## 3. Deviation from the Equilibrium Level Credit-to-GDP Ratio

## Source details

**Canonical URL:** [3. Deviation from the Equilibrium Level Credit-to-GDP Ratio](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026137-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026137-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026137-source-pdf.pdf.json)

---

### Long-run equilibrium framework and data
- Dependent variable:
  - domestic bank credit to the private non‑financial sector as a share of GDP (percent), at quarterly frequency.
- Key regressors:
  - Log GDP per capita (PPP), CPI inflation (year‑over‑year), real lending interest rate (average nominal lending rate minus inflation), Financial Institutions Index (FII, rescaled 0–100).
- Data sources and coverage:
  - Credit and GDP: BIS; IMF databases; national authorities when needed.
  - GDP per capita (PPP): World Bank.
  - Lending rates and inflation: central bank statistics.
  - FII: IMF Financial Institutions Strength index.
  - Quarterly data are interpolated where necessary. Data cover the period up until the first quarter of 2025 (2025Q1). The panel is unbalanced (e.g., Kyrgyz Republic from 1996Q1; Armenia from 2000Q1; Kazakhstan from 2003Q1; Azerbaijan from 2005Q1; Tajikistan from 2008Q1; Uzbekistan from 2019Q1).
- Time series properties:
  - Credit-to-GDP and log GDP per capita are I(1).
  - Inflation and the real lending rate are I(0).
  - FII is I(1) at annual frequency.
  - First differences of all variables are stationary.
- Estimation approach:
  - Panel cointegration framework using both a fixed effects (FE) panel OLS estimator and a Mean Group Estimator (MGE / ARDL error-correction representation) to allow for slope heterogeneity and to recover long-run equilibrium relationships without spurious regression.

### Panel estimation results (long-run coefficients)
- Large Economies — FE (Table 1 long-run estimates):
  - Ln (PPP): 9.47*** (3.31)
  - Inflation: -0.60* (0.31)
  - Real Rate: -1.05** (0.41)
  - FII: 0.25** (0.10)
  - Note: FE uses fixed effects with Driscoll–Kraay standard errors.
- Emerging Markets — MGE:
  - Ln (PPP): 8.91** (4.54)
  - Inflation: -1.68*** (0.35)
  - Real Rate: -0.59*** (0.16)
  - FII: 0.55* (0.31)
  - EC Term (average speed of adjustment, λ): -0.06** (0.32)
  - Note: MGE is the Mean Group Estimator from an ARDL error-correction model.
- Interpretations from coefficients:
  - Income per capita (Ln PPP) is the strongest positive structural driver of financial deepening.
  - Inflation has a significantly negative effect on credit-to-GDP.
  - Higher real lending rates are associated with lower credit-to-GDP.
  - Stronger Financial Institutions (higher FII) are associated with higher long-run credit-to-GDP.
  - Error-correction terms are negative and statistically significant, supporting cointegration between credit and fundamentals.

### Constructing equilibrium paths and country deviations
- Benchmarking approach:
  - Use long-run coefficients from the emerging-market MGE model to construct country-specific equilibrium credit paths for the Caucasus and Central Asia (CCA).
  - Equilibrium band: central path uses the median country-specific constant from the reference panel; upper and lower bounds use the largest and smallest constants respectively, producing a spectrum of plausible equilibrium levels conditional on fundamentals.
- Main country-level assessments (emerging-market benchmark):
  - Armenia and Georgia:
    - Actual credit-to-GDP increased rapidly in 2005–2021 and approached or exceeded equilibrium during the COVID-2019 pandemic (denominator effects).
    - Subsequent declines and partial recoveries in 2023–2024 indicate financial deepening approaching maturity given current income levels.
  - Kyrgyz Republic and Uzbekistan:
    - Gradual convergence. Kyrgyz actual credit-to-GDP lies between lower band and median benchmark (within equilibrium range).
    - Uzbekistan increased rapidly early in transformation but stabilized as GDP growth outpaced credit growth.
  - Kazakhstan and Azerbaijan:
    - Remain well below their equilibrium ranges, reflecting legacy scars from mid-2000s credit booms and banking stress (partial recoveries only).
    - Azerbaijan and Kazakhstan’s credit-to-GDP ratios are far below levels suggested by fundamentals after post‑shock adjustments.
  - Tajikistan:
    - Credit-to-GDP has been within the equilibrium range but with limited increase, indicating shallow banking intermediation.
- Overall equilibrium-region findings:
  - Most CCA countries remain under‑credited relative to fundamentals.
  - Convergence dynamics are uneven: Armenia and Georgia closest to equilibrium; Kazakhstan, Azerbaijan, and Tajikistan lagging; Kyrgyzstan and Uzbekistan in catch-up phases.
  - Different vulnerability types identified:
    - overheating risk in Armenia and Georgia as credit nears equilibrium;
    - weak intermediation/post‑crisis balance-sheet constraints in Kazakhstan and Azerbaijan;
    - limited intermediation in Tajikistan;
    - transitional risks in Uzbekistan (directed/preferential lending and rapid household lending).

### Statistical credit-gap measures (speed of credit growth)
- Filters used to capture cyclical deviations (parameters preserved exactly):
  - Hodrick–Prescott (two-sided) filter: smoothing parameter λ = 14,400 (implied cycle length ~10–15 years).
  - Hamilton (2018) regression filter: lookback horizon h = 20; lag order p = 4 (5-year horizon, AR(4)).
  - Christiano–Fitzgerald (CF) band-pass: cycle bounds low = 32; high = 120 (8–30 years).
  - Moving average (backward): window length = 20 quarters (5-year trailing average).
- Rationale for parameter choices:
  - Two-sided HP with λ = 14,400 targets medium-length credit cycles; chosen because the one-sided BIS convention (λ = 400,000) produced very persistent negative gaps in the CCA sample.
- Stylized results (2023–2025Q1):
  - South Caucasus:
    - Armenia: MA and HP gaps positive in late 2024–early 2025; other gaps hover below zero; five-year growth rate near zero → credit growth slowed sharply from post-pandemic peak and aligns broadly with trend.
    - Georgia: uniformly negative gaps and near-zero five-year growth rate → mature credit cycle.
  - Azerbaijan and Kazakhstan: gaps clustered around zero with some filters slightly positive; moderate recoveries but not indicative of excessive expansion.
  - Kyrgyz Republic and Tajikistan: early-stage upswing in some filters (MA and HP) with five-year growth rates turning modestly positive; signals of gradual recovery rather than booms.
  - Uzbekistan: moving-average and CF slightly positive, Hamilton around zero, HP negative; five-year growth rate fell markedly → post-liberalization boom cooled toward a soft landing.

### Disaggregation: household vs. corporate credit
- Household credit findings:
  - Generally near or slightly above trend across most economies.
  - Armenia: household gaps positive but contained; five-year growth moderate.
  - Azerbaijan: household lending expanding above trend; five-year growth increasing.
  - Georgia: household credit plateaued; mixed filter signals; five-year growth negative (deceleration).
  - Kazakhstan: all household gaps positive; sustained above‑trend growth; vulnerabilities from policy/subsidized lending risk.
  - Kyrgyz Republic: household gaps negative and declining; five-year growth down.
  - Tajikistan: household gaps trending upward and positive; early expansion from low base.
  - Uzbekistan: household credit aligned with trend; five-year growth positive but declining; rapid microlending growth noted.
- Corporate credit findings:
  - Generally subdued recovery with variation across countries:
    - Armenia: corporate gaps negative or close to zero; narrowing gaps recently.
    - Azerbaijan: early-stage recovery; mixed filter signals; five-year growth improved since 2020.
    - Georgia: gaps at or below zero; decelerated five-year growth.
    - Kazakhstan: marginally negative gaps for most filters; nascent recovery.
    - Kyrgyz Republic: marginally positive gaps; five-year growth around 30 percent.
    - Tajikistan: corporate cycle subdued; HP marginally positive; five-year growth improved but negative overall.
    - Uzbekistan: limited history; filters slightly negative; five-year growth flat.

### Multivariate filter (MVF / MVK) approach and Kazakhstan case
- Motivation:
  - Combine cyclical identification with macroeconomic conditioning to distinguish financial deepening from cyclical excess while avoiding pitfalls of purely univariate filters.
- MVF structure (illustrated for Kazakhstan):
  - IS curve: output gap depends on lags/leads, credit gap, real-rate gap, external conditions.
  - Phillips curve: inflation depends on expected and lagged inflation, output gap, relative-price gap.
  - Okun’s law linking output gap and unemployment gap.
  - Credit equation: credit gap responds to lagged output, own persistence, and the real rate gap.
  - External block (Russia) provides foreign demand, inflation, and interest-rate conditions for Kazakhstan.
- Kazakhstan-specific MVF notes:
  - Estimation sample: 2003:Q1–2024:Q2.
  - Model uses quarterly domestic and Russian variables listed in the framework (exact variable names preserved: KZ_RGDP, KZ_NGDP, KZ_CPI, KZ_UNR, KZ_STN, KZ_LTN, KZ_CREDIT, RU_RGDP_SA, RU_CPI, RU_STN, RU_LTN).
  - Calibrated steady-state values include:
    - gss — 4.0
    - gss E — 2.0
    - UNRss — 6.5
    - πss — 4.0
    - Rss E — 1.5
    - φss — 0.0
    - gss c — 3.0
  - Calibrated structural parameters include:
    - τ — 0.08
    - ρgc — 1.0
    - τE — 0.08
  - Estimation uses Bayesian priors and prior-only estimation with system priors; system prior: half-life of the business cycle centered at 0.5 (approximately 8 quarters / 2 years) with standard deviation 0.05; active-set algorithm tolerance = 10−7.
- Comparison outcome for Kazakhstan:
  - MVF produces a smoother and more plausible estimate of the credit trend than one-sided and two-sided HP filters, with the MVF trend peaking earlier and declining more gradually after the 2007–08 boom.
  - MVF trend estimates include a relatively narrow confidence band (MVF trend ± two standard deviations of the estimated credit cycle).
  - Note: MVF estimates real credit trend (nominal credit deflated by GDP deflator) while earlier statistical filters estimate trends in credit-to-GDP ratios.
- Shock standard deviations (prior modes preserved exactly) include notable domestic priors:
  - σY — 0.25
  - σG — 0.05
  - σLGDP_BAR — 0.05
  - σPIE — 0.25
  - σC_GAP — 0.25
  - σC_BAR — 0.05
  - σGC — 0.06
  - σUNR_GAP — 0.25
  - σUNR_BAR — 0.05
  - σG_UNR_BAR — 0.05
  - σPHI — 0.25
  - σPI — 0.15
  - Real exchange rate shocks:
    - σZ_GAP — 0.60
    - σZ_BAR — 0.60

### Predictive performance (“horse race”) of gap measures
- Objective:
  - Evaluate which credit gap measures best predict systemic banking stress using AUC (area under the ROC curve) per Drehmann and Yetman (2020).
- Methodology highlights:
  - Event window: identify systemic banking stress episodes (12‑quarter horizons preceding significant slowdowns or stress); exclude signals during ongoing stress episodes.
  - Rolling forecast evaluation (expanding-sample) to mimic real-time policymaking; bootstrap resampling for standard errors and Wald tests for differences.
  - Crisis definition: use Laeven and Valencia (2018) taxonomy supplemented with IMF Article IV reports, FSSAs, and contemporaneous policy documents for the CCA.
- Identified systemic stress episodes used in evaluation:
  - Azerbaijan: 2015Q1–2017Q4
  - Kazakhstan: 2008Q3–2010Q4 and 2014Q4–2017Q4
  - Kyrgyz Republic: 2010Q2–2010Q4
  - Tajikistan: 2015Q4–2016Q4
- Key empirical findings:
  - Credit gaps display significant predictive content at short-to-medium horizons (up to eight quarters before a crisis) though performance varies by country and filter.
  - Azerbaijan and Tajikistan: CF and MA filters yielded AUC profiles close to unity up to six–eight quarters before crises (likely reflecting sharp pre-crisis credit accelerations but also small-sample bias).
  - Kazakhstan: Hamilton and MVK filters achieved stable high AUCs (around 0.9) over 4–8 quarters; MA and HP gaps weaker.
  - Kyrgyzstan: Hamilton and one-sided HP filters dominated, with AUCs approaching 1 at horizons of 6–8 quarters.
  - Caveat: high AUCs in many cases reflect very small numbers of crisis episodes per country; limited crisis observations constrain the generalizability of out‑of‑sample predictive dominance tests.
- Policy implication from horse race:
  - Credit gap indicators can be informative early‑warning components within macroprudential surveillance but should not be treated as mechanical policy triggers.
  - Persistent/widening positive gaps across multiple filters are more informative than single-quarter deviations.
  - Pooling across countries or panel-based estimation can increase robustness of predictive assessments.

### Filter parameters and technical settings (exact values)
- Hodrick–Prescott (two-sided): smoothing parameter λ = 14,400 (implied cycle length ~10–15 years).
- Hamilton (2018): lookback horizon h = 20; lag order p = 4 (5-year horizon, AR(4)).
- Christiano–Fitzgerald: cycle bounds (quarters) low = 32; high = 120 (8–30 years).
- Moving average (backward): window length = 20 quarters (5-year trailing average).
- Note on BIS convention:
  - one-sided HP with λ = 400,000 targets ~30-year financial cycles; authors did not adopt this for main CCA analysis.

### Banking sector crisis episodes (Annex III country summaries)
- Azerbaijan (2015–2017):
  - Trigger: 2014 oil-price collapse causing two manat devaluations in 2015.
  - Effects: sharply raised FX credit risk, dollarization, and pressure on bank capital and liquidity.
  - Policy actions: creation of a dedicated financial regulator (FIMSA); sectoral restructuring; revocation of licenses of weaker banks; overhaul of the International Bank of Azerbaijan (IBA).
  - 2017: IBA halted foreign-debt payments and entered court-sanctioned restructuring under Azeri law; recognition proceedings in English courts document the process and timing.
  - Evaluation window dated: 2015Q1–2017Q4.
  - IMF characterization: economy recovering “from a banking crisis and recession.”
- Kazakhstan (2008–2010; 2014–2017):
  - First wave (2008Q3–2010Q4): pre-GFC wholesale-funded credit boom; funding dried up; multiple large banks required intervention; NPLs surged; system-wide banking shock requiring comprehensive resolution and recapitalization.
  - Second wave (2014Q4–2017Q4): oil-price weakness and 2015 move to a free float produced large tenge depreciation revealing legacy asset-quality problems; cleanup culminated in the 2017 rescue and merger of Kazkommertsbank (KKB) with Halyk; staff reports noted substantial expected public support and sector consolidation.
- Kyrgyz Republic (2010):
  - Event: domestic political upheaval in April–June 2010 triggered rapid deposit outflows and system-wide liquidity stress.
  - Measures: temporary administration in key banks; nationalization of Asia Universal Bank (AUB) and placement under conservatorship.
  - IMF note: 30 percent deposit decline within April 2010 and supervisory measures across the system.
  - Evaluation window dated: 2010Q2–2010Q4 as the core crisis phase.
- Tajikistan (2015–2016):
  - Underlying weaknesses: directed lending, governance problems, rising loan losses.
  - Resolution: late-2016 public recapitalization of two large banks amounting to about 6 percent of GDP.
  - Evaluation window dated: 2015Q4–2016Q4 capturing run-up and immediate resolution.
- Armenia, Georgia, Uzbekistan:
  - LV report: no systemic banking crises in these countries through 2017.
  - Observations: periods of rapid credit growth (notably Georgia in 2017–2021) and high dollarization, but macroprudential tightening and supervisory measures helped avoid generalized bank failures or large fiscal recapitalizations.
  - Treatment in analysis: these countries are treated as non-crisis cases for horizon-specific evaluation; their histories are used to define tranquil observations and to document vulnerability build-ups that did not crystallize into systemic breaks.

### Conclusions and policy implications summarized
- Structural vs cyclical synthesis:
  - Equilibrium benchmarks (emerging-market MGE) indicate most CCA countries remain under‑credited relative to fundamentals.
  - Statistical and MVK-based credit gaps indicate a shift from post‑crisis deleveraging toward moderate recovery, with limited evidence of broad overheating.
  - Household credit is the main driver of recent credit growth; corporate lending remains weaker in many countries.
- Key policy messages:
  - Use a dual framework: combine fundamentals-based equilibrium benchmarks (level) with multiple cyclical gap measures (momentum), and complement with multivariate filtering to condition trends on macro variables.
  - Treat credit gap indicators as monitoring tools within a broader macroprudential toolkit rather than as standalone policy triggers, due to sensitivity to filter choice, endpoint bias, and small-sample limitations.
  - Strengthen supervisory scrutiny—especially where household credit is expanding above trend—to preserve lending standards and guard against overheating in specific segments.
  - Continue institutional strengthening (banking sector repair, financial infrastructure) to support safe financial deepening in countries below equilibrium (e.g., Kazakhstan, Azerbaijan, Tajikistan).

*Source: IMF staff estimates and analysis as presented in "3. Deviation from the Equilibrium Level Credit-to-GDP Ratio" (IMF Working Paper).*

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

### References

### Annexes (as listed)
- Annex I. Equilibrium (Large Economies) ........................................................................................................ 35
- Annex II: Multivariate Filter Estimation for Kazakhstan ................................................................................ 37
- Annex III. Banking Sector Crisis Episodes ..................................................................................................... 44

### Figures (as listed)
- Figure 1. Evolution of Total, Household, and Corporate Credit Across CCA Economies ................................................ 7
- Figure 2. Credit Decomposition by Currency of Denomination ........................................................................................ 8
- Figure 3. Deviation from the Equilibrium Level Credit -to-GDP Ratio ............................................................................ 16
- Figure 4. Credit Gap Measures for the CCA Economies ............................................................................................... 19
- Figure 5. Household Credit Gap Measures for the CCA Economies ............................................................................. 23
- Figure 6. Corporate Credit Gap Measures for the CCA Economies .............................................................................. 24
- Figure 7. Comparison of MVK and Other Gap measures for the CCA Economies ........................................................ 26
- Figure 8. AUC by Horizon .............................................................................................................................................. 30

### Pagination and Publication indicator
- References page number: 32
- IMF WORKING PAPERS CCA Credit Expansion
- INTERNATIONAL MONETARY FUND
- Page indicator: 3

*Source: wpiea2026137-source-pdf - References (canonical URL: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026137-source-pdf.pdf)*

### 3. Deviation from the Equilibrium Level Credit -to-GDP Ratio ...........................................................

### 3. Deviation from the Equilibrium Level Credit-to-GDP Ratio

### Long-run equilibrium framework and data
- Dependent variable: domestic bank credit to the private non‑financial sector as a share of GDP (percent), at quarterly frequency.  
- Key regressors: Log GDP per capita (PPP), CPI inflation (year-over-year), real lending interest rate (average nominal lending rate minus inflation), Financial Institutions Index (FII, rescaled 0–100).  
- Data sources and coverage:
  - Credit and GDP: BIS; IMF databases; national authorities when needed.
  - GDP per capita (PPP): World Bank.
  - Lending rates and inflation: central bank statistics.
  - FII: IMF Financial Institutions Strength index.
  - Quarterly data are interpolated where necessary. Data cover the period up until the first quarter of 2025 (2025Q1). The panel is unbalanced (e.g., Kyrgyz Republic from 1996Q1; Armenia from 2000Q1; Kazakhstan from 2003Q1; Azerbaijan from 2005Q1; Tajikistan from 2008Q1; Uzbekistan from 2019Q1).
- Time series properties:
  - Credit-to-GDP and log GDP per capita are I(1).
  - Inflation and the real lending rate are I(0).
  - FII is I(1) at annual frequency.
  - First differences of all variables are stationary.
- Estimation approach:
  - Panel cointegration framework using both a fixed effects (FE) panel OLS estimator and a Mean Group Estimator (MGE / ARDL error-correction representation) to allow for slope heterogeneity and to recover long-run equilibrium relationships without spurious regression.

### Panel estimation results (long-run coefficients)
- Table 1 (long-run estimates):
  - Large Economies — FE:
    - Ln (PPP): 9.47*** (3.31)
    - Inflation: -0.60* (0.31)
    - Real Rate: -1.05** (0.41)
    - FII: 0.25** (0.10)
    - Note: FE uses fixed effects with Driscoll–Kraay standard errors.
  - Emerging Markets — MGE:
    - Ln (PPP): 8.91** (4.54)
    - Inflation: -1.68*** (0.35)
    - Real Rate: -0.59*** (0.16)
    - FII: 0.55* (0.31)
    - EC Term (average speed of adjustment, λ): -0.06** (0.32)
    - Note: MGE is the Mean Group Estimator from an ARDL error-correction model.
- Interpretations from coefficients:
  - Income per capita (Ln PPP) is the strongest positive structural driver of financial deepening.
  - Inflation has a significantly negative effect on credit-to-GDP.
  - Higher real lending rates are associated with lower credit-to-GDP.
  - Stronger Financial Institutions (higher FII) are associated with higher long-run credit-to-GDP.
  - Error-correction terms are negative and statistically significant, supporting cointegration between credit and fundamentals.

### Constructing equilibrium paths and deviations for CCA countries
- Benchmarking approach:
  - Use long-run coefficients from the emerging-market MGE model to construct country-specific equilibrium credit paths for the Caucasus and Central Asia (CCA).
  - Equilibrium band: central path uses the median country-specific constant from the reference panel; upper and lower bounds use the largest and smallest constants respectively, producing a spectrum of plausible equilibrium levels conditional on fundamentals.
- Main country-level assessments (based on emerging-market benchmark):
  - Armenia and Georgia: actual credit-to-GDP increased rapidly in 2005–2021 and approached or exceeded equilibrium during the COVID-2019 pandemic (denominator effects). Subsequent declines and partial recoveries in 2023–2024 indicate financial deepening approaching maturity given current income levels.
  - Kyrgyz Republic and Uzbekistan: gradual convergence. Kyrgyz actual credit-to-GDP lies between lower band and median benchmark (within equilibrium range). Uzbekistan increased rapidly early in transformation but stabilized as GDP growth outpaced credit growth.
  - Kazakhstan and Azerbaijan: remain well below their equilibrium ranges, reflecting legacy scars from mid-2000s credit booms and banking stress (partial recoveries only). Azerbaijan and Kazakhstan’s credit-to-GDP ratios are far below levels suggested by fundamentals after post‑shock adjustments.
  - Tajikistan: credit-to-GDP has been within the equilibrium range but with limited increase, indicating shallow banking intermediation.
- Overall equilibrium-region findings:
  - Most CCA countries remain under‑credited relative to fundamentals.
  - Convergence dynamics are uneven: Armenia and Georgia closest to equilibrium; Kazakhstan, Azerbaijan, and Tajikistan lagging; Kyrgyzstan and Uzbekistan in catch-up phases.
  - Different vulnerability types: overheating risk in Armenia and Georgia as credit nears equilibrium; weak intermediation/post‑crisis balance-sheet constraints in Kazakhstan and Azerbaijan; limited intermediation in Tajikistan; transitional risks in Uzbekistan (directed/preferential lending and rapid household lending).

### Statistical credit-gap measures (speed of credit growth)
- Purpose: capture cyclical deviations (momentum) of credit-to-GDP from trend using four univariate filters plus a moving-average benchmark:
  - Hodrick–Prescott (two-sided) filter: smoothing parameter λ = 14,400 (implied cycle length ~10–15 years).
  - Hamilton (2018) regression filter: lookback horizon h = 20; lag order p = 4 (5-year horizon, AR(4)).
  - Christiano–Fitzgerald (CF) band-pass: cycle bounds low = 32; high = 120 (8–30 years).
  - Moving average (backward): window length = 20 quarters (5-year trailing average).
- Rationale for parameter choices:
  - Two-sided HP with λ = 14,400 targets medium-length credit cycles; chosen because the one-sided BIS convention (λ = 400,000) produced very persistent negative gaps in the CCA sample.
- Stylized results (2023–2025Q1):
  - South Caucasus:
    - Armenia: MA and HP gaps positive in late 2024–early 2025; other gaps hover below zero; five-year growth rate near zero → credit growth slowed sharply from post-pandemic peak and aligns broadly with trend.
    - Georgia: uniformly negative gaps and near-zero five-year growth rate → mature credit cycle.
  - Azerbaijan and Kazakhstan: gaps clustered around zero with some filters slightly positive; moderate recoveries but not indicative of excessive expansion.
  - Kyrgyz Republic and Tajikistan: early-stage upswing in some filters (MA and HP) with five-year growth rates turning modestly positive; signals of gradual recovery rather than booms.
  - Uzbekistan: moving-average and CF slightly positive, Hamilton around zero, HP negative; five-year growth rate fell markedly → post-liberalization boom cooled toward a soft landing.

### Disaggregation: household vs. corporate credit
- Household credit:
  - Generally near or slightly above trend across most economies.
  - Armenia: household gaps positive but contained; five-year growth moderate.
  - Azerbaijan: household lending expanding above trend; five-year growth increasing.
  - Georgia: household credit plateaued; mixed filter signals; five-year growth negative (deceleration).
  - Kazakhstan: all household gaps positive; sustained above‑trend growth; vulnerabilities from policy/subsidized lending risk.
  - Kyrgyz Republic: household gaps negative and declining; five-year growth down.
  - Tajikistan: household gaps trending upward and positive; early expansion from low base.
  - Uzbekistan: household credit aligned with trend; five-year growth positive but declining; rapid microlending growth noted.
- Corporate credit:
  - Generally subdued recovery with variation:
    - Armenia: corporate gaps negative or close to zero; narrowing gaps recently.
    - Azerbaijan: early-stage recovery; mixed filter signals; five-year growth improved since 2020.
    - Georgia: gaps at or below zero; decelerated five-year growth.
    - Kazakhstan: marginally negative gaps for most filters; nascent recovery.
    - Kyrgyz Republic: marginally positive gaps; five-year growth around 30 percent.
    - Tajikistan: corporate cycle subdued; HP marginally positive; five-year growth improved but negative overall.
    - Uzbekistan: limited history; filters slightly negative; five-year growth flat.

### Multivariate filter (MVF / MVK) approach
- Motivation: combine cyclical identification with macroeconomic conditioning to distinguish financial deepening from cyclical excess while avoiding the pitfalls of purely univariate filters.
- MVF structure (illustrated for Kazakhstan):
  - IS curve: output gap depends on lags/leads, credit gap, real-rate gap, external conditions.
  - Phillips curve: inflation depends on expected and lagged inflation, output gap, relative-price gap.
  - Okun’s law linking output gap and unemployment gap.
  - Credit equation: credit gap responds to lagged output, own persistence, and the real rate gap.
  - External block (e.g., Russia) provides foreign demand, inflation, and interest-rate conditions for Kazakhstan.
- Comparison outcome:
  - For Kazakhstan, the MVF produces a smoother and more plausible estimate of the credit trend than one-sided and two-sided HP filters, with the MVF trend peaking earlier and declining more gradually after the 2007–08 boom. MVF trend estimates include a relatively narrow confidence band (MVF trend ± two standard deviations of the estimated credit cycle).
  - Note: MVF estimates real credit trend (nominal credit deflated by GDP deflator) while statistical filters in the earlier analysis estimate trends in credit-to-GDP ratios.

### Predictive performance (“horse race”) of gap measures
- Objective: evaluate which credit gap measures best predict systemic banking stress using AUC (area under the ROC curve) per Drehmann and Yetman (2020).
- Methodology highlights:
  - Event window: identify systemic banking stress episodes (12‑quarter horizons preceding significant slowdowns or stress); exclude signals during ongoing stress episodes.
  - Rolling forecast evaluation (expanding-sample) to mimic real-time policymaking; bootstrap resampling for standard errors and Wald tests for differences.
  - Crisis definition: use Laeven and Valencia (2018) taxonomy supplemented with IMF Article IV reports, FSSAs, and contemporaneous policy documents for the CCA.
- Identified systemic stress episodes used in evaluation:
  - Azerbaijan: 2015Q1–2017Q4
  - Kazakhstan: 2008Q3–2010Q4 and 2014Q4–2017Q4
  - Kyrgyz Republic: 2010Q2–2010Q4
  - Tajikistan: 2015Q4–2016Q4
- Key empirical findings:
  - Credit gaps display significant predictive content at short-to-medium horizons (up to eight quarters before a crisis) though performance varies by country and filter.
  - Azerbaijan and Tajikistan: CF and MA filters yielded AUC profiles close to unity up to six–eight quarters before crises (likely reflecting sharp pre-crisis credit accelerations but also small-sample bias).
  - Kazakhstan: Hamilton and MVK filters achieved stable high AUCs (around 0.9) over 4–8 quarters; MA and HP gaps weaker.
  - Kyrgyzstan: Hamilton and one-sided HP filters dominated, with AUCs approaching 1 at horizons of 6–8 quarters.
  - Caveat: high AUCs in many cases reflect very small numbers of crisis episodes per country; limited crisis observations constrain the generalizability of out‑of‑sample predictive dominance tests.
- Policy implication from horse race:
  - Credit gap indicators can be informative early‑warning components within macroprudential surveillance but should not be treated as mechanical policy triggers. Persistent/widening positive gaps across multiple filters are more informative than single-quarter deviations.
  - Pooling across countries or panel-based estimation can increase robustness of predictive assessments.

### Filter parameters and technical settings (exact values)
- Hodrick–Prescott (two-sided): smoothing parameter λ = 14,400 (implied cycle length ~10–15 years).
- Hamilton (2018): lookback horizon h = 20; lag order p = 4 (5-year horizon, AR(4)).
- Christiano–Fitzgerald: cycle bounds (quarters) low = 32; high = 120 (8–30 years).
- Moving average (backward): window length = 20 quarters (5-year trailing average).
- Note on BIS convention: one-sided HP with λ = 400,000 targets ~30-year financial cycles; authors did not adopt this for main CCA analysis.

### Conclusions and policy implications summarized
- Structural (equilibrium) vs cyclical (gap) views:
  - Equilibrium benchmarks (emerging-market MGE) indicate most CCA countries remain under‑credited relative to fundamentals.
  - Statistical and MVK-based credit gaps indicate a shift from post‑crisis deleveraging toward moderate recovery, with limited evidence of broad overheating.
  - Household credit is the main driver of recent credit growth; corporate lending remains weaker in many countries.
- Key policy messages:
  - Use a dual framework: combine fundamentals-based equilibrium benchmarks (level) with multiple cyclical gap measures (momentum), and complement with multivariate filtering to condition trends on macro variables.
  - Treat credit gap indicators as monitoring tools within a broader macroprudential toolkit rather than as standalone policy triggers, due to sensitivity to filter choice, endpoint bias, and small-sample limitations.
  - Strengthen supervisory scrutiny—especially where household credit is expanding above trend—to preserve lending standards and guard against overheating in specific segments.
  - Continue institutional strengthening (banking sector repair, financial infrastructure) to support safe financial deepening in countries below equilibrium (e.g., Kazakhstan, Azerbaijan, Tajikistan).

*Source: IMF staff estimates and analysis as presented in "3. Deviation from the Equilibrium Level Credit-to-GDP Ratio" (IMF Working Paper).*

### References

### References

### Bibliographic references
- List of cited works including working papers, journal articles, books, and court decisions (authors and titles as provided). Notable entries referenced in the text include:
  - Andrle, M., Garcia-Saltos, R., & Ho, G. (2014). "A Model-Based Analysis of Spillovers: The Case of Poland and the Euro Area," IMF Working Paper No. 14/186.
  - Baba, C., Dell’Erba, S., Detragiache, E., Harrison, O., Mineshima, A., Musayev, A., and Shahmoradi, A. (2020). “How Should Credit Gaps Be Measured? An Application to European Countries,” IMF Working Paper No. WP/20/6.
  - Borio, C., and Lowe, P. (2002). "Assessing the risk of banking crises." BIS Quarterly Review, December, 43–54.
  - Drehmann, M., Borio, C., & Tsatsaronis, K. (2011). "Anchoring Countercyclical Capital Buffers: The Role of Credit Aggregates," BIS Working Paper No. 355.
  - Hamilton, J. D. (2018). "Why You Should Never Use the Hodrick-Prescott Filter." Review of Economics and Statistics, 100(5), 831–843.
  - Hodrick, R. J., and Prescott, E. C. (1997). "Postwar U.S. Business Cycles: An Empirical Investigation." Journal of Money, Credit and Banking, 29(1), 1–16.
  - Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., and Shin, Y. (1992). "Testing the Null Hypothesis of Stationarity against the Alternative of a Unit Root..."
  - Laeven, L., & Valencia, F. (2018). "Systemic Banking Crises Revisited," IMF Working Paper No. 18/206.
  - Maddala, G. S., and Kim, I.-M. (1998). Unit Roots, Cointegration, and Structural Change. Cambridge University Press.
  - Pesaran, M. H., Shin, Y., & Smith, R. P. (1999). "Pooled Mean Group Estimation of Dynamic Heterogeneous Panels," Journal of the American Statistical Association, 94(446), 621–634.
- IMF country reports and press releases cited (selected):
  - International Monetary Fund (2010a). Republic of Kazakhstan: 2010 Article IV Consultation Staff Report. IMF Country Report No. 10/241, July 27, 2010.
  - International Monetary Fund (2010b). Public Information Notice: IMF Executive Board Concludes 2010 Article IV Consultation with the Republic of Kazakhstan. July 27, 2010.
  - International Monetary Fund (2011a). Kyrgyz Republic: Selected Issues. IMF Country Report No. 11/156, June 2, 2011 (notes 30% deposit decline in April 2010 and measures around AUB).
  - International Monetary Fund (2017d). IMF Executive Board Concludes 2017 Article IV Consultation with the Republic of Tajikistan. Press Release No. 17/429, November 9, 2017 (records 6.1% of GDP bank recapitalization in 2016).
  - International Monetary Fund (2019). Republic of Azerbaijan: 2019 Article IV Consultation Press Release and Staff Report. IMF Country Report No. 19/241, September 18, 2019.
  - International Monetary Fund (2021). Republic of Tajikistan: 2017 Article IV Consultation Staff Report (eLibrary re-posting with DSA discussion). September 3, 2021.

### Annex I (Figures)
- Figure 1. Deviation from the Equilibrium Level Credit-to-GDP Ratio, Large Economies reference panel (In percent of credit to GDP).
- Figure 2. Deviation from the Equilibrium Level Credit-to-GDP Ratio, Large Economies Reference panel (without FII).

### Annex II — Multivariate Filter Estimation for Kazakhstan: model overview and implementation
- Estimation sample: 2003:Q1–2024:Q2.
- Framework: MVF model following Baba et al. (2019), building on Carabenciov et al. (2008), Andrle et al. (2014), and Krznar and Matheson (2017).
- Notation:
  - Lower-case letters denote deviations from trends (gaps); lower-case with * or BAR denote trends; upper-case denote observed levels.
  - Subscript "E" refers to the foreign economy (Russia).

- I. Domestic economy (Kazakhstan)
  - A. Output
    - Output gap: yt = Yt − Yt* where Yt ≡ 100 × ln(RGDPt).
    - Potential output law of motion: Yt* = Yt−1* + 1/4 gt + εt_LGD P_BAR.
    - Trend growth: gt = ρg ⋅ gt−1 + (1−ρg) ⋅ gss + εt_G.
    - IS specification: yt = βlead ⋅ yt+1 + βlag ⋅ yt−1 + βr ⋅ rt + βyE ⋅ ytE + βc ⋅ εt c + βz ⋅ zt + εt Y.
    - rt is real interest rate gap; ytE is Russian output gap; εt c is the credit shock; zt is the real exchange rate gap (Russia–Kazakhstan inflation differential).
  - B. Phillips curve
    - πt = (1−αlag)⋅πt+1 + αlag⋅πt−1 + αy⋅yt + αz⋅zt + εt_PI E.
  - C. Credit
    - Credit gap: ct = Ct − Ct* where Ct ≡ 100×ln(CREDt/DEFt).
    - Credit-gap equation: ct = θ1⋅yt−1 + θ2⋅ct−1 + θ3⋅rt−1 + εt_C_GAP.
    - Trend credit: Ct* = Ct−1* + 1/4 gtc + εt_C_BAR.
    - Trend credit growth: gtc = ρgc⋅gt−1 c + (1−ρgc)⋅gss c + εt_GC.
  - D. Okun's Law
    - Unemployment gap: ut = UNRt − UNRt*.
    - Dynamic Okun: ut = τ2⋅ut−1 + (1−τ2)⋅(−τ1)⋅yt + εt_UNR_GAP.
    - NAIRU evolution: UNRt* = (1−τ3)⋅UNRss + τ3⋅(UNRt−1* + gt u) + εt_UNR_BAR.
    - gt u = (1−τ4)⋅gt−1 u + εt_G_UNR_BAR.

- II. Foreign economy (Russia)
  - A. Output
    - Russian output gap: ytE = βleadE⋅yt+1E + βlagE⋅yt−1E + βrE⋅rtE + βy⋅yt + εt_Y_E.
    - Feedback from Kazakhstan to Russia: βy = 0.10 (deviation from small-open-economy assumption).
    - Russia potential output: YtE* = Yt−1E* + 1/4 gtE + εt_LGD P_BAR_E.
    - gtE = ρgE⋅gt−1E + (1−ρgE)⋅gss E + εt_G_E.
  - B. Phillips curve
    - πtE = (1−αlagE)⋅πt+1E + αlagE⋅πt−1E + αyE⋅ytE + εt_PIE_E.
  - C. Monetary policy rule (Russia)
    - itE = γlagE⋅it−1E + (1−γlagE)[rtE* + πttar + γinflE⋅(πtE − πttar) + γyE⋅ytE] + εt_INT_E.
  - D. Real interest rate (Russia)
    - RtE = itE − πt+1E.
    - rtE = RtE − rtE*.
    - Natural real rate: rtE* = ρrE⋅rt−1E* + (1−ρrE)⋅Rss E + εt_R_BAR_E.

- III. Risk premium and real exchange rate
  - Risk premium (φt, spread of Kazakhstan long-term rates over Russian long-term rates): φt = ρφ⋅φt−1 + (1−ρφ)⋅φss + εt_PHI (AR(1)).
  - Real exchange rate gap (z gap): zt = ρgz⋅zt−1 + εt_Z_GAP.
  - Real exchange rate trend: Zt* = ρz*⋅Zt−1* + (1−ρz*)⋅Zss + εt_Z_BAR.

- IV. Data
  - Quarterly data for Kazakhstan (domestic) and Russia (foreign) over 2003:Q1–2024:Q2.
  - Variables and sources (as listed):
    - KZ_RGDP Real GDP, Kazakhstan — National authorities
    - KZ_NGDP Nominal GDP, Kazakhstan — National authorities
    - KZ_CPI Consumer Price Index, Kazakhstan — National authorities
    - KZ_UNR Unemployment rate, Kazakhstan — National authorities
    - KZ_STN Short-term nominal interest rate, Kazakhstan — National authorities
    - KZ_LTN Long-term nominal interest rate, Kazakhstan — National authorities
    - KZ_CREDIT Credit to private non-financial sector, Kazakhstan — National authorities
    - RU_RGDP_SA Real GDP (seasonally adjusted), Russia — National authorities
    - RU_CPI Consumer Price Index, Russia — National authorities
    - RU_STN Short-term nominal interest rate, Russia — Central Bank of Russia
    - RU_LTN Long-term nominal interest rate, Russia — National authorities

- V. Calibrated parameters (fixed values)
  - Steady-state values (parameter — value — description):
    - gss — 4.0 — Average annual real GDP growth, Kazakhstan
    - gss E — 2.0 — Average annual real GDP growth, Russia
    - UNRss — 6.5 — Average unemployment rate, Kazakhstan
    - πss — 4.0 — Medium-term inflation target (KZ/RU)
    - Rss E — 1.5 — Long-run equilibrium real interest rate, Russia
    - φss — 0.0 — Long-run risk premium
    - gss c — 3.0 — Steady-state trend credit growth (set equal to GDP trend growth)
  - Calibrated structural parameters:
    - τ — 0.08 — Trend-growth adjustment in IS curve
    - ρgc — 1.0 — Persistence of trend credit growth (random walk)
    - τE — 0.08 — Trend-growth adjustment, Russia

- VI. Estimated parameters and prior distributions (Bayesian, prior-only estimation with system priors)
  - System prior: half-life of the business cycle centered at 0.5 (approximately 8 quarters / 2 years) with standard deviation 0.05.
  - Estimation algorithm: active-set algorithm with tolerance parameters set to 10−7.
  - Selected priors and prior modes (examples preserved exactly):
    - Domestic Output (IS Curve) — Kazakhstan:
      - βlead: Beta prior, Prior Mode 0.15, Prior Std. Dev. 0.05, Bounds [0.001, 0.99]
      - βlag: Beta prior, Prior Mode 0.65, Prior Std. Dev. 0.10, Bounds [0.001, 0.99]
      - βr: Gamma prior, Prior Mode 0.08, Prior Std. Dev. 0.03, Bounds [0.001, ∞)
      - βyE: Gamma prior, Prior Mode 0.05, Prior Std. Dev. 0.03, Bounds [0.001, ∞)
      - βc: Gamma prior, Prior Mode 0.15, Prior Std. Dev. 0.05, Bounds [0.001, ∞)
      - βz: Gamma prior, Prior Mode 0.05, Prior Std. Dev. 0.02, Bounds [0.001, ∞)
    - Domestic Phillips Curve — Kazakhstan:
      - αy: Gamma prior, Prior Mode 0.12, Prior Std. Dev. 0.05, Bounds [0.001, 1]
      - αlag: Beta prior, Prior Mode 0.60, Prior Std. Dev. 0.10, Bounds [0.001, 1]
      - αz: Beta prior, Prior Mode 0.05, Prior Std. Dev. 0.02, Bounds [0.001, 1]
    - Credit Gap — Kazakhstan:
      - θ1: Beta prior, Prior Mode 0.10, Prior Std. Dev. 0.05, Bounds [0.001, 0.99]
      - θ2: Beta prior, Prior Mode 0.80, Prior Std. Dev. 0.08, Bounds [0.001, 0.99]
      - θ3: Gamma prior, Prior Mode 0.05, Prior Std. Dev. 0.02, Bounds [0.001, ∞)
    - Okun's Law — Kazakhstan:
      - τ1: Normal prior, Prior Mode 0.12, Prior Std. Dev. 0.10, Bounds [0.001, 2]
      - τ2: Normal prior, Prior Mode 0.70, Prior Std. Dev. 0.10, Bounds [0.001, 0.99]
      - τ3: Normal prior, Prior Mode 0.08, Prior Std. Dev. 0.08, Bounds [0.001, 0.99]
      - τ4: Normal prior, Prior Mode 0.08, Prior Std. Dev. 0.08, Bounds [0.001, 0.99]
    - Risk Premium:
      - ρφ: Beta prior, Prior Mode 0.70, Prior Std. Dev. 0.10, Bounds [0.01, 1]
    - Foreign Output (IS Curve) — Russia:
      - βlagE: Beta prior, Prior Mode 0.50, Prior Std. Dev. 0.10, Bounds [0.001, 0.99]
      - βleadE: Beta prior, Prior Mode 0.30, Prior Std. Dev. 0.08, Bounds [0.001, 0.99]
      - βrE: Gamma prior, Prior Mode 0.03, Prior Std. Dev. 0.01, Bounds [0.001, ∞)
      - τE: Beta prior, Prior Mode 0.08, Prior Std. Dev. 0.03, Bounds [0.001, 0.99]
    - Foreign Phillips Curve — Russia:
      - αyE: Gamma prior, Prior Mode 0.07, Prior Std. Dev. 0.03, Bounds [0.001, 1]
      - αlagE: Beta prior, Prior Mode 0.50, Prior Std. Dev. 0.10, Bounds [0.01, 1]
    - Monetary Policy Rule — Russia (Taylor Rule):
      - γlagE: Normal prior, Prior Mode 0.80, Prior Std. Dev. 0.10, Bounds [0.001, 1]
      - γinflE: Normal prior, Prior Mode 1.70, Prior Std. Dev. 0.10, Bounds [0.01, 2]
      - γyE: Normal prior, Prior Mode 0.20, Prior Std. Dev. 0.10, Bounds [0.01, 1]
    - Autoregressive Processes:
      - ρrE: Beta prior, Prior Mode 0.70, Prior Std. Dev. 0.10, Bounds [0.01, 1]
      - ρπ: Beta prior, Prior Mode 0.40, Prior Std. Dev. 0.10, Bounds [0.01, 1]

- VII. Shock standard deviations — prior distributions
  - All shock standard deviations estimated with inverse-gamma prior distributions (uninformative scale).
  - Domestic shocks (prior modes and descriptions):
    - σY — 0.25 — Output gap shock
    - σG — 0.05 — Trend growth shock
    - σLGDP_BAR — 0.05 — Potential output level shock
    - σPIE — 0.25 — Inflation shock
    - σC_GAP — 0.25 — Credit gap shock
    - σC_BAR — 0.05 — Trend credit level shock
    - σGC — 0.06 — Trend credit growth shock
    - σUNR_GAP — 0.25 — Unemployment gap shock
    - σUNR_BAR — 0.05 — NAIRU level shock
    - σG_UNR_BAR — 0.05 — NAIRU growth shock
    - σPHI — 0.25 — Risk premium shock
    - σPI — 0.15 — Inflation target shock
  - Foreign shocks (Russia) (prior modes and descriptions):
    - σY_E — 0.18 — Output gap shock (RU)
    - σG_E — 0.06 — Trend growth shock (RU)
    - σLGDP_BAR_E — 0.04 — Potential output level shock (RU)
    - σPIE_E — 0.18 — Inflation shock (RU)
    - σINT_E — 0.20 — Monetary policy shock (RU)
    - σR_BAR_E — 0.20 — Natural real rate shock (RU)
  - Real exchange rate shocks:
    - σZ_GAP — 0.60 — Real exchange rate gap shock
    - σZ_BAR — 0.60 — Real exchange rate trend shock

- Estimation note:
  - The model is used primarily as a filtering device; the likelihood is not evaluated (prior-only estimation with system priors).

### Annex III — Banking sector crisis episodes (country summaries and dating)
- Azerbaijan (2015–2017)
  - Trigger: 2014 oil-price collapse causing two manat devaluations in 2015.
  - Effects: sharply raised FX credit risk, dollarization, and pressure on bank capital and liquidity.
  - Policy actions: creation of a dedicated financial regulator (FIMSA); sectoral restructuring; revocation of licenses of weaker banks; overhaul of the International Bank of Azerbaijan (IBA).
  - 2017: IBA halted foreign-debt payments and entered court-sanctioned restructuring under Azeri law; recognition proceedings in English courts document the process and timing.
  - Evaluation window dated: 2015Q1–2017Q4.
  - IMF characterization: economy recovering “from a banking crisis and recession.”
- Kazakhstan (2008–2010; 2014–2017)
  - First wave (2008Q3–2010Q4): pre-GFC wholesale-funded credit boom; funding dried up; multiple large banks required intervention; NPLs surged; system-wide banking shock requiring comprehensive resolution and recapitalization.
  - Second wave (2014Q4–2017Q4): oil-price weakness and 2015 move to a free float produced large tenge depreciation revealing legacy asset-quality problems; cleanup culminated in the 2017 rescue and merger of Kazkommertsbank (KKB) with Halyk; staff reports noted substantial expected public support and sector consolidation.
- Kyrgyz Republic (2010)
  - Event: domestic political upheaval in April–June 2010 triggered rapid deposit outflows and system-wide liquidity stress.
  - Measures: temporary administration in key banks; nationalization of Asia Universal Bank (AUB) and placement under conservatorship.
  - IMF note: 30 percent deposit decline within April 2010 and supervisory measures across the system.
  - Evaluation window dated: 2010Q2–2010Q4 as the core crisis phase.
- Tajikistan (2015–2016)
  - Underlying weaknesses: directed lending, governance problems, rising loan losses.
  - Resolution: late-2016 public recapitalization of two large banks amounting to about 6 percent of GDP.
  - IMF materials: record the fiscal cost and its impact on debt dynamics.
  - Evaluation window dated: 2015Q4–2016Q4 capturing run-up and immediate resolution.
- Armenia, Georgia, Uzbekistan
  - LV report: no systemic banking crises in these countries through 2017.
  - Observations: periods of rapid credit growth (notably Georgia in 2017–2021) and high dollarization, but macroprudential tightening and supervisory measures helped avoid generalized bank failures or large fiscal recapitalizations.
  - Treatment in analysis: these countries are treated as non-crisis cases for horizon-specific evaluation; their histories are used to define tranquil observations and to document vulnerability build-ups that did not crystallize into systemic breaks.

*Credit Expansion in the Caucasus and Central Asia — Working Paper No. WP/2026/137*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026137-source-pdf.pdf_