## wpiea2020065-print-pdf

## Source details

**Canonical URL:** [wpiea2020065-print-pdf](https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020065-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2020/english/wpiea2020065-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2020/english/wpiea2020065-print-pdf.pdf.json)

---

### Introduction: purpose and research question
- Motivation: The Global Financial Crisis (GFC) highlighted costs of systemic financial crises and increased use of macroprudential policies to prevent build-up of financial vulnerabilities.
- Research gap: Few studies systematically investigate the impact of macroprudential policies on the likelihood of financial crises; literature mostly focuses on intermediate financial variables.
- Primary objective: Examine the impact of macroprudential policies on the likelihood of systemic banking crises while accounting for potential adverse effects on economic growth and the net effect on crisis probability.
- Approach summary:
  - Step 1: Dynamic logit model à la Candelon et al. (2014) including a lagged crisis index and other predictors to estimate impact on banking-crisis likelihood.
  - Step 2: Bivariate VAR-like system (crisis probability and output growth) with impulse response analysis (GIRF) to measure direct (stabilizing) and indirect (via growth) effects of macroprudential policies.

### Data, indices, and sample
- Sample and period:
  - More than 100 countries; sample period reported as 2000−2016 at yearly frequency (summary statistics note refers to period 2000-2017).
- Key databases:
  - Macroprudential policies index (MPI): Cerutti et al. (2017) database based on IMF GMPI survey covering 134 countries (33 advanced economies, 63 emerging, 38 developing).
  - Banking crisis dummy: Laeven and Valencia (2018).
- MPI construction: composite index using 12 macroprudential instruments (CTC, LEV, DP, LTV, DTI, CG, FC, RR, TAX, SIFI, INTER, CONC).
- MPI sub-indices and aggregates:
  - LTV CAP and RRREV as MPI-related indicators.
  - MPI-Bor (borrower-facing): LTV CAP and DTI.
  - MPI-Fin (financial-institution-facing): DP, CTC, LEV, SIFI, INTER, CONC, FC, RRREV, CG, TAX.
- Descriptive statistics (selected):
  - Panel A (All countries, Obs. 2,178): Mean GDP growth 4.06; Crisis mean 0.06; FD mean 0.35; Debt-to-GDP mean 54.33; MPI mean 2.00; MPI-Borrow mean 0.33; MPI-Fin mean 1.66; KA mean 0.59; Trade-to-GDP mean 83.36.
  - Panel B (Advanced Economies, Obs. 576): Mean Crisis 0.14; MPI mean 1.86; MPI-Borrow mean 0.45; MPI-Fin mean 1.41; FD mean 0.66; GDP growth mean 2.41; Trade-to-GDP mean 98.60.
  - Panel C (Emerging Market Economies, Obs. 1,044): Mean Crisis 0.04; MPI mean 2.46; MPI-Borrow mean 0.39; MPI-Fin mean 2.06; FD mean 0.31; GDP growth mean 4.33; Trade-to-GDP mean 80.03.
  - Panel D (Low-Income Developing Countries, Obs. 558): Mean Crisis 0.02; MPI mean 1.27; MPI-Borrow mean 0.10; MPI-Fin mean 1.17; FD mean 0.13; GDP growth mean 5.26; Trade-to-GDP mean 73.85.
- MPI distribution:
  - AEs: MPI ranges 0 to 6; mean 1.86; standard deviation 1.53.
  - EMEs: MPI ranges 0 to 10; mean 2.46; standard deviation 1.89.
- Policy-tightening scenarios: delta = 1, 2, and 3 designated small, moderate, and large macroprudential tightenings.

### Methodology and model specification
- Dynamic panel logit (early-warning system) with fixed effects:
  - Pr(y_it = 1) = F(π_i,t) with π_i,t = g_i,t−1 β_1 + x_i,t−1^> β_2 + MPI_i,t−1 β_3 + δ π_i,t−1 + η_i (equation (1)).
  - Growth equation: g_i,t = Φ_1 . g_i,t−1 + Φ_2 . MPI_i,t−1 + Φ_3 . x_t−1 + Ψ π_it−1 + c_i + ν_i,t (equation (2)).
  - System estimated by Maximum Likelihood Estimator (MLE); crisis variable enters as a continuous probability in the multivariate system.
- Estimation details:
  - ML estimation with iterative pseudo-demeaning algorithm (Stammann et al. (2016)).
  - Analytical bias correction of Hahn and Kuersteiner (2011).
  - GIRF computed as in Appendix A; 68% confidence intervals via Moving-Block Bootstrap (Appendix B).
- Identification of effects:
  - Direct effect: MPI → lower crisis probability (eq. 1).
  - Indirect effect: MPI → lower growth (eq. 2) → higher crisis probability (back to eq. 1).
  - GIRF used to quantify timing and net outcome.

### Main empirical results — crisis probability and growth (baseline)
- Dynamic panel logit (Table 3) — MPI and crisis probability:
  - MPI coefficient (logit):
    - All: −0.606 (0.175) ***
    - AE: −0.411 (0.292) (not significant)
    - EME: −0.899 (0.286) ***
  - GDP Growth coefficient (on crisis prob):
    - All: −0.192 (0.039) ***
    - AE: −0.283 (0.065) ***
    - EME: −0.18 (0.059) ***
  - Crisis Index (autoregressive):
    - All: 0.56 (0.113) ***
    - AE: 0.534 (0.143) ***
    - EME: 0.2 (0.185) (not significant)
  - FD (Financial development) coefficient:
    - All: 12.574 (3.308) ***
    - AE: 11.884 (4.06) ***
    - EME: 18.289 (7.205) **
  - Economic magnitudes (from estimates):
    - One unit increase in MPI lowers conditional probability of systemic banking crisis by:
      - 7.29 percentage points (pooled sample).
      - 5.21 percentage points (Advanced Economies).
      - 8.83 percentage points (Emerging Market Economies).
    - One percentage point increase in real GDP growth lowers crisis probability by:
      - 2.31 percentage points (pooled sample).
      - 3.58 percentage points (AEs).
      - 1.78 percentage points (EMEs).
- Growth regressions (Table 4) — MPI and real GDP growth:
  - MPI coefficient (OLS with country FE):
    - All: −0.237 (0.079) ***
    - AE: −0.037 (0.132) (not significant)
    - EME: −0.318 (0.122) ***
  - Interpreted magnitudes:
    - One unit increase in MPI lowers economic growth by 23.7 basis points (pooled sample).
    - EMEs: one unit MPI lowers growth by 31.8 basis points.
    - No statistically significant MPI effect on growth in AEs.
  - Other notable coefficients:
    - Crisis Index depresses growth: All −0.204 (0.092) **; AE −0.202 (0.117) *; EME −0.147 (0.150) (not significant).
    - Debt-to-GDP loads positively on growth across samples (e.g., All: 0.011 (0.004) ***).

### GIRF analysis — direct vs indirect effects and timing
- Shock definition: positive variation by one unit of MPI corresponds to activation of an additional macroprudential instrument (delta = 1 small; delta = 2 moderate; delta = 3 large).
- GIRF procedure: initialize at unconditional means, simulate baseline and shock paths, compute differences (Appendix A); 68% CIs via Moving-Block Bootstrap (Appendix B).
- Advanced Economies — aggregate MPI GIRF (Figures 1):
  - Immediate effects on conditional probability of systemic banking crisis:
    - delta = 1: decrease of 3.93 percentage points.
    - delta = 2: decrease of 6.74 percentage points.
    - delta = 3: decrease of 8.7 percentage points.
  - 68% CIs for crisis-probability effects:
    - Small: from −1.31 to −23.9 percentage points.
    - Moderate: from −2.24 to −36.21 percentage points.
    - Large: from −2.84 to −41.48 percentage points.
  - Horizon and persistence: No statistically significant effect on crisis probability at horizons 1 to 8 years after tightening.
  - Effect on real GDP growth: No evidence of a significant effect (68% CI).
- Emerging Market Economies — aggregate MPI GIRF (Figures 2):
  - Immediate effects on conditional probability of systemic banking crisis:
    - delta = 1: decrease of 7.85 percentage points.
    - delta = 2: decrease of 11.47 percentage points.
    - delta = 3: decrease of 13.02 percentage points.
  - 68% CIs for crisis-probability effects:
    - Small: from -1.21 to -11 percentage points.
    - Moderate: from -1.67 to -17.06 percentage points.
    - Large: from -1.88 to -20.6 percentage points.
  - Effect on real GDP growth (initial impact, basis points):
    - delta = 1: contraction of 31.8 basis points.
    - delta = 2: contraction of 63.6 basis points.
    - delta = 3: contraction of 95.4 basis points.
  - 68% CIs for growth effects (basis points):
    - Small: from -17.5 to -56.4 basis points.
    - Moderate: from -35.1 to -112.7 basis points.
    - Large: from -52.6 to 169.1 basis points.
  - Horizon and persistence: No statistically significant effects beyond the initial introduction.

### Robustness and heterogeneity: borrower- vs financial-institution-targeted policies
- MPI-Fin vs MPI-Bor (Table 5 and Appendix C):
  - MPI-Fin drives much of aggregate MPI effect; MPI-Fin reduces crisis probability significantly for both AEs and EMEs.
  - MPI-Bor has limited contribution in AEs; larger effects in pooled and EME samples.
- Economic magnitudes (one unit change effects on crisis probability):
  - MPI-Bor: lowers probability by 8.91 percentage points (pooled) and by 16.21 percentage points (EMEs).
  - MPI-Fin: lowers probability by 9.38 percentage points (pooled), by 7.43 percentage points (AEs), and by 8.91 percentage points (EMEs).
- Effects on real GDP growth (one unit change):
  - MPI-Bor: contraction of real GDP growth by 36.6 basis points (pooled).
  - MPI-Fin: contraction of real GDP growth by 28.8 basis points (pooled) and by 36.8 basis points (EMEs).
- GIRF by instrument category (Appendix C Figures 3–6):
  - Advanced Economies: MPI-Fin tightening reduces crisis probability with effects statistically significant up to two years; no significant GDP effects detected.
  - Emerging Market Economies: MPI-Fin tightening reduces crisis probability and produces larger initial GDP contractions (e.g., small tightening: −36.8 basis points).

### Post-GFC (interaction) analysis
- Specification: include MPI × 1{t−1≥2008} interaction to capture post-2008 shifts.
- Banking crisis (Table 7):
  - MPI baseline: All −0.606 (0.175) ***; AE −0.411 (0.292) (n.s.); EME −0.899 (0.286) ***.
  - MPI × post-2008: All −0.061 (0.15); AE 0.239 (0.253); EME −0.43 (0.256) *.
  - Interpretation: For EMEs, effectiveness of MPPs in lowering crisis probability improved after 2008 (interaction negative and significant at 10%).
- Real GDP growth (Table 8):
  - MPI baseline on growth: All −0.237 (0.079) ***; AE −0.037 (0.132) (n.s.); EME −0.318 (0.122) ***.
  - MPI × post-2008: All −0.434 (0.075) ***; AE −0.644 (0.199) ***; EME −0.419 (0.092) ***.
  - Interpretation: Sensitivity of real activity to MPI changed after the GFC; macroprudential policies depress real economic activity more in the post-GFC period.
- Model fit:
  - Interaction specifications often increase adjusted R2 and improve fit for EMEs.

### Summary conclusions and policy-relevant implications
- Core findings:
  - Macroprudential policy tightenings reduce the conditional probability of systemic banking crises in both Advanced and Emerging Market Economies.
  - Macroprudential tightening depresses real GDP growth modestly (e.g., pooled sample −0.237 per one unit MPI; EMEs −0.318), creating an indirect destabilizing channel.
  - GIRF analysis indicates the direct stabilizing effect generally dominates the indirect growth-reducing effect, producing a net positive effect on financial stability.
- Heterogeneity:
  - The mitigating effect on banking crises is more pronounced in EMEs than in AEs.
  - MPI-Fin instruments are primary drivers of aggregate MPI effects; MPI-Bor effects are limited in AEs but larger in EMEs.
- Robustness:
  - Results hold across aggregate MPI, MPI-Fin, MPI-Bor, GIRFs, and post-GFC interaction specifications.
- Policy nuance:
  - The growth cost of macroprudential tightening is non-negligible (e.g., 23.7 basis points per one unit MPI increase pooled); calibration should consider country context, instrument type, and the trade-off between stability gains and output costs.
  - Post-GFC, the trade-off intensified, particularly in EMEs, implying careful policy design and sequencing after major global shocks.

*Source: wpiea2020065-print-pdf*

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

### wpiea2020065-print-pdf - References

### Introduction: purpose and research question
- Motivation: The Global Financial Crisis (GFC) highlighted the costs of systemic financial crises and led to increased use of macroprudential policies to prevent build-up of financial vulnerabilities and strengthen financial-system resilience.
- Research gap: Apart from Choi et al. (2018), few studies have systematically investigated the impact of macroprudential policies on the likelihood of financial crises; most literature focuses on intermediate financial variables (credit, housing) rather than crisis incidence.
- Primary objective: Examine the impact of macroprudential policies on the likelihood of systemic banking crises, while accounting for potential adverse effects of these policies on economic growth and the net effect on crisis probability.
- Approach summary:
  - Step 1: Dynamic logit model á la Candelon et al. (2014) including a lagged crisis index and other predictors to estimate impact on banking-crisis likelihood.
  - Step 2: Bivariate vector autoregressive (VAR) model comprising the crisis index and output growth (growth regression uses lagged regressors consistent with Cerutti et al. (2017) except for lagging).
  - Use of impulse response analysis (GIRF) to measure direct (stabilizing) and indirect (via growth) effects of macroprudential policies on crisis probability.

### Key empirical findings and statistics
- Sample and data:
  - Sample: more than 100 countries.
  - Period: 2000-2017.
  - Macroprudential policies database: Cerutti et al. (2017).
  - Banking crisis database: Laeven and Valencia (2018).
- Main quantitative results (baseline estimations):
  - A one unit increase in the macroprudential policies index (MPI) lowers the conditional probability of a systemic banking crisis by 7.29 percentage points on average (sample including both advanced and emerging market economies).
  - A one unit rise in the MPI lowers economic growth by 23.7 basis points.
- Heterogeneity:
  - Negative effects of macroprudential policies on both growth and banking crises are most significant in emerging market economies.
- Net effect:
  - Generalized impulse response function (GIRF) analysis of the dynamic system composed of output growth and crisis probability reveals that the direct stabilizing effect of macroprudential policies dominates their indirect effect through lower output growth.

### Literature context and contributions
- Two strands advanced:
  1. Direct assessment of macroprudential policies on incidence of financial crises — departure from the prevalent focus on intermediate financial variables (credit growth, house prices, non-performing loans).
  2. Joint assessment of real-economy effects (growth) and feedback from growth to financial stability — extends literature documenting damaging effects of macroprudential tightening on growth by linking these effects back to crisis probability.
- Methodological contribution:
  - Use of a mixed-measurement dynamic model (Creal et al. 2014) that simultaneously includes a dichotomous (crisis probability) and a continuous (growth) equation; estimation by Maximum Likelihood Estimator (MLE).
  - First paper to empirically assess the effect of various macroprudential tools on the probability of financial crisis within this simultaneous framework.

### Methodology and model specification
- Univariate early-warning system (EWS) dynamic panel model with fixed effects:
  - Pr(y_it = 1) = F(π_i,t)
  - π_i,t = g_i,t−1 β_1 + x_i,t−1^> β_2 + MPI_i,t−1 β_3 + δ π_i,t−1 + η_i    (equation (1) as specified)
  - Where Pr(y_t = 1) is the conditional probability of a banking crisis at time t given information at t−1, g represents GDP growth rate, MPI denotes macroprudential instruments, and x_t−1 is a set of control variables (Financial development, macroeconomic variables, etc).
- Extension to incorporate growth (bivariate system):
  - g_i,t = Φ_1 . g_i,t−1 + Φ_2 . MPI_i,t−1 + Φ_3 . x_t−1 + Ψ π_it−1 + c_i + ν_i,t    (equation (2) as specified)
  - System of equations (1) and (2) estimated by Maximum Likelihood Estimator (MLE).
  - Differences versus Cerutti et al. (2017): (i) probability of crisis and growth are weakly exogenous rather than strictly exogenous; (ii) crisis variable enters as a continuous probability rather than a binary indicator.
- Expected parameter signs:
  - MPI expected to load negative and statistically significant in the crisis-probability equation (eq. 1).
  - MPI expected to have a negative and significant coefficient in the growth equation (eq. 2) due to credit constraints reducing economic activity.
- Identification of direct and indirect effects:
  - Use impulse response function (IRF) / GIRF analysis to measure the overall effect of the opposing direct stabilizing and indirect growth-reducing impacts of macroprudential policies on crisis probability.

### Implications and interpretation
- Evidence indicates macroprudential policies have both a stabilizing direct effect (lowering crisis probability) and a contractionary indirect effect via lower growth.
- Quantitatively, the direct stabilizing effect dominates in the dynamic system studied, implying net reductions in crisis probability following macroprudential tightening despite modest negative effects on growth.
- Policy-relevant nuance: the growth-cost of macroprudential measures is non-negligible (23.7 basis points per one unit MPI increase) and is more pronounced in emerging market economies; thus policy design should consider country context and calibrate tools to balance stability gains against output costs.

*Source: wpiea2020065-print-pdf - References (content excerpt).*

### 4. Empirical Results

### 4. Empirical Results

### 4.1. Data
- Sample period: 2000−2016 at a yearly frequency (summary statistics note refers to period 2000-2017).
- Banking crisis dummy: Laeven and Valencia (2018) database of systemic financial crises.
- Macroprudential policy index (MPI) source: Cerutti et al. (2017) database based on IMF GMPI survey covering 134 countries (33 advanced economies, 63 emerging, 38 developing).
- MPI construction: composite index using 12 macroprudential instruments:
  - General Countercyclical Capital Buffer/Requirement (CTC)
  - Leverage Ratio for banks (LEV)
  - Time-Varying/Dynamic Loan-Loss Provisioning (DP)
  - Loan-to-Value Ratio (LTV)
  - Debt-to-Income Ratio (DTI)
  - Limits on Domestic Currency Loans (CG)
  - Limits on Foreign Currency Loans (FC)
  - Reserve Requirement Ratios (RR)
  - Levy/Tax on Financial Institutions (TAX)
  - Capital Surcharges on SIFIs (SIFI)
  - Limits on Interbank Exposures (INTER)
  - Concentration Limits (CONC)
- Additional MPI-related indicators:
  - LTV CAP: sub-set of LTV measures used as strict caps on new loans.
  - RRREV: subset of RR measures imposing wedges on foreign currency deposits or adjusted countercyclically.
- Aggregated MPI indices considered:
  - Borrower-facing: LTV CAP and DTI
  - Financial-institution-facing: DP, CTC, LEV, SIFI, INTER, CONC, FC, RRREV, CG, TAX
- Other variables:
  - Financial development index (FD): Svirydzenka (2016)
  - Macroeconomic controls: GDP growth rate, inflation (World Bank WDI)
  - External factors: trade (percent of GDP), Chinn-Ito capital account index (KA)
- Descriptive statistics highlights (Table 2):
  - Panel A (All countries, Obs. 2,178): Mean GDP growth 4.06; Crisis mean 0.06; FD mean 0.35; Debt-to-GDP mean 54.33; MPI mean 2.00; MPI-Borrow mean 0.33; MPI-Fin mean 1.66; KA mean 0.59; Trade-to-GDP mean 83.36.
  - Panel B (Advanced Economies, Obs. 576): Mean Crisis 0.14; MPI mean 1.86; MPI-Borrow mean 0.45; MPI-Fin mean 1.41; FD mean 0.66; GDP growth mean 2.41; Trade-to-GDP mean 98.60.
  - Panel C (Emerging Market Economies, Obs. 1,044): Mean Crisis 0.04; MPI mean 2.46; MPI-Borrow mean 0.39; MPI-Fin mean 2.06; FD mean 0.31; GDP growth mean 4.33; Trade-to-GDP mean 80.03.
  - Panel D (Low-Income Developing Countries, Obs. 558): Mean Crisis 0.02; MPI mean 1.27; MPI-Borrow mean 0.10; MPI-Fin mean 1.17; FD mean 0.13; GDP growth mean 5.26; Trade-to-GDP mean 73.85.

### 4.2. Macroprudential Policies and Banking crises
- Estimation approach:
  - Dynamic panel logit model (Kauppi and Saikkonen (2008); Candelon et al. (2012)); includes lagged crisis index to capture persistence.
  - ML estimation using iterative pseudo-demeaning algorithm (Stammann et al. (2016)).
  - Analytical bias correction of Hahn and Kuersteiner (2011) applied.
  - Countries without at least one systemic banking crisis over the sample discarded prior to estimation.
- Key coefficient estimates (Table 3: pooled sample and subsamples):
  - MPI coefficient:
    - All: −0.606 (standard error 0.175), significance: ***
    - AE: −0.411 (0.292), not statistically significant
    - EME: −0.899 (0.286), significance: ***
  - GDP Growth coefficient:
    - All: −0.192 (0.039), ***
    - AE: −0.283 (0.065), ***
    - EME: −0.18 (0.059), ***
  - Crisis Index (autoregressive):
    - All: 0.56 (0.113), ***
    - AE: 0.534 (0.143), ***
    - EME: 0.2 (0.185), not significant
  - FD (Financial development) coefficient:
    - All: 12.574 (3.308), ***
    - AE: 11.884 (4.06), ***
    - EME: 18.289 (7.205), **
  - Debt-to-GDP, KA, Trade-to-GDP: coefficients not consistently significant across specifications.
  - Model fit metrics (All/AE/EME): AIC 532.396 / 321.042 / 170.112; BIC 727.937 / 429.619 / 238.075; Pseudo R2 0.28 / 0.29 / 0.328.
  - Observations used in estimation: Total observations 20,575; Effective observations 6,298; # Effect. Obs. 629 / 357 / 221 (All / AE / EME).
- Economic magnitudes derived from estimates:
  - A one unit change in the MPI index lowers the conditional probability of a systemic banking crisis by:
    - 7.29 percentage points on average (pooled sample).
    - 5.21 percentage points for Advanced Economies.
    - 8.83 percentage points for Emerging Market Economies.
  - A one percentage point increase in real GDP growth rate lowers crisis probability by:
    - 2.31 percentage points (pooled sample).
    - 3.58 percentage points for AEs.
    - 1.78 percentage points for EMEs.
- Interpretations and conclusions:
  - MPI has a statistically significant negative association with crisis probability in the pooled sample and in EMEs; effect in AEs is negative but not statistically significant.
  - Higher financial development (FD) is associated with a higher probability of banking crises across samples.
  - Macroprudential policies, economic growth, and financial development are the most informative predictors of banking crises in these estimations.
  - The autoregressive Crisis Index coefficient is below 1, indicating no absorbing crisis state; static models would be misspecified.

### 4.3. Macroprudential policies, Economic Growth and banking crises
- Estimation approach:
  - Growth regression via OLS with one-way country fixed effects.
  - Dependent variable: year-on-year real GDP growth.
  - Regressors considered as lagged values; estimation period 2001-2017.
- Key coefficient estimates (Table 4: All / AE / EME):
  - MPI coefficient:
    - All: −0.237 (0.079), ***
    - AE: −0.037 (0.132), not significant
    - EME: −0.318 (0.122), ***
  - GDP Growth (lagged) coefficient:
    - All: 0.306 (0.022), ***
    - AE: 0.368 (0.044), ***
    - EME: 0.372 (0.031), ***
  - Crisis Index coefficient:
    - All: −0.204 (0.092), **
    - AE: −0.202 (0.117), *
    - EME: −0.147 (0.150), not significant
  - FD coefficient:
    - All: −6.320 (1.792), ***
    - AE: −6.863 (2.977), **
    - EME: −5.510 (2.446), **
  - Debt-to-GDP coefficient:
    - All: 0.011 (0.004), ***
    - AE: 0.023 (0.009), ***
    - EME: 0.028 (0.008), ***
  - Trade-to-GDP coefficient:
    - All: 0.002 (0.005), not significant
    - AE: −0.031 (0.011), ***
    - EME: 0.021 (0.009), **
  - Model fit metrics:
    - Observations: 2,057 (All), 544 (AE), 986 (EME)
    - R2: 0.361 (All), 0.299 (AE), 0.347 (EME)
    - Adjusted R2: 0.319 (All), 0.247 (AE), 0.301 (EME)
- Economic magnitudes and interpretation:
  - MPI depresses economic growth:
    - Pooled sample: a one unit increase in MPI lowers economic growth by 23.7 basis points (i.e., −0.237).
    - EMEs: a one unit increase in MPI lowers economic growth by 31.8 basis points (i.e., −0.318).
    - No statistically significant MPI effect on growth in AEs (−0.037, not significant).
  - Crisis Index (banking crises) depresses growth in the pooled sample and in AEs.
  - Debt-to-GDP loads positively on growth across samples.
  - Trade openness has opposite effects: negative in AEs and positive in EMEs.

### 4.4. Direct versus Indirect effects: A Multivariate analysis
- Conceptual framing:
  - Direct effect: MPI reduces banking crisis probability by mitigating banking sector vulnerabilities (equation (1) effect).
  - Indirect effect: MPI depresses economic growth (equation (2) effect), and lower growth raises crisis probability per equation (1) estimates.
  - Net effect question: whether direct stabilizing effect of MPI dominates or the indirect destabilizing effect via growth dominates.
- Analytical strategy:
  - Use Generalized Impulse Response Function (GIRF) analysis to disentangle direct and indirect effects and to uncover the time horizon over which one effect dominates.
- Shock definition:
  - A positive variation by one unit of the MPI corresponds to activation of an additional macroprudential policy instrument (examples: introduction of an LTV cap or a capital surcharge for SIFIs).
  - This definition reflects a substantial policy tightening as opposed to minor recalibrations.
- Expected outcomes to be evaluated by GIRF (as laid out in text):
  - Impact of a one-unit MPI tightening on:
    - Conditional probability of a systemic banking crisis (direct and indirect channels).
    - Real GDP growth (as the mediating variable for the indirect channel).
- Empirical motivation:
  - Prior results indicate MPI reduces crisis probability (direct) but also depresses growth (indirect), creating potential ambiguity for net effect on financial stability.
  - GIRF analysis intended to quantify timing and magnitude of these opposing channels.

*Source: wpiea2020065-print-pdf - 4. Empirical Results*

### Appendix B describes in details the method used to obtain it at an horizonh.

### wpiea2020065-print-pdf - Appendix B describes in details the method used to obtain it at an horizonh.

### MPI distribution and definitions
- Advanced Economies (AEs): MPI ranges from 0 to 6; mean 1.86; standard deviation 1.53.
- Emerging Market Economies (EMEs): MPI ranges from 0 to 10; mean 2.46; standard deviation 1.89.
- Policy-tightening scenarios: delta = 1, 2, and 3 are respectively designated small, moderate, and large macroprudential policy tightenings.
- MPI sub-indices:
  - MPI-Bor: instruments targeted at borrowers’ leverage and financial positions (includes loan-to-value caps and debt-to-income ratios).
  - MPI-Fin: instruments targeted at financial institutions’ assets or liabilities (contains dynamic loan-loss provisioning, countercyclical capital buffers, leverage ratios for banks, capital surcharges on SIFIs, limits on interbank exposures, concentration limits, limits on domestic and foreign currency loans, reserve requirements, and tax on financial institutions).

### GIRF results — Advanced Economies (Figures 1 / MPI aggregate)
- Effect on conditional probability of a systemic banking crisis (initial impact just after MPI tightening):
  - Small tightening (delta = 1): decrease of 3.93 percentage points.
  - Moderate tightening (delta = 2): decrease of 6.74 percentage points.
  - Large tightening (delta = 3): decrease of 8.7 percentage points.
- 68% confidence intervals (Moving-Block Bootstrap, Appendix B) for crisis-probability effects:
  - Small: from −1.31 to −23.9 percentage points.
  - Moderate: from −2.24 to −36.21 percentage points.
  - Large: from −2.84 to −41.48 percentage points.
- Horizon and persistence:
  - No statistically significant effect on the probability of crisis at horizons ranging from 1 to 8 years after the initial tightening.
- Effect on real GDP growth:
  - No evidence of a significant effect on real GDP growth when considering a 68% confidence interval.

### GIRF results — Emerging Market Economies (Figures 2 / MPI aggregate)
- Effect on conditional probability of a systemic banking crisis (initial impact):
  - Small tightening (delta = 1): decrease of 7.85 percentage points.
  - Moderate tightening (delta = 2): decrease of 11.47 percentage points.
  - Large tightening (delta = 3): decrease of 13.02 percentage points.
- 68% confidence intervals for crisis-probability effects:
  - Small: from -1.21 to -11 percentage points.
  - Moderate: from -1.67 to -17.06 percentage points.
  - Large: from -1.88 to -20.6 percentage points.
- Effect on real GDP growth (initial impact, in basis points):
  - Small tightening: contraction of 31.8 basis points.
  - Moderate tightening: contraction of 63.6 basis points.
  - Large tightening: contraction of 95.4 basis points.
- 68% confidence intervals for real GDP growth effects (basis points):
  - Small: from -17.5 to -56.4 basis points.
  - Moderate: from -35.1 to -112.7 basis points.
  - Large: from -52.6 to 169.1 basis points.
- Horizon and persistence:
  - No statistically significant effect on crisis probability or real economic activity for periods following the initial introduction of the policy tightening.

### Interpretation and net effect
- Across both AEs and EMEs, macroprudential policy tightenings are associated with a net positive effect on financial stability (lower conditional probability of systemic banking crises).
- In AEs the stabilizing direct effect tends to dominate without detectable adverse GDP effects at the 68% CI level.
- In EMEs macroprudential tightening improves financial stability but produces an initially measurable contraction in real GDP growth; nonetheless, the stabilizing direct effect still appears to dominate the indirect harmful effect on economic activity.

### Robustness: borrower- vs financial-institution-targeted policies (section 5.1 and Table 5)
- Aggregate results confirm that macroprudential policy tools significantly decrease the probability of systemic banking crisis.
- MPI-Fin drives much of the aggregate MPI effect:
  - MPI-Fin reduces crisis probability significantly for both AEs and EMEs.
  - MPI-Bor does not contribute significantly to the conditional probability of a systemic banking crisis in Advanced Economies.
- Economic magnitudes (one unit change effects on crisis probability):
  - MPI-Bor: lowers probability of crisis by 8.91 percentage points (all countries pooled) and by 16.21 percentage points (EMEs).
  - MPI-Fin: reduces conditional probability of a banking crisis by 9.38 percentage points (all countries pooled), by 7.43 percentage points (AEs), and by 8.91 percentage points (EMEs).
- Effects on real GDP growth (one unit change):
  - MPI-Bor: contraction of real GDP growth by 36.6 basis points (all countries pooled).
  - MPI-Fin: contraction of real GDP growth by 28.8 basis points (all countries pooled) and by 36.8 basis points (EMEs).
- Other macroeconomic controls:
  - Debt-to-GDP (positive sign) and Trade-to-GDP (negative sign for AEs and positive sign for EMEs) retain statistical significance consistent with baseline estimations.
- MPI-Fin descriptive statistics:
  - AEs: MPI-Fin ranges 0 to 6; mean 1.41; standard deviation 1.25.
  - EMEs: MPI-Fin ranges 0 to 8; mean 2.06; standard deviation 1.51.
- MPI-Bor has a much narrower range (composed of two categories) with many country-year zeros.

### Key statistics and estimation details (selected)
- Estimation period: 2001-2017.
- Confidence intervals reported: 68% (Moving-Block Bootstrap, Appendix B).
- Bias correction: analytical bias correction of Hahn & Kuersteiner (2011) applied; ML estimation with iterative pseudo-demeaning algorithm of Stammann et al. (2016).
- Table 5 selected coefficients (economic significance preserved as reported):
  - MPI-Bor: −0.705 (All), 0.368 (AE), −1.556 (EME) with reported standard errors (0.428)(0.719)(0.651).
  - MPI-Fin: −0.78 (All), −0.59 (AE), −1.094 (EME) with reported standard errors (0.219)(0.343)(0.356).
  - GDP Growth effects on crisis probability reported (multiple columns) with coefficients around −0.179 to −0.188 and standard errors shown.
  - Financial development (FD) coefficients reported (e.g., 11.061, 13.465, 15.464) with standard errors (3.156)(4.174)(6.852), etc.
- Model fit indicators in Table 5 include AIC and BIC values and McFadden pseudo R2 for each specification.

### Summary conclusions
- Macroprudential policy tightenings reduce the conditional probability of systemic banking crises in both Advanced and Emerging Market Economies.
- The stabilizing effect on financial stability is evident for small, moderate, and large MPI shocks.
- In EMEs, tighter macroprudential policy is associated with a measurable short-run contraction in real GDP growth at the time of introduction, but the net effect favors financial stability.
- Policies targeted at financial institutions’ assets or liabilities (MPI-Fin) are the primary drivers of the aggregate MPI effect on crisis probability; borrower-targeted tools (MPI-Bor) show limited effect in AEs but larger economic-significance estimates in pooled and EME samples.

*Source: wpiea2020065-print-pdf - Appendix B describes in details the method used to obtain it at an horizonh.*

### Appendix C for completeness.

### Appendix C

### Macroprudential borrower and financial indices — regression results (Table 6)
- Dependent variable: year-on-year real GDP growth rate. Estimation by OLS with one-way country fixed effects; estimation period 2001-2017. Standard errors clustered at country level.
- Key coefficient estimates (All / AE / EME), standard errors in parentheses; ***, **, * indicate significance at the 1-, 5-, and 10-percent levels:
  - MPI-Bor: −0.366 ∗∗ (0.186) / −0.191 (0.268) / −0.439 (0.282)
  - MPI-Fin: −0.288 ∗∗∗ (0.103) / 0.019 (0.183) / −0.368 ∗∗ (0.156)
  - GDP Growth (lag): 0.307 ∗∗∗ (0.022) / 0.361 ∗∗∗ (0.044) / 0.379 ∗∗∗ (0.031)
  - Crisis Index: −0.274 ∗∗ (0.114) / −0.231 ∗∗ (0.115) / −0.032 (0.183)
  - FD: −6.827 ∗∗∗ (1.765) / −6.632 ∗∗ (2.972) / −7.316 ∗∗∗ (2.297)
  - Debt-to-GDP: 0.011 ∗∗∗ (0.004) / 0.024 ∗∗∗ (0.009) / 0.028 ∗∗∗ (0.008)
  - KA: −0.593 (0.628) / 0.893 (1.373) / −0.786 (0.780)
  - Trade-to-GDP: 0.001 (0.005) / −0.030 ∗∗∗ (0.010) / 0.020 ∗∗ (0.009)
- Sample and fit:
  - Observations: 2,057 / 544 / 986 (All / AE / EME)
  - R2: 0.361 / 0.301 / 0.344 (All / AE / EME)
  - Adjusted R2: 0.319 / 0.249 / 0.298

### GIRF analysis of financial macroprudential tightening (Figures 3 and 4)
- Method: Generalized Impulse Response Functions computed as in Appendix A; 68% confidence intervals from Moving-Block Bootstrap (Appendix B). Horizons reported up to eight years after tightening.
- Advanced Economies (Figure 3):
  - A tightening in MPI-Fin reduces the conditional systemic banking crisis probability; effect statistically significant up to two years.
  - Crisis probability two years after MPI-Fin tightening is lower than no-policy-change case by:
    - 2.1 percentage points for a small policy tightening,
    - 3.83 percentage points for a moderate tightening,
    - 5.31 percentage points for a large tightening.
  - No evidence of a significant effect of MPI-Fin tightening on real GDP growth in AEs; noted larger estimation uncertainty (wide 68% CIs).
- Emerging Market Economies (Figure 4):
  - MPI-Fin tightening produces a statistically significant decrease in systemic banking crisis probability; magnitude comparable to aggregate MPI results.
  - Real GDP growth effects are more pronounced than for aggregate MPI. Initial decreases in real GDP growth after MPI-Fin tightening:
    - Small tightening: −36.8 basis points (68% CI: −27.42 to −75.8 basis points)
    - Moderate tightening: −73.62 basis points (68% CI: −54.83 to −151.59 basis points)
    - Large tightening: −110.43 basis points (68% CI: −82.25 to −227.39 basis points)
  - No further statistically significant effects at horizons longer than the initial introduction.

### Impact of the Global Financial Crisis (GFC) on transmission (Tables 7 and 8)
- Approach: Add interaction MPI × 1{t−1≥2008} to capture potential post-2008 changes in sensitivity of crisis probability and real GDP growth to MPI.
- Banking crises — dynamic panel logit (Table 7). ML estimation with iterative pseudo-demeaning; bias-corrected estimates; standard errors clustered at country level.
  - Baseline MPI coefficients (left panel): MPI −0.606 ∗∗∗ (0.175) / −0.411 (0.292) / −0.899 ∗∗∗ (0.286) for All / AE / EME.
  - MPI × 1{t−1≥2008} (right panel): −0.061 (0.15) / 0.239 (0.253) / −0.43 ∗ (0.256) for All / AE / EME.
  - GDP Growth: −0.192 ∗∗∗ (0.039) / −0.283 ∗∗∗ (0.065) / −0.18 ∗∗∗ (0.059)
  - Crisis Index (persistence): 0.56 ∗∗∗ (0.113) / 0.534 ∗∗∗ (0.143) / 0.20 (0.185)
  - FD: 12.574 ∗∗∗ (3.308) / 11.884 ∗∗∗ (4.06) / 18.289 ∗∗ (7.205)
  - Debt-to-GDP: −0.009 (0.007) / −0.026 ∗∗ (0.012) / 0.013 (0.015)
  - Trade-to-GDP: 0.013 (0.012) / 0.024 (0.021) / 0.045 (0.028)
  - Model fit and sample:
    - AIC: 532.396 / 321.042 / 170.112 (baseline); 531.382 / 325.101 / 165.986 (with interaction)
    - BIC: 727.937 / 429.619 / 238.075 (baseline); 731.368 / 437.556 / 237.347 (with interaction)
    - Pseudo R2: 0.28 / 0.29 / 0.328 (baseline); 0.285 / 0.284 / 0.36 (with interaction)
    - Effective Observations (# Effect. Obs.): 629 / 357 / 221
    - Total Observations: 2,057 / 544 / 986
- Real GDP growth — linear panel (Table 8). OLS with one-way country fixed effects; standard errors clustered at country level.
  - Baseline MPI coefficients (left panel): MPI −0.237 ∗∗∗ (0.079) / −0.037 (0.132) / −0.318 ∗∗∗ (0.122)
  - MPI × 1{t−1≥2008} (right panel): −0.434 ∗∗∗ (0.075) / −0.644 ∗∗∗ (0.199) / −0.419 ∗∗∗ (0.092)
  - GDP Growth (lag): 0.306 ∗∗∗ (0.022) / 0.368 ∗∗∗ (0.044) / 0.372 ∗∗∗ (0.031)
  - Crisis Index: −0.204 ∗∗ (0.092) / −0.202 ∗ (0.117) / −0.147 (0.150)
  - FD: −6.320 ∗∗∗ (1.792) / −6.863 ∗∗ (2.977) / −5.510 ∗∗ (2.446)
  - Debt-to-GDP: 0.011 ∗∗∗ (0.004) / 0.023 ∗∗∗ (0.009) / 0.028 ∗∗∗ (0.008)
  - Trade-to-GDP: 0.002 (0.005) / −0.031 ∗∗∗ (0.011) / 0.021 ∗∗ (0.009)
  - Sample and fit:
    - Observations: 2,057 / 544 / 986
    - R2: 0.361 / 0.299 / 0.347 (baseline); 0.372 / 0.315 / 0.362 (with interaction)
    - Adjusted R2: 0.319 / 0.247 / 0.301 (baseline); 0.331 / 0.262 / 0.317 (with interaction)

- Interpretation of GFC effects:
  - For banking crises:
    - The GFC did not significantly alter the effectiveness of MPPs in lowering crisis probability in Advanced Economies.
    - For EMEs, the effectiveness of MPPs in lowering systemic banking crisis probability improved after 2008 (MPI × post-2008 coefficient negative and significant at the 10-percent level).
    - The modified specification shows better fit for EMEs (lower AIC/BIC and higher pseudo R2).
  - For real GDP growth:
    - The sensitivity of real economic activity to MPI changed markedly after the GFC: MPI exhibits a statistically significant positive sign in pooled and AE samples in the baseline, but the MPI × post-2008 interaction coefficients are negative and significant in all cases.
    - Conclusion: macroprudential policies depress real economic activity in the post-GFC period (pooled sample and EMEs).
    - The alternative specification increases adjusted R2 in all three samples, indicating improved fit.

### Main findings and conclusions (Section 6 and synthesis)
- Macroprudential policies (MPPs):
  - Have a statistically and economically significant direct negative effect on the probability of banking crises.
  - Depress economic growth (indirect channel).
  - Despite the adverse effect on growth, MPPs have a positive net effect on financial stability: net reduction in incidence of systemic banking crises.
- Heterogeneity:
  - The mitigating effect of MPPs on banking crises is more pronounced in Emerging Market Economies than in Advanced Economies.
  - Borrower-based macroprudential tools appear more effective in EMEs; financial-based macroprudential tools are more useful in AEs for taming banking crises.
- Robustness and policy implications:
  - Results hold across alternative indices (aggregate MPI, MPI-Fin, MPI-Bor) and robustness checks including GIRFs and post-GFC interaction specifications.
  - The trade-off between financial stability gains and growth costs intensified after the Global Financial Crisis, particularly in EMEs.

*Appendix C, wpiea2020065-print-pdf - Appendix C for completeness.*

### References

### wpiea2020065-print-pdf - References

### Appendix A — Generalized Impulse Response Function (GIRF) of a tightening shock to macroprudential supervision
- Purpose: compute response of the crisis probability and of real GDP growth to an exogenous shock to the macroprudential index (MPI) using a GIRF approach.
- Procedure (for h = 1, ... , H):
  - For h = 1:
    - Initialize all explanatory variables to their unconditional means.
    - Country fixed effect parameters from equations (1) and (2) are set to the mean of the estimated parameters in each case.
    - Compute two paths of the bivariate system (conditional probability of crisis and real GDP growth):
      - Baseline path (initialized as above).
      - Shock path: add a one-period shock, δ > 0, to the initial value of the MPI.
  - For h > 1:
    - Exogenous variables set to their initial (unconditional mean) value.
    - Endogenous variables (real GDP growth and latent crisis index) are updated recursively using their lag value following the two-equation system (equations (1) and (2)).
  - GIRF definition:
    - The GIRF for crisis probability and real GDP growth = difference between the path with the one-period macroprudential tightening (positive shock) and the baseline path.
- Implementation note:
  - Country-specific averages are first computed and then a cross-sectional average is taken (footnote 11).

### Appendix B — Moving-block bootstrap procedure for GIRFs
- Model setup:
  - Dynamic panel logit with fixed effect for crisis indicator (equation (B.1)):
    - y1,nt = 1{y*1,nt ≥ 0}
    - y*1,nt = α1,n + x'nt−1 β1 + y'nt−1 ρ1 + u1,nt
  - Dynamic linear panel with fixed effects for real GDP growth (equation (B.2)):
    - y2,nt = α2,n + x'nt−1 β2 + y'nt−1 ρ2 + u2,nt
  - Estimated parameter vectors:
    - θ̂1 ≡ [α̂'1, β̂'1, ρ̂'1]'
    - θ̂2 ≡ [α̂'2, β̂'2, ρ̂'2]'
- Construction of matrix Zt to preserve cross-sectional dependence:
  - Zt ≡ [û1t, û2t, Xt−1] with rows stacking countries:
    - Each country row contains: û1,n,t; û2,n,t; x1,n,t−1; x2,n,t−1; ...; xK,n,t−1
  - For countries with no systemic banking crisis in the sample (and thus not used in (B.1) estimation), fitted residuals of the model are replaced by a missing value.
- Moving-block bootstrap (MBB) setup:
  - Given block length b and time-series dimension T, generate T − b + 1 overlapping blocks:
    - Z1,...,Zb ; Z2,...,Zb+1 ; ... ; ZT−b+1,...,ZT
- Bootstrap algorithm to compute bootstrapped confidence intervals for GIRFs:
  1. Draw (with replacement) (T_burn + T)/b blocks from the T − b + 1 overlapping blocks, where T_burn is a burn-in period. Call the resulting sample replicate Z(r).
  2. Compute recursively the model-implied endogenous variables y(r)1,nt and y(r)2,nt using equations (B.1)-(B.2) with estimated parameters θ̂1 and θ̂2. Keep only the last T observations to avoid dependence on initial conditions.
  3. Re-estimate models (B.1)-(B.2) and store θ̂(r)1 and θ̂(r)2.
  4. Compute and store the implied GIRFs Îy(H, δ, Ωt−1)(r) using the approach in Appendix A.
  5. Repeat steps 1–4 for r = 1, ... , R (R = total number of bootstrap replications).
  6. For h = 0,1, ... , H, construct a 68% confidence interval around the GIRF for horizon h by taking the 16- and 84-percentile of the distribution of the R replicas of Îy(h, δ, Ωt−1)(r).

### Appendix C — Targeting borrowers vs. financial institutions: GIRFs for MPI-Borrower (figures summary)
- Figures C.5 and C.6 present GIRF analyses of borrower macroprudential tightenings for:
  - Advanced Economies (Figure C.5)
  - Emerging Market Economies (Figure C.6)
- Shock scenarios:
  - Positive borrower macroprudential shocks denoted by delta = 1, delta = 2, delta = 3.
- Outcomes displayed:
  - Top panel: impulse response of the probability of a systemic banking crisis (red).
  - Bottom panel: impulse response of real GDP growth (green).
- Horizon:
  - Responses displayed up to eight years after the policy tightening (horizon shown ranges up to eight years).
- Inference:
  - Generalized Impulse Response Functions computed following Appendix A.
  - Shaded areas correspond to the 68% confidence intervals obtained from the Moving-Block Bootstrap procedure detailed in Appendix B.

*Source: wpiea2020065-print-pdf - References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2020/english/wpiea2020065-print-pdf.pdf_