Country-level econometric evidence (panel regression results)
- Regression specification (summary): RCOc,t+h regressed on ∆(Credit-to-GDP)c,t, FCIc,t, real GDP growthc,t, with country and year fixed effects; standard errors clustered at country level; horizons h=0,1.
- Main findings (Table 2, Panel A and robustness panels):
  - Greater credit expansions are associated with statistically significant higher RCO contemporaneously and one year ahead.
  - Lead relationship stronger: coefficients larger and more significant for h=1.
  - Quantitative example (column (2)):
    - An increase in the credit-to-GDP ratio by one standard deviation (4.9 percentage points) is associated with an increase in RCO by 0.2 standard deviations (a value of 0.26) one year ahead.
  - Inclusion of year fixed effects reduces the magnitude and significance of credit-expansion coefficients, indicating part of explanatory power stems from global cycles.
  - FCI is positively related to contemporaneous and future RCO, but statistically significant only when year fixed effects are excluded.
    - Quantitatively, a one standard deviation increase in FCI is associated with a 0.1 standard deviation increase in RCO (0.1–0.15).
  - Interaction of credit expansion size and FCI: interaction coefficients not significant (unreported).
  - Real GDP growth coefficient: not statistically significant in country-level regressions (no contemporaneous or lagged relationship).
- Robustness:
  - Panel B (restricted sample), Panel C (synthetic EDF RCO), Panel D (additional macro controls) yield similar conclusions.
  - Relationship between RCO and credit expansions positive and almost always statistically significant: 11 out of 12 cases.
  - Relationship with FCI mostly positive but significant in only 3 cases out of 12; significance disappears with year fixed effects.
  - Real GDP growth significant in only one specification without year fixed effects.

### Micro-level (bank-level) evidence on lending behavior
- Bank-level panel specification (summary): bank growth rate regressed on lagged growth, bank riskiness decile, interaction of bank riskiness and country credit expansion, bank-level controls (liquid asset ratio, bank loan market share), bank and country-year fixed effects; standard errors clustered at bank level.
- Key findings (Table 3):
  - On average over the cycle, a riskier bank expands its loan portfolio more slowly than a safer bank.
  - This negative effect of bank risk on loan growth weakens (becomes less negative) when aggregate credit expansion is stronger.
  - Quantitative illustration:
    - A one-standard-deviation increase in the credit-to-GDP ratio is associated with relatively faster annual lending growth by 1.1 percentage points for a bank at the second decile of the z-score distribution relative to a bank at the 8th decile.
  - Similar patterns for total assets and total debt growth; no analogous pattern for total equity growth.
- Robustness:
  - Panel B of Table 3: interaction coefficient similar across alternative samples and synthetic EDF riskiness indicator.

### RCO and downside risks to GDP growth (quantile regressions / growth-at-risk)
- Approach:
  - Panel quantile regressions for percentiles τ ∈ {20, 50, 80} of 1- to 3-year-ahead average cumulative GDP growth; explanatory variables enter as three-year moving averages.
  - Baseline controls include change in credit-to-GDP ratio, FCI, real GDP growth, and global financial conditions.
  - Estimation primarily via Canay (2011); robustness checks using Machado and Santos-Silva (2019) and other methods.
- Main results (Tables 4 and 5):
  - Greater RCO shifts the left tail (20th percentile) and, to a lesser extent, the median (50th percentile) of the one-year ahead growth distribution leftward.
  - The leftward shift remains significant at two- and three-year horizons and extends to the right tail (80th percentile) at the three-year horizon.
  - Quantitative finding:
    - A one-standard-deviation increase in RCO moves the left tail of the two-year ahead average annual real GDP growth distribution by 31 basis points.
  - Stronger credit growth worsens future downside risks across quantiles.
  - Looser global financial conditions lead to leftward shifts, especially in low quantiles; looser domestic financial conditions shift distributions rightward.
- Robustness (Table 5, Panel B):
  - Using Machado and Santos-Silva (2017) method, restricted sample, and synthetic EDF-based RCO:
    - Coefficient for the 20th percentile is negative at all horizons and significant at 2-year and 3-year horizons.
    - Coefficients for the 50th percentile are negative and mostly significant at 2- and 3-year horizons.
- Overall: greater RCO is a robust predictor of increased downside risks to economic activity.

### Channels: Why does RCO predict downside risks?
- Three candidate channels:
  1. Asset quality channel: riskier banks expanding faster lend to riskier borrowers, weakening future loan portfolio quality.
  2. Sentiment channel: high RCO captures buoyant sentiment in banking or financial markets.
  3. Resilience channel: concentration of credit in riskier, less resilient banks increases likelihood of lending contraction in adverse shocks due to frictions in borrower reallocation.
- Asset quality evidence (local-projections regressions, Equations (5) and (6)):
  - Dependent variables: future flow of loan-loss provisions and future change in nonperforming loan (NPL) ratio, expressed in basis points.
  - Covariates include bank riskiness decile, a dummy for whether bank loan growth is above median, interaction term, bank controls (loan market share, liquidity), bank and country-year fixed effects.
  - Findings (Table 6):
    - Banks that grow their loan book relatively faster suffer a deterioration in the flow of loan-loss provisions down the road.
    - The deterioration is stronger when the bank is ex-ante riskier (positive and significant interaction coefficient).
    - Results echo evidence in Fahlenbrach et al. (2018) for U.S. banks.

### Key quantitative takeaways
- Standard deviations and variability:
  - FCI: mean 0; standard deviation 1.
  - RCO baseline SD: 1.3; IQR: 1.6.
  - RCO synthetic EDF SD: 1.4; IQR: 1.8.
- Credit expansion and RCO:
  - One SD increase in credit-to-GDP (4.9 percentage points) → RCO increases by 0.2 SD (0.26) one year ahead.
- Bank-level amplification:
  - One SD increase in credit-to-GDP → relatively faster lending growth by 1.1 percentage points for riskier-vs-safer decile comparison.
- Growth-at-risk:
  - One SD increase in RCO → left tail of two-year ahead average annual GDP growth shifts by 31 basis points.

### Interpretation and implications (from findings)
- RCO is procyclical with respect to aggregate credit expansions and less strongly related to financial conditions.
- Large credit expansions not only increase aggregate leverage but also change the composition of lending toward relatively riskier banks, adding a compositional channel to macrofinancial vulnerability.
- RCO provides incremental predictive power for downside growth risks beyond the size of the credit expansion and standard macro-financial controls.
- Asset-quality deterioration following faster lending by riskier banks is an important mechanism linking RCO to future downside risks.

*IMF Working Paper — excerpted content from "The Riskiness of Credit Origins and Downside Risks to Economic Activity"*

### conclusions when we assess the quality of banks’ loan portfolios using the cumulative change in the NPL ratio

### conclusions when we assess the quality of banks’ loan portfolios using the cumulative change in the NPL ratio

### Syndicated loans / Leveraged Loan Share (LLS) and bank riskiness
- Constructed Leveraged Loan Share (LLS): share of a bank’s syndicated lending origination classified as leveraged or highly leveraged in a year; computable only for banks participating in the syndicated loans market.
- Local projections regression (Equation (7)) findings (Table 6, columns (7)-(9)):
  - Riskier banks have a relatively lower LLS when their portfolio growth is relatively low.
  - The sign of this relationship reverses when portfolio growth is relatively high; this stronger risk-taking occurs up to 2 years ahead.
  - Caveat: smaller bank sample size for this exercise.
- Footnotes:
  - 20: Information on whether banks retain these loans on their balance sheets or sell them after origination is not publicly available.
  - 21: The analysis also indicates that RCO’s predictive power is more persistent that RCA’s.

### Aggregate asset-quality proxy (RCA) versus RCO in growth-at-risk
- RCA: aggregate measure of the riskiness of borrowers (from Brandao et al. (2022)), constructed similarly to RCO and captures extent to which more leveraged borrowers are expanding their debt relatively faster; included as a three-year moving average.
- Panel quantile regressions for the twentieth percentile at horizons h=1,2,3 (Table 7):
  - When RCA and RCO enter separately, both are significant at horizons up to two years.
  - Controlling for RCA renders RCO coefficients:
    - Insignificant at the 1-year horizon.
    - Decreased in size at longer horizons without affecting significance.
  - Interpretation: either RCO is a better aggregate proxy for asset quality or other channels beyond aggregate asset-quality are at play.

### Sentiment channel: banking sector sentiment and investor sentiment
- Data and sample:
  - Quarterly changes in bank lending standards (BLS) from bank loan officer surveys collapsed to annual frequency by simple average and standardized within a year.
  - Data available for a subsample of 31 countries; more than five countries have data only after 2003; sample size reduced by about half for these exercises.
- Local projections (Equation (8)) linking RCO to future BLS (Table 8):
  - A higher RCO helps predict future tightening in lending standards at horizons of 1 to 2 years ahead, including when current bank lending standards are controlled for (columns (7)-(9)).
  - Indicates RCO’s predictive power for downside risks to growth may work via predicting banking sector sentiment.
- Investor sentiment / financial conditions (Equation (9), Table 9):
  - A higher RCO helps predict future tightening in financial conditions at 1 and 2 years ahead.
  - When concurrent financial conditions are controlled for, predictive power appears only at the two-year horizon.
- Additional evidence (Table 10):
  - Including the change in financial conditions at t+1 in growth-at-risk regressions reduces the magnitude of the RCO coefficient and eliminates its significance at horizons of up to two years.
- Footnote:
  - 22: Survey construction and collapsing method described (see source).

### Resilience channel: RCO, banking-sector stock returns, and future lending
- Panel quantile regressions of banking-sector cumulative returns (Equation (10), Table 11):
  - Focused quantiles: twentieth, fiftieth, eightieth, and tenth percentiles.
  - RCO generally predicts adverse banking sector stock return performance, significantly mostly for the tenth percentile at horizons of up to two years.
  - Baseline: a one-decile increase in RCO predicts a shift in the very left tail of the one-year-ahead stock return distribution by about 480 basis points.
- Bank-level loan growth local projections (Equation (11), Table 12):
  - Explored horizons up to four years (h=1,2,3,4); systemic banking crisis dummy from Laeven and Valencia (2020).
  - Riskier banks tend to exhibit lower rates of future loan growth, especially at shorter horizons (columns (1)-(4)).
  - Lower growth rates of riskier banks are especially pronounced during systemic financial crises (columns (5)-(8)).
  - A one-decile-increase in the bank-level riskiness measure is associated with a deterioration in gross loan growth performance of 0.45 percentage points if a crisis hits the following year.
  - The deterioration is present at longer horizons too, although its magnitude is about half.
- Aggregate fragility controls and RCO (Table 13):
  - Augmented growth-at-risk regressions with:
    - Average riskiness (negative of average z-score) across banks.
    - Skewness of the asset-weighted leverage distribution (as in Coimbra and Rey 2018).
  - Findings:
    - Average riskiness is always significant and indicates a higher aggregate z-score is a source of resilience.
    - Asset-weighted leverage skewness does not have a particularly strong relationship with the left tail of future growth.
    - RCO coefficient magnitude declines by about 20 percent at the 1-year and 3-year horizons when controlling for the two resilience indicators, suggesting a resilience channel in aggregate.
  - Footnote:
    - 24: Loan growth rates are serially correlated; top lenders keep growing faster for the subsequent two years and revert to the mean after 3 years.
    - 25: Skewness is negatively related to future median GDP growth at the three-year horizon (unreported).
    - 26: Robustness check including RCA on the right-hand side leaves results in Table 13 almost identical.

### Overall conclusions and policy relevance
- Constructed RCO: measures extent to which aggregate credit growth is driven by relatively riskier banks.
- Main empirical findings:
  - RCO is procyclical and uncovers a new dimension of the credit cycle.
  - RCO helps predict future macroeconomic outcomes and downside risks to GDP growth after controlling for the size of credit expansion and financial conditions.
- Three inter-related channels underpinning RCO’s predictive power:
  1. Asset-quality channel: faster growth by riskier banks appears to target riskier borrowers, adversely affecting asset quality.
  2. Sentiment channel: RCO predicts future tightening in bank lending standards and financial conditions, acting as a measure of banking sector and investor sentiment.
  3. Resilience channel: concentration of lending in riskier banks reduces aggregate lending capacity following large adverse shocks because riskier banks cut lending relatively more.
- Policy implications:
  - Highlights importance of accounting for lender heterogeneity (in addition to borrower heterogeneity) in empirical and theoretical models of the credit cycle.
  - Findings are relevant for prudential authorities and for calibration of macroprudential capital buffers, linking micro- and macro-prudential approaches.
- Suggested direction for future research:
  - Use more granular datasets to further analyze mechanisms through which RCO matters, including lender–borrower matching over the credit cycle.

*IMF Working Paper: The Riskiness of Credit Origins and Downside Risks to Economic Activity — conclusions when we assess the quality of banks’ loan portfolios using the cumulative change in the NPL ratio*

### References

### References

### Key literature cited
- Extensive bibliography on banking, credit cycles, and financial stability, including works by Adrian et al., Allen et al., Brunnermeier and Sannikov, Diamond and Rajan, Dell’Ariccia et al., Jiménez et al., Jordà et al., Laeven and Valencia, Mian and Sufi, Schularick and Taylor, and many others.
- Methodological and econometric sources referenced include Canay (2011), Machado and Santos Silva (2019), and Koop and Korobilis (2014).
- Empirical and survey studies on bank lending behavior, loan growth, capital regulation, and bank risk are cited across journals such as American Economic Review, Journal of Finance, Journal of Financial Intermediation, Review of Financial Studies, and Journal of Banking and Finance.

### Summary statistics (Table 1 — bank- and country-level variables)
- Bank-level (overall bank-year):
  - Z-score: Mean 147, Standard Deviation 252, Median 59, 25th percentile 144, 75th percentile 25, Number of observations 39,730.
  - Synth EDF: Mean 0.00, Standard Deviation 0.30, Median 0.00, 25th percentile -0.14, 75th percentile 0.14, Number of observations 41,085.
  - Loan Growth (percent): Mean 7.82, Standard Deviation 16.96, Median 6.40, 25th percentile -0.97, 75th percentile 15.67, Number of observations 43,091.
  - LLP (percent): Mean 1.02, Standard Deviation 1.67, Median 0.49, 25th percentile 0.13, 75th percentile 1.20, Number of observations 44,515.
  - NPL (percent): Mean 4.43, Standard Deviation 5.40, Median 2.51, 25th percentile 0.97, 75th percentile 5.66, Number of observations 32,004.
  - Δ(NPL) (percent): Mean 0.21, Standard Deviation 2.05, Median 0.02, 25th percentile -0.33, 75th percentile 0.55, Number of observations 30,298.
- Country-level (overall country-year):
  - RCO (z-score, deciles): Mean 0.0, Standard Deviation 1.3, Median 0.0, 25th percentile -0.8, 75th percentile 0.8, Number of observations 833.
  - RCO (synth. EDF, deciles): Mean 0.0, Standard Deviation 1.4, Median 0.0, 25th percentile -0.9, 75th percentile 0.9, Number of observations 855.
  - Δ(Credit/GDP) (percent): Mean 1.1, Standard Deviation 4.9, Median 1.0, 25th percentile -1.2, 75th percentile 3.2, Number of observations 1,280.
  - FCI: Mean 0.0, Standard Deviation 1.0, Median 0.1, 25th percentile -0.5, 75th percentile 0.7, Number of observations 1,320.
  - GDP growth (percent): Mean 3.1, Standard Deviation 3.6, Median 3.0, 25th percentile 1.4, 75th percentile 5.1, Number of observations 1,310.
  - CA/GDP (percent): Mean 0.1, Standard Deviation 2.5, Median 0.0, 25th percentile -0.9, 75th percentile 0.9, Number of observations 1,303.

### RCO cyclicality and macro links (Table 2 — country-level OLS)
- Baseline (Panel A): Δ(Credit/GDP) coefficients:
  - h=0: 0.029* (standard error 0.016)
  - h=1: 0.051*** (standard error 0.017)
- Baseline (Panel A): FCI coefficients:
  - h=0: 0.096* (standard error 0.056)
  - h=1: 0.146** (standard error 0.065)
- Restricted sample (Panel B): Δ(Credit/GDP):
  - h=0: 0.025* (standard error 0.013)
  - h=1: 0.036** (standard error 0.013)
- Alternative RCO measure (Panel C) and with macro controls (Panel D) report similar patterns where Δ(Credit/GDP) and FCI show positive and in some cases statistically significant associations with RCO. Standard errors and significance levels are reported per cell; * p<10 percent, ** p<5 percent, *** p<1 percent.

### Bank-level relationships: Riskiness and balance-sheet dynamics (Table 3)
- Baseline bank-level OLS (Panel A) — Riskiness (negative z-score) effects:
  - Loan Growth: coefficient -0.348*** (standard error 0.035)
  - Asset Growth: coefficient -0.379*** (standard error 0.034)
  - Debt Growth: coefficient -0.459*** (standard error 0.039)
  - Equity Growth: coefficient 0.279*** (standard error 0.044)
- Interaction Riskiness X Δ(Credit/GDP):
  - Loan Growth: 0.047*** (standard error 0.009)
  - Asset Growth: 0.038*** (standard error 0.009)
  - Debt Growth: 0.034*** (standard error 0.010)
  - Equity Growth: 0.001 (standard error 0.012)
- Panel B confirms robustness using Synthetic EDF and a restricted sample; interaction coefficients remain positive and statistically significant.

### RCO and Growth-at-Risk (Table 4 — quantile regressions)
- One-year ahead (h=1), τ=20:
  - Real GDP Growth: 0.349*** (0.085)
  - Δ(Credit/GDP): -0.116*** (0.026)
  - FCI: 0.441* (0.265)
  - Global FCI: -0.352 (0.282)
  - RCO: -0.186* (0.099)
  - Observations 678
- Two-year ahead (h=2), τ=20:
  - RCO: -0.311*** (0.091), Observations 642
- Three-year ahead (h=3), τ=20:
  - RCO: -0.278*** (0.062), Observations 604
- Coefficients vary across quantiles (τ=20, 50, 80) and horizons; negative RCO coefficients at lower quantiles indicate higher RCO is associated with worse downside growth outcomes.

### Robustness and alternative specifications (Table 5)
- Additional macro controls (Panel A) — RCO coefficients (h=1):
  - Q=20: -0.154 (0.109)
  - Q=50: -0.118 (0.086)
  - Q=80: 0.039 (0.103)
- Robustness to method/samples (Panel B):
  - MSS method (row a) h=1 Q=20: -0.206 (2.651)
  - Restricted sample (row b) h=1 Q=20: -0.235* (0.123)
  - RCO Synthetic EDF (row c) h=1 Q=20: -0.322*** (0.124)
- Results show the negative association between RCO and downside growth is robust across methods, samples, and alternative RCO measures, particularly at lower quantiles and multi-year horizons.

### Asset quality channel — bank loan portfolio riskiness (Table 6)
- Flow of Loan Loss Provisions (basis points):
  - Riskiness coefficients:
    - h=1: -0.705* (0.393)
    - h=2: -1.322*** (0.391)
    - h=3: -1.333*** (0.456)
  - Riskiness X High Loan Growth:
    - h=1: 0.777* (0.441)
    - h=2: 1.421*** (0.433)
    - h=3: 1.401*** (0.462)
- Change in Non-Performing Loans (basis points):
  - Riskiness:
    - h=1: -4.694*** (0.813)
    - h=2: -5.280*** (0.730)
    - h=3: -4.895*** (0.671)
  - Riskiness X High Loan Growth: positive and highly significant across h=1 to h=3.
- Leveraged Loans Share (percent):
  - Riskiness:
    - h=1: -0.725* (0.379)
    - h=2: -0.560* (0.298)
    - h=3: -0.494 (0.364)
  - Interaction terms indicate Riskiness combined with High Loan Growth increases loan portfolio risk measures.

### Controlling for riskiness of credit allocation in Growth-at-Risk (Table 7)
- When including RCA (riskiness of credit allocation) and RCO together for the 20th percentile:
  - RCA (where reported) and RCO both show negative coefficients at h=1, h=2, and h=3.
  - Example (h=1, column (2)): RCO -0.238** (0.104); RCA -0.453* (0.264).
  - Coefficients indicate both credit-origin risk (RCO) and allocation risk (RCA) contribute to downside growth risks.

### Sentiment channel: RCO and future lending standards and financial conditions (Tables 8–10)
- Table 8 (BLS — change in bank lending standards):
  - RCO effects on future lending standards:
    - h=1: 0.142* (0.072)
    - h=2: 0.115** (0.052)
    - h=3: 0.143* (0.073)
  - BLS (lagged) strongly predictive: e.g., column (2) BLS 0.406*** (0.066).
- Table 9 (FCI — financial conditions changes):
  - RCO effects on future changes in financial conditions:
    - h=1: -0.053** (0.025)
    - h=2: -0.024 (0.015)
    - h=3: -0.050** (0.021)
  - FCI itself strongly persistent: e.g., FCI coefficient -0.294*** (0.032) in one specification.
- Table 10 (including ΔFCI(t+1) in Growth-at-Risk):
  - ΔFCI (t+1) positive and significant for τ=20 across horizons:
    - h=1: 0.289** (0.117) when included alone; 0.299*** (0.110) in alternative column.
    - h=2: 0.799*** (0.110)
    - h=3: 0.540*** (0.109)
  - RCO retains negative coefficients in joint specifications; example h=2 column (3): RCO -0.311*** (0.089).

### Resilience channel: banking sector returns and loan growth (Tables 11–12)
- Table 11 (bank sector stock returns — quantile regressions):
  - Panel A (baseline), h=1, τ=10: RCO -4.788** (2.359)
  - Panel A, h=2, τ=10: RCO -3.346** (1.382)
  - Panel A, h=3, τ=10: RCO -2.798* (1.440)
  - Negative RCO coefficients among lower quantiles indicate worse future banking-sector returns associated with higher RCO.
- Table 12 (bank riskiness and future loan growth — local projections):
  - High Loan Growth effects:
    - h=1: 2.461*** (0.174)
    - h=2: 0.425** (0.167)
    - h=3: -0.118 (0.186)
    - h=4: -0.630*** (0.188)
  - Riskiness effects (baseline):
    - h=1: -0.224*** (0.035)
    - h=2: -0.107*** (0.037)
    - h=3: -0.069* (0.039)
    - h=4: -0.010 (0.038)
  - Interaction Riskiness X Crisis:
    - h=1: -0.445*** (0.084)
    - h=2: -0.247** (0.096)
    - h=3: -0.184** (0.092)
    - h=4: -0.244** (0.097)
  - Results indicate higher bank riskiness predicts weaker loan growth, amplified during crises.

### Controlling for banking-sector fragility indicators in Growth-at-Risk (Table 13)
- Average Riskiness (t+1) coefficients (significant negatives):
  - h=1: -0.004*** (0.001)
  - h=2: -0.006*** (0.001)
  - h=3: -0.005*** (0.001)
- RCO retains negative and statistically significant coefficients when included alongside leverage skewness and average riskiness:
  - Example h=2 column (5): RCO -0.302*** (0.089).
- Leverage Skewness (t+1) coefficients are less consistently significant but reported alongside other indicators.

### Figures and distributions
- Figure 1 (descriptive): Evolution of global distributions of the two RCO measures (z-score-based and Synthetic-EDF-based) and country-specific z-score RCO evolution for United States, Germany, Ireland, Spain — presented as authors’ calculations.

*Source: wpiea2024072-print-pdf - References (IMF Working Paper content unit).*

### Appendix 1. List of Countries Included in the Sample

### Appendix 1. List of Countries Included in the Sample

### Sample composition: countries and number of banks
- ARGENTINA 68
- AUSTRALIA 46
- AUSTRIA 69
- BELGIUM 44
- BRAZIL 78
- BULGARIA 22
- CANADA 59
- CHILE 30
- CHINA 112
- COLOMBIA 37
- CZECH REPUBLIC 28
- DENMARK 33
- FINLAND 12
- FRANCE 109
- GERMANY 227
- GREECE 17
- HUNGARY 29
- INDIA 65
- INDONESIA 108
- IRELAND 28
- ITALY 163
- JAPAN 223
- KOREA (SOUTH), REPUBLIC OF 28
- MALAYSIA 53
- MEXICO 47
- NETHERLANDS 48
- NEW ZEALAND 17
- NORWAY 51
- PERU 22
- PHILIPPINES 35
- POLAND 46
- PORTUGAL 51
- RUSSIAN FEDERATION 103
- SOUTH AFRICA 21
- SPAIN 83
- SWEDEN 29
- SWITZERLAND 238
- THAILAND 30
- TURKEY 30
- UNITED KINGDOM 90
- UNITED STATES 400
- VIETNAM 42

### Appendix 2 — Synthetic EDF estimation: methodology and interpretation
- Composite bank risk index constructed from balance-sheet fundamentals following CAMEL/CAEL approach:
  - Capital adequacy (principal component of total equity/total assets and z-score)
  - Loan loss provisions to total assets
  - Return on assets
  - Cost-to-income ratio
  - Liquidity (principal component of loans/assets, loans/deposits, liquid assets/total assets, liquid assets/deposits)
- Estimation specification (OLS with country-year fixed effects exploited to identify cross-sectional relations within country-year). All variables winsorized at 1 percent and standardized (mean zero, std dev 1) prior to estimation.
- Synthetic risk index constructed using estimated coefficients without country-year fixed effects; normalized to have mean zero within a country. Rationale: index captures cross-sectional differences in bank riskiness within country-year and allows construction for banks without EDF data while the empirical analysis controls for country-year fixed effects.
- Estimation note: LASSO OLS with EBIC selection was used; the method selected all variables so final specification is OLS.

### Appendix Table 2.1 — Bank Synthetic Risk Index (estimated coefficients)
- R R R R G G (capitalization) -0.230*** (0.0364)
- L L L L G G (provisions) 0.139*** (0.0301)
- R R R R R R (profitability) -0.0668* (0.0368)
- R R R F F (cost-to-income) 0.0847*** (0.0303)
- L L F F L L (liquidity) 0.157*** (0.0419)
- Observations 7595
- R2 0.700
- Adjusted R2 0.665

### Appendix 3 — Data appendix: key dataset construction steps and sample counts
- Bank-level dataset: FitchConnect full vintage downloads on June 11, 2018 and March 3, 2021; extensive cleaning steps (selection by market sector and specialization; drop missing total assets; keep annual statements; prioritize statement quality and audit status; match BvD and Fitch IDs; prioritize accounting standards; match global ultimate owner (GUO) and apply historical ownership rules; keep consolidated statements when available; drop subsidiaries with GUO in same country; identify mergers via Orbis M&A; tag extreme loan growth (>=200 percent) and accounting changes; merge with CreditEdge EDFs; add unique 2018 vintage bank-years to 2021 vintage).
- Exclusions for analysis: bank-years with change in consolidation level, large merger, accounting standards change, or missing loan growth.
- Baseline sample construction criteria:
  - Banks with total assets above 0.5 percent of the total assets of their country’s largest bank during at least one year and with at least five years of data.
  - For each country, keep years with at least ten banks meeting the above criteria; keep countries with at least five such years.
  - Keep only countries where a Financial Conditions Index (FCI) could be built.
- Resulting baseline sample counts:
  - 44,515 total bank-year observations
  - 39,730 bank-year observations with required information to construct the z-score
  - 3,071 banks
  - 42 countries
  - Coverage period 1990 to 2019
- Alternative (restricted) sample: country-years with at least 20 banks used for robustness checks.
- Loan-level dataset (Dealogic syndicated loans) construction:
  - Manual fuzzy matching of lenders to bank-level dataset; exclude missing allocation amounts or missing loan generation date; retain banks participating in at least 20 syndicated loan tranches in a year.
  - Final loan-level regression sample: 1676 bank-year observations from 1993 to 2019, covering 99 banks from 17 countries.
  - Average LLS in regression sample is 22.6 percent.
- Financial Conditions Indices (FCIs):
  - Estimated for 1990–2019 at quarterly frequency for 42 advanced and emerging market economies using up to eight price-based indicators: term spread; corporate spread; sovereign spread; interbank spread; first difference in real long-term rate; equity returns; equity volatility; house price returns.
  - FCIs estimated using Koop and Korobilis (2014) code to control for current macroeconomic conditions and allow dynamic interactions between FCIs and macroeconomic conditions.

*Appendix 1. List of Countries Included in the Sample — IMF Working Paper No. WP/2024/072*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024072-print-pdf.pdf_