## _wp0893

## Source details

**Canonical URL:** [_wp0893](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp0893.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp0893.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp0893.pdf.json)

---

### I. INTRODUCTION — scope, motivation, and contribution
- Banking crises in the 1990s included transition-economy meltdowns, the East Asian crises in 1997–99, and the Dominican Republic crisis of 2003.
- Historical impact: banking crises have imposed huge fiscal costs of resolution and/or sharp output losses.
- Contribution:
  - Applies a binary classification tree (BCT) technique to analyze banking crises in a sample of 50 emerging market and developing countries during 1990–2005.
  - Explicitly considers financial soundness indicators (FSI) alongside macroeconomic and external indicators.
- Explanatory variable categories considered:
  - Macroeconomic environment: growth, inflation, nominal depreciation, government balance.
  - External vulnerability: official foreign exchange reserve (FX) cover of broad money, export growth, terms of trade (TOT) growth.
  - Monetary conditions: credit growth, real deposit rate, foreign interest rate, existence of explicit deposit insurance, de facto exchange rate regime.
  - Banking sector health: liability dollarization (FX deposits in total official FX reserves), net FX open position, bank liquidity, equity strength, asset quality, two proxies for bank profitability.

### II. BINARY CLASSIFICATION TREE (BCT) METHODOLOGY — approach, advantages, limitations
- Method overview:
  - Nonparametric technique that iteratively splits a parent node into child nodes by selecting variables and thresholds that increase homogeneity (crisis vs. noncrisis).
  - Default model selection: V-fold cross validation with 10 folds; authors may overrule default tree size for in-sample fit and interpretability.
- Variable importance and masking:
  - “Variable importance index” ranks indicators by overall discriminatory power; mitigates but does not eliminate masking.
- Advantages:
  - No assumptions about functional form.
  - Captures nonlinearities and threshold effects and interactions.
  - Handles correlated indicators and missing values (uses surrogates).
  - Robust to outliers and monotonic transformations.
- Limitations:
  - Does not establish sample-wide marginal relationships (unlike logit/probit).
  - Masking problem may hide useful variables from final tree.
  - Optimal tree via V-fold may require judgment-based pruning/growth for economic meaningfulness.

### III. SAMPLE, CRISIS DEFINITION, AND DATA
- Sample: annual observations from 50 emerging market and developing countries during 1990–2005.
- Banking crisis definition: episodes involving exhaustion of bank capital/closure/merger/nationalization; extensive bank runs; or large-scale central bank liquidity support.
- Crisis observations: 127 annual crisis observations comprising 38 crisis episodes.
- Among 50 countries, 31 experienced at least one banking crisis during 1990–2005; crisis durations typically 2–3 years, some exceeding six years (e.g., Ecuador and Jamaica).
- Control group: 19 countries did not experience any systemic banking crisis during the sample period.
- Baseline sample adjustments:
  - To avoid post-crisis contamination, crisis observations after the first year of any given crisis episode are deleted, reducing sample size from 800 to 711 observations, with 38 crises.
  - Unconditional crisis probability in reduced sample: 5.3 percent.

### IV. STYLIZED FACTS BEFORE AND DURING CRISES (selected)
- Macroeconomic fundamentals prior to banking crises (vs. tranquil times):
  - Real GDP growth and fiscal balances are lower.
  - Inflation and nominal depreciation are higher.
- External sector prior to crises:
  - Official FX cover of broad money, export growth, and TOT growth are lower.
- Monetary conditions prior to crises:
  - Higher foreign interest and higher domestic real deposit rates.
  - Real credit growth much higher in run-up to banking crisis.
- Banking sector indicators prior to crises:
  - Banking system liquidity dries up (higher private credit to deposits).
  - Liability dollarization increases (higher FX deposits to reserves).
  - FX open positions widen.
  - Larger share of countries had explicit deposit insurance before a banking crisis.
  - Capital ratios and profitability are lower before a crisis.
  - NPL ratio is relatively better before a crisis and worsens when crisis is underway (possible masking by rapid credit growth).

### V. BASELINE BCT RESULTS — performance, primary splitters, and key terminal nodes
- Baseline tree: 8 terminal nodes.
  - In-sample fit: signals all but 1 of the 38 crises (97 percent) and has false signals for 19 percent of noncrisis outcomes.
- Parent-node primary split:
  - Inflation threshold: 19 percent
    - Inflation > 19 percent → conditional probability of crisis increases from 5.3 percent to over 16 percent.
    - Inflation < 19 percent → conditional probability of crisis declines to 2.8 percent.
- Five most important candidate variables identified (out of 19), in order:
  - nominal depreciation
  - bank profitability (interest rate spread)
  - inflation
  - liability dollarization (FX deposits to official FX reserves)
  - bank liquidity (private credit to deposits)
- Key crisis-prone terminal nodes (exact thresholds and conditional probabilities):
  - Terminal node 1:
    - Conditions: inflation greater than 18.7 percent AND TOT growth less than 3.3 percent
    - Conditional probability of crisis: over 21 percent
    - Median diagnostics: nominal depreciation 35.2; interest rate spread 13.0; inflation 44.7; liability dollarization index 67.5; illiquidity index 74.7
    - Comprises 22 crisis episodes including Argentina (1990), Brazil (1990, 1994), Bulgaria (1996), Colombia (1999), Ecuador (1996), Guyana (1993), Jamaica (1994), Jordan (1990), Kenya (1992), Lebanon (1990), Mexico (1992), Nicaragua (1990), Nigeria (1993), Peru (1993), Poland (1991), Tanzania (1990), Turkey (1991, 1994, 2000), and Venezuela (1993)
  - Terminal node 2:
    - Conditions: interest rate spread less than 3.1 percent AND export growth less than 11.8 percent
    - Conditional probability of crisis: 21 percent
    - Median diagnostics: nominal depreciation 0.0; interest rate spread 2.1; inflation 5.4; liability dollarization index 20.0; illiquidity index 103.3
    - Comprises crises in Argentina (1994, 2001), China (1999), Indonesia (1992, 1997), Korea (1997), Sri Lanka (1990), Malaysia (1997), and Thailand (1997)
  - Terminal node 3:
    - Conditions: FX deposits to official FX reserves more than 140 percent AND nominal depreciation more than 9 percent
    - Conditional probability of crisis: 25 percent
    - Median diagnostics: nominal depreciation 11.8; interest rate spread 10.6; inflation 10.8; liability dollarization index 201.8; illiquidity index 96.0
    - Contains crises in Bolivia (1994), Czech Republic (1991), Dominican Republic (2003), Nicaragua (2000), Philippines (1998), and Uruguay (2002)
  - Terminal node 4:
    - Condition: private credit to deposits more than 150 percent
    - Conditional probability of crisis: 100 percent
    - Observation: only one outcome in this node (Croatia, 1996) — treated as an outlier
- Diagnostic groupings / thematic vulnerabilities identified:
  - Macroeconomic Instability (high inflation, weak TOT growth).
  - Low Bank Profitability.
  - High FX Risk with Nominal Depreciation.
  - High FX Risk with Low Liquidity.

### VI. VARIABLE RANKING AND MEDIANS (selected numeric highlights)
- Variable importance scores (selected):
  - Baseline Rank 1: Nominal depreciation — Baseline Score 100
  - Baseline Rank 2: Profitability: Interest rate spread — Baseline Score 92
  - Baseline Rank 3: Inflation — Baseline Score 73
  - Baseline Rank 4: Liability dollarization — Baseline Score 45
  - Baseline Rank 5: Liquidity — Baseline Score 55
- Representative medians from terminal nodes:
  - Median nominal depreciation under “FX-risk” induced crises: 12 percent
  - Median nominal depreciation under “macroeconomic instability” crises: 35 percent (reported as 35.2 in table)
  - Median nominal depreciation under noncrisis outcome: 28 percent
  - Median liability dollarization under “FX-risk” induced crises: over 200 percent (reported as 201.8 in table)
  - Median liability dollarization under “macroeconomic instability” crises: 68 percent (reported as 67.5 in table)

### VII. POLICY IMPLICATIONS FOR SUPERVISION AND MONITORING
- Conditions associated with zero probability of banking crisis in baseline results:
  - Inflation less than 19 percent
  - Interest rate margin more than 3 percent
  - FX deposits to official FX reserves less than 140 percent
- Emphasis:
  - Monitor combinations of key indicators (inflation, profitability, FX exposure, liquidity) rather than single indicators in isolation.
  - FX risk (liability dollarization) amplifies the effect of nominal depreciation on crisis probability (twin-crisis mechanism).
  - Excessive credit activity relative to deposits (low liquidity) can be crisis-prone even with low FX exposure.

### VIII. “SEVERE” CRISES SPECIFICATION — definition and main results
- Severe crisis definition: banking crises associated with a recession or negative real GDP growth.
- Sample: reduced to 689 observations with 16 crisis episodes (unconditional probability described as a little over 2 percent).
- Severe-crises BCT performance: correctly classifies all crises but gives false alarms for 35 percent of noncrises episodes.
- Primary splitters for severe crises (higher thresholds than baseline): nominal depreciation, liability dollarization, bank illiquidity.
- Severe-crisis terminal nodes (selected):
  - Terminal node 1:
    - Condition: nominal depreciation over 9.8 percent (in absence of accompanying vulnerabilities)
    - Effect: almost triples probability of a severe crisis from 2.3 percent to 6.3 percent
    - Includes crises: Argentina (1990), Brazil (1990), Colombia (1999), Czech Republic (1991), Dominican Republic (2003), Jordan (1990), Kenya (1992), Lebanon (1990), Nicaragua (1990), Philippines (1998), Poland (1991), Turkey (1994), and Uruguay (2002)
  - Terminal node 2:
    - Condition: FX deposits to official FX reserves over 179 percent (even if nominal depreciation < 10 percent)
    - Conditional probability of crisis: 4 percent
    - Includes crises: Argentina (2001) and Bulgaria (1996)
  - Terminal node 3:
    - Condition: private credit to deposits exceeding 178 percent
    - Conditional probability of crisis: 100 percent
    - Includes crisis: Thailand (1997)

### IX. ROBUSTNESS CHECKS AND OUT-OF-SAMPLE PREDICTION
- Additional indicators tested:
  - Real exchange rate volatility (standard deviation of CPI-deflated real exchange rate over a four-year period)
  - Contagion dummy (1 if there was a banking crisis in the region in the previous year)
  - Lagged growth rate of stock market prices
- Result: classification tree almost identical to baseline; real exchange rate volatility (as outlier) replaces illiquidity at one crisis terminal node; contagion dummy and stock market growth add no contribution.
- History (past crises) test: variable measuring number of past crises per country added — no change in classification tree.
- Out-of-sample predictions (BCT estimated with data up to given year):
  - 2001 prediction: model predicts 3 of the 9 crises in 2001.
  - 2002 prediction: model predicts 4 of the 8 crises in 2002.
  - 2003 prediction: model predicts 2 of the 3 crises in 2003.
- Caribbean exclusion test:
  - BCT excluding 13 Caribbean countries correctly forecasts the three banking crises in the region (Guyana 1993, Jamaica 1994, Dominican Republic 2003).
  - Interpretation: Caribbean resilience largely due to macroeconomic (price) and exchange rate stability offered by fixed exchange rate regime.

### X. COMPARISON WITH TRADITIONAL LOGIT ANALYSIS — reconciliation with BCT findings
- Baseline logit (continuous variables) identifies significant average predictors:
  - Increase in inflation
  - Real domestic interest rate pressure
  - Increase in FX open position (NFA to GDP)
  - Decline in bank profitability
- Anomalous logit finding: increase in NPLs appears to reduce probability of a banking crisis (interpreted as rapid credit growth masking NPLs).
- Reconciliation:
  - Logit captures average effects across sample; BCT captures threshold-dependent, interaction effects.
  - Logit augmented with BCT-identified dummies (inequalities corresponding to BCT splits) improves fit and makes dummies for: inflation > 18.7 percent; nominal depreciation > 9.1 percent; FX deposits > 140 percent; interest rate spread < 3 percent — all statistically significant in increasing crisis probability.
  - Parsimonious logit with five BCT predictors plus terminal-node dummies: terminal-node dummies statistically significant; indicators significant mainly when crossing identified thresholds.

### XI. LIMITATIONS, CAVEATS, AND CONCLUSION
- Limitations:
  - Institutional indicators excluded due to lack of time series data (central bank independence, supervisory/regulatory strength, share of foreign ownership).
  - BCT findings may be localized to terminal nodes and require judgment for generalization.
- Methodological and practical implications:
  - BCT provides a perspective on nonlinear interactions and threshold effects among macroeconomic, FX, and FSI indicators.
  - BCT is robust to outliers and useful for monitoring banking sector vulnerabilities by identifying critical combinations and thresholds that standard regressions may miss.

*Source: _wp0893 - References*

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

### References

### I. INTRODUCTION
- Scope and motivation:
  - Banking crises in the 1990s included transition-economy meltdowns, the East Asian crises in 1997–99, and the Dominican Republic crisis of 2003.
  - Historically, banking crises have imposed huge fiscal costs of resolution and/or sharp output losses.
- Contribution:
  - Applies a binary classification tree (BCT) technique to analyze banking crises in a sample of 50 emerging market and developing countries during 1990–2005.
  - Explicitly considers financial soundness indicators (FSI) alongside macroeconomic and external indicators.
- Explanatory variable categories considered:
  - Macroeconomic environment: growth, inflation, nominal depreciation, government balance.
  - External vulnerability: official foreign exchange reserve (FX) cover of broad money, export growth, terms of trade (TOT) growth.
  - Monetary conditions: credit growth, real deposit rate, foreign interest rate, existence of explicit deposit insurance, de facto exchange rate regime.
  - Banking sector health: liability dollarization (FX deposits in total official FX reserves), net FX open position, bank liquidity, equity strength, asset quality, two proxies for bank profitability.

### Key empirical findings (Baseline BCT model)
- Five most important candidate variables identified (out of 19):
  - nominal depreciation
  - bank profitability
  - inflation
  - liability dollarization
  - bank liquidity
- Three key crisis-prone condition sets and associated probabilities:
  - Macroeconomic instability:
    - High annual inflation (greater than 19 percent) combined with relatively low TOT growth (less than 3¼ percent) increases probability of crisis from 5.3 percent to 21.4 percent.
  - Low bank profitability:
    - Low interest profitability (spread between lending and deposit rates of less than 3 percent) combined with modest export growth (less than 12 percent) increases probability of crisis to over 20 percent.
  - High foreign exchange (FX) risk:
    - Liability dollarization (FX deposits to official FX reserves more than 140 percent) combined with either:
      - relatively high depreciation (greater than 9 percent) → probability of crisis increases to 25 percent; or
      - low bank liquidity (private credit to deposits higher than 150 percent) → probability of crisis increases to 100 percent.
- Alternative “severe crises” specification (banking crises accompanied by recessions) identifies:
  - nominal depreciation (beyond 10 percent)
  - liability dollarization (beyond 179 percent of official FX reserves)
  - low bank liquidity (private credit to deposits more than 178 percent)
  - as key precursors to banking crises.

### II. LITERATURE SURVEY (selected synthesis)
- Two main traditional empirical approaches:
  - Signals approach (e.g., Kaminsky and Reinhart, 1999): contrasts indicator behavior before/after crises and identifies threshold-based signals.
  - Limited dependent variable regression approach (logit/probit) (e.g., Demirgüc and Detragiache, 1998): computes probability of crisis using multivariate models.
- Commonly identified correlates of banking crises in prior literature:
  - Low real GDP growth, high real interest rates, high inflation, low foreign reserve cover of broad money, positive credit growth, increased bank exposure to private sector.
  - Real GDP per capita (proxy for institutional development): higher per capita income lowers probability of crisis.
  - Existence of explicit deposit insurance: lowers probability of crises.
  - Bank-level FSI (e.g., return on assets, loan loss reserves to capital, loan growth, net interest income) are significant in micro-level studies.
- Mixed findings across studies on:
  - Role of domestic financial liberalization and rising interest rates.
  - Role of rapid credit growth as a precursor to crises.
- Motivation for BCT:
  - An indicator may be critical only at certain threshold values and/or when combined with other conditions; BCT captures such nonlinear, combinational effects.

### III. BINARY CLASSIFICATION TREE METHODOLOGY
- Overview of BCT:
  - Nonparametric technique that iteratively splits a parent node into child nodes by selecting variables and thresholds that increase homogeneity (crisis vs. noncrisis) in child nodes.
  - Splitting continues until stopping rules are met (pure nodes, single observations) or until trade-off between tree cost and fit warrants stopping.
- Model selection and validation:
  - Default algorithm: V-fold cross validation with 10 folds (sample divided into 10 roughly equal parts; largest possible tree built on 9/10 and tested on remaining 1/10; repeated so each part serves as test once).
  - Authors may overrule the default tree size to ensure in-sample fit and economic interpretability.
- Variable importance and masking:
  - “Variable importance index” ranks indicators by their overall discriminatory power, mitigating the “masking problem” where a variable with useful information may not appear in the final tree because it is outperformed at a split.
- Advantages of BCT:
  - No assumptions about functional form.
  - Captures nonlinearities and threshold effects.
  - Handles highly correlated candidate indicators and datasets with missing values (uses surrogates for primary splitters).
  - Robust to outliers and unaffected by monotonic transformations of indicators.
- Limitations of BCT:
  - Does not establish sample-wide general relationships; findings can be localized to terminal nodes.
  - Cannot provide marginal contributions to overall probability as logit/probit can.
  - Masking problem may still hide useful variables from appearing in the tree.
  - Optimal tree chosen by V-fold may require judgment-based pruning or growth for economic meaningfulness.

### Robustness checks and model performance
- BCT robustness examined via:
  - Addition of indicators for economic volatility, regional contagion pressures, and forward-looking information.
  - Out-of-sample prediction accuracy assessments.
  - Comparison with traditional logit model results.
- Conclusion from robustness analysis:
  - BCT performs reasonably well and is a useful tool for monitoring banking sector vulnerabilities given its ability to identify complex nonlinear relationships.

### Organization of the paper (as described)
- Section II: literature summary.
- Section III: BCT methodology and recent crisis literature using it.
- Section IV: empirical analysis — stylized facts and BCT results.
- Section V: robustness checks.
- Section VI: conclusion.

*Source: _wp0893 - References*

### References to other relevant studies using the BCT technique can be found in http://www.salford-

### _wp0893 - References to other relevant studies using the BCT technique can be found in http://www.salford-

### BCT literature and context
- Binary Classification Tree (BCT) used to assess the role of structural factors (governance, rule of law) relative to macroeconomic variables in explaining currency and banking crises.
- Manasse and Roubini (2005) use BCTs for sovereign debt crises, finding combinations of vulnerabilities (debt unsustainability, illiquidity, macroeconomic risks).
- Chamon, Manasse and Prati (2007) use BCT to analyze reversals in capital inflows and find crises are underpinned by combinations of weaknesses; low FX reserve cover combined with a high share of external debt to GDP significantly increases probability of capital account crises.
- Frankel and Wei (2004) also apply binary regression trees to currency crises.

### Empirical analysis — sample and crisis definition
- Sample: annual observations from 50 emerging market and developing countries during 1990–2005.
- Banking crisis definition: episode involving banking sector problems resulting in exhaustion of much of the capital and closure/merger/nationalization of banks; extensive bank runs; or large scale liquidity support by the central bank to avoid a run on deposits.
- Crisis observations identified: 127 annual crisis observations comprising 38 crisis episodes.
- Among 50 countries, 31 experienced at least one banking crisis during 1990–2005; crisis durations typically 2–3 years, with some exceeding six years (e.g., Ecuador and Jamaica).
- Control group: 19 countries did not experience any systemic banking crisis during the sample period.

### Explanatory variables used in BCT analysis
- Macroeconomic fundamentals:
  - Real GDP growth
  - Inflation
  - Nominal exchange rate depreciation
  - Fiscal overall balance (ratio to GDP) of the central government
- External liquidity:
  - Growth of exports of goods and services
  - Terms of trade (TOT) growth
  - Official foreign exchange (FX) reserve cover of broad money
- Monetary conditions:
  - Growth of real private credit
  - Real domestic interest rate
  - Foreign short-term interest rate (U.S. one-month LIBOR)
  - De facto fixed peg exchange rate regime
  - Existence of explicit deposit insurance
- Financial soundness indicators (FSIs):
  - Net foreign assets (NFA) to GDP (net open position)
  - FX deposits to official FX reserves (liability dollarization proxy)
  - Private credit to banking system deposits (liquidity condition)
  - Banking system equity over assets (capital adequacy)
  - Share of nonperforming loans (asset quality)
  - Interest rate spread (lending minus deposit rates) and pre-tax return over average assets (profitability)

### Stylized facts before and during crises (summary of Table 2)
- Macroeconomic fundamentals prior to banking crises (vs. tranquil times):
  - Real GDP growth and fiscal balances are lower
  - Inflation and nominal depreciation are higher
- External sector prior to crises:
  - Official FX cover of broad money, export growth, and TOT growth are lower
- Monetary conditions prior to crises:
  - Higher foreign interest and higher domestic real deposit rates
  - Real credit growth much higher in run-up to banking crisis
- Banking sector indicators prior to crises:
  - Banking system liquidity dries up (higher private credit to deposits)
  - Liability dollarization increases (higher FX deposits to reserves)
  - FX open positions widen
  - Larger share of countries had explicit deposit insurance before a banking crisis (consistent with moral hazard concerns)
  - De facto exchange rate regime rigidity is lower before a banking crisis (contrary to expectation)
  - Capital ratios and profitability are lower before a crisis
  - NPL ratio is relatively better before a crisis and worsens when crisis is underway (possible dampening effect of rapid credit growth on observed NPLs)

### Baseline BCT results — model performance and sample adjustments
- Indicator variables lagged (except foreign interest rate).
- To avoid post-crisis contamination, crisis observations after the first year of any given crisis episode are deleted, reducing sample size from 800 to 711 observations, with 38 crises.
- Unconditional crisis probability in reduced sample: 5.3 percent.
- Baseline tree: 8 terminal nodes.
  - In-sample fit: signals all but 1 of the 38 crises (97 percent) and has false signals for 19 percent of noncrisis outcomes.

### Key crisis-prone terminal nodes (baseline tree)
- Parent-node split:
  - Inflation threshold: 19 percent
    - Inflation > 19 percent → conditional probability of crisis increases from 5.3 percent to over 16 percent
    - Inflation < 19 percent → conditional probability of crisis declines to 2.8 percent
- Identified terminal nodes (with exact thresholds and conditional probabilities):
  - Terminal node 1:
    - Conditions: inflation greater than 18.7 percent AND TOT growth less than 3.3 percent
    - Conditional probability of crisis: over 21 percent
    - Comprises 22 crisis episodes, including: Argentina (1990), Brazil (1990, 1994), Bulgaria (1996), Colombia (1999), Ecuador (1996), Guyana (1993), Jamaica (1994), Jordan (1990), Kenya (1992), Lebanon (1990), Mexico (1992), Nicaragua (1990), Nigeria (1993), Peru (1993), Poland (1991), Tanzania (1990), Turkey (1991,1994, 2000), and Venezuela (1993)
  - Terminal node 2:
    - Conditions: interest rate spread less than 3.1 percent AND export growth less than 11.8 percent
    - Conditional probability of crisis: 21 percent
    - Comprises crises in: Argentina (1994, 2001), China (1999), Indonesia (1992, 1997), Korea (1997), Sri Lanka (1990), Malaysia (1997), and Thailand (1997)
  - Terminal node 3:
    - Conditions: FX deposits to official FX reserves more than 140 percent AND nominal depreciation more than 9 percent
    - Conditional probability of crisis: 25 percent
    - Contains crises in: Bolivia (1994), Czech Republic (1991), The Dominican Republic (2003), Nicaragua (2000), The Philippines (1998), and Uruguay (2002)
  - Terminal node 4:
    - Condition: private credit to deposits more than 150 percent
    - Conditional probability of crisis: 100 percent
    - Observation: only one outcome in this node (Croatia, 1996) — treated as an outlier

### Variable ranking and interpretation
- Top five ranked variables (out of 19 candidate indicators), in order:
  - Nominal depreciation
  - Interest profitability (interest rate spread / bank profitability)
  - Inflation
  - Liability dollarization (FX deposits to official FX reserves)
  - Bank liquidity (private credit to deposits)
- Interpretation highlights:
  - FX risk (liability dollarization) amplifies the effect of nominal depreciation on crisis probability (twin-crisis mechanism).
  - Excessive credit activity relative to deposits (low liquidity) can be crisis-prone even with low FX exposure.
  - Macroeconomic fundamentals (inflation, nominal depreciation) often dominate over micro FSIs in explaining crises, though select FSIs for FX risk, profitability, and liquidity remain key.
  - Conditional thresholds matter: some indicators are crisis-prone only when combined with other vulnerabilities.

### Examples of conditional thresholds and medians (from Table 4 commentary)
- Median nominal depreciation under “FX-risk” induced crises: 12 percent
- Median nominal depreciation under “macroeconomic instability” crises: 35 percent
- Median nominal depreciation under noncrisis outcome: 28 percent
- Median liability dollarization under “FX-risk” induced crises: over 200 percent
- Median liability dollarization under “macroeconomic instability” crises: 68 percent
- Example implication: a 12 percent nominal depreciation can induce banking crises in presence of high dollarization, whereas higher depreciation may not trigger a crisis in low-inflation, low-dollarization environments.

### Policy implication and guidance for bank supervision
- Conditions associated with zero probability of banking crisis in baseline results:
  - Inflation less than 19 percent
  - Interest rate margin more than 3 percent
  - FX deposits to official FX reserves less than 140 percent
- Emphasis: monitor combinations of key indicators (inflation, profitability, FX exposure, liquidity) rather than single indicators in isolation.

### “Severe” banking crises — alternative specification and findings
- Severe crisis definition: banking crises associated with a recession or negative real GDP growth.
- Sample reduced to 689 observations with 16 crisis episodes (unconditional probability described as a little over 2 percent).
- Severe-crises BCT performance: correctly classifies all crises but gives false alarms for 35 percent of noncrises episodes.
- Primary splitters for severe crises (higher thresholds than baseline): nominal depreciation, liability dollarization, bank illiquidity.
- Identified severe-crisis terminal nodes:
  - Terminal node 1:
    - Condition: nominal depreciation over 9.8 percent (in absence of accompanying vulnerabilities)
    - Effect: almost triples probability of a severe crisis from 2.3 percent to 6.3 percent
    - Includes crises: Argentina (1990), Brazil (1990), Colombia (1999), Czech Republic (1991), The Dominican Republic (2003), Jordan (1990), Kenya (1992), Lebanon (1990), Nicaragua (1990), The Philippines (1998), Poland (1991), Turkey (1994), and Uruguay (2002)
  - Terminal node 2:
    - Condition: FX deposits to official FX reserves over 179 percent (even if nominal depreciation < 10 percent)
    - Conditional probability of crisis: 4 percent
    - Includes crises: Argentina (2001) and Bulgaria (1996)
  - Terminal node 3:
    - Condition: private credit to deposits exceeding 178 percent
    - Conditional probability of crisis: 100 percent
    - Includes crisis: Thailand (1997)
- Table 3 (third and fourth columns) confirms nominal depreciation, liability dollarization and liquidity among top-ranked variables for severe crises; inflation, profitability, and real credit growth ranked high though may not appear in the severe-crisis tree.

### Robustness checks
- Alternative indicators added:
  - Real exchange rate volatility (standard deviation of CPI-deflated real exchange rate over a four-year period)
  - Contagion dummy (1 if there was a banking crisis in the region in the previous year)
  - Lagged growth rate of stock market prices (forward-looking information)
- Result: classification tree almost identical to baseline; real exchange rate volatility (as outlier) replaces illiquidity at crisis terminal node 4; contagion dummy and stock market growth add no contribution to model fit.
- History (past crises) test:
  - Variable measuring number of past crises per country added
  - Result: no change in classification tree; past crisis variable does not contribute to overall fit
- Out-of-sample predictive performance:
  - BCT estimated with data up to 2000, 2001, and 2002 to predict crises in 2001, 2002, and 2003 respectively
  - Out-of-sample predictions (years 2001–03) performance:
    - 2001: model predicts 3 of the 9 crises in 2001
    - 2002: model predicts 4 of the 8 crises in 2002
    - 2003: model predicts 2 of the 3 crises in 2003
- Caribbean out-of-sample exercise:
  - Caribbean in sample: 13 countries; only three systemic banking crises in sample period — Guyana (1993), Jamaica (1994), Dominican Republic (2003)
  - Purpose: assess whether absence of crises in Caribbean was due to fundamentals or luck
  - Result: out-of-sample forecast for Caribbean is good — model correctly signaled all new crises that occurred in the Caribbean during the sample period

*Source: _wp0893 - References to other relevant studies using the BCT technique can be found in http://www.salford- (IMF working paper content provided).*

### 2002. The classification tree produces three terminal crisis nodes similar to the original

### _wp0893 - 2002. The classification tree produces three terminal crisis nodes similar to the original

### Out-of-sample prediction for 2002
- The classification tree for 2002 produces three terminal crisis nodes similar to the baseline specification:
  - Terminal node 1: high inflation combined with modest TOT growth.
  - Terminal node 2: low interest profitability and modest export growth.
  - Terminal node 3: two measures of high FX risk—high liability dollarization combined with a wide FX open position (relatively low share of NFA to GDP).
- Model performance and specific outcomes:
  - The model successfully predicts 4 out of the 8 crises in 2002: Argentina, Ecuador, Korea, and Turkey.
  - Missed crises in 2002: Indonesia, Philippines, Thailand, and Uruguay.
  - Uruguay is marginally missed at terminal node 3, where Uruguay’s NFA to GDP ratio is higher than the threshold of 3 percent identified by the tree.
  - The tree signals a banking crisis in Nicaragua, although Nicaragua had just come out of a banking crisis in 2001.

### Out-of-sample prediction for 2003
- BCT model based on 1990–2002 yields three terminal nodes for 2003 prediction:
  - Terminal node 1: high inflation combined with modest TOT growth.
  - Terminal node 2: low interest profitability and modest export growth.
  - Terminal node 3: high liability dollarization combined with relatively high illiquidity.
- Model performance for 2003:
  - Predicts 2 of the 3 crises in 2003: the ongoing crisis in Uruguay and the new crisis in the Dominican Republic.
  - Misses the ongoing crisis in the Philippines.
  - Continues to signal crises in Argentina and Turkey, even though these countries exited their banking crises in 2002.

### Out-of-sample prediction for the Caribbean (exclusion test)
- BCT based on 1990–2005 that excludes 13 Caribbean countries produces three crisis-prone terminal nodes and correctly forecasts the three banking crises in the region:
  - Terminal node 1: highly inflationary state; predicts Guyana (1993) and Jamaica (1994).
  - Terminal node 2: low interest profitability combined with relatively modest export growth; this node signals a banking crisis in The Bahamas for the entire sample period (noted as odd).
  - Terminal node 3: high liability dollarization combined with nominal depreciation; correctly signals the Dominican Republic banking crisis in 2003.
- Interpretation:
  - Results stress that a key reason underpinning the resilience of Caribbean banking systems is macroeconomic (price) and exchange rate stability offered by the fixed exchange rate regime.

### Consistency of BCT results with traditional logit analysis
- Baseline logit model (column (i)) — variables that significantly increase the probability of crises:
  - Increase in inflation.
  - Real domestic interest rate pressure.
  - Increase in FX open position.
  - Decline in bank profitability.
- Anomalous finding in baseline logit:
  - An increase in NPLs appears to reduce the probability of a banking crisis (interpreted as rapid credit growth masking asset quality problems).
- Baseline logit in-sample fit:
  - Able to correctly classify 82 percent of crises.
- Discrepancy with BCT:
  - Other than inflation, key BCT splitters (liability dollarization, interest rate spread, liquidity) are not significant in the baseline logit, and interest profitability has the opposite sign.
  - Reconciliation: logit identifies variables that trigger crises on average; BCT recognizes indicators signal crises only after crossing particular thresholds and in combination.
- Logit specification including BCT-identified threshold dummies (column (ii)):
  - Inequalities/dummies added corresponding to primary splitters in the BCT baseline model:
    - inflation > 18.7 percent
    - nominal depreciation > 9.1 percent
    - interest rate spread > 3.1 percent
    - liability dollarization > 140 percent of official FX reserves
    - liquidity (private credit > 150 percent of deposits)
    - export growth > 11.8 percent
    - TOT growth > 3.3 percent
  - Resulting findings (column (ii)):
    - Dummies corresponding to an increase in: inflation above 18.7 percent, nominal depreciation beyond 9 percent, FX deposits beyond 140 percent of official FX reserves, and decline in interest rate spread below 3 percent are all statistically significant in increasing the probability of crisis.
    - Inclusion of these dummies improves statistical significance of other indicators (official FX reserve cover of broad money, real private credit growth, equity level).
    - Improved in-sample fit: identifies 94 percent of the crisis episodes.
- Parsimonious logit specification (column (iii)):
  - Includes only the five key predictors from the BCT analysis and dummy variables for the crisis terminal nodes.
  - Findings:
    - All terminal node dummies are statistically significant.
    - The five key indicator variables are either not statistically significant or have the wrong sign.
  - Interpretation:
    - Predictors of banking crisis are significant only when they cross the identified thresholds, not at all values—confirming the importance of BCT in identifying critical thresholds.

### Key indicators, thresholds, and interactions emphasized by the BCT
- Financial market condition indicators identified as important predictors:
  - Nominal depreciation
  - Interest profitability
  - Inflation
  - Liability dollarization
  - Bank illiquidity
- Crisis triggers are combinations of vulnerabilities rather than single factors.
- Foreign exchange risk emerges as a primary vulnerability:
  - Banks become crisis-prone when faced with nominal depreciation combined with high liability dollarization.
  - Banks can be crisis-prone even with limited depreciation if high liability dollarization coincides with low bank liquidity.
- Conditional thresholds matter more than unconditional variable levels: indicators trigger banking sector problems only after crossing specific thresholds.

### Limitations and caveats
- Institutional indicators excluded due to lack of time series data:
  - Central bank independence
  - Strength of supervisory and regulatory framework
  - Share of foreign ownership of the banking system

### Conclusion (methodological and practical implications)
- The Binary Classification Tree (BCT) methodology:
  - Provides a new perspective on banking crises by identifying nonlinear interactions and threshold effects among macroeconomic, foreign exchange, and financial soundness indicators.
  - Is robust to outliers among explanatory variables.
  - Is a unique and useful tool to monitor bank vulnerabilities, as it identifies critical combinations and thresholds that standard regressions may miss.

*Source: _wp0893 - 2002. The classification tree produces three terminal crisis nodes similar to the original*

### 19. De facto exchange rate regime: Annual Report on Exchange Arrangements and Exchange

### De facto exchange rate regime: Annual Report on Exchange Arrangements and Exchange Restrictions, various issues (IMF)

### Overview and sample
- Sample period: 1990–2005.
- Total observations reported in baseline classification tree: 711.
- Number of crises in baseline: 38.
- Prob. of crisis (baseline sample): 5.3%.

### Banking crisis episodes (Table 1)
- Argentina: 1990, 1994–95, 2001–02 — Average Years in Crisis 1.7
- Bolivia: 1994–95 — 2.0
- Brazil: 1990, 1994–99 — 3.5
- Bulgaria: 1996–97 — 2.0
- China: 1999 — 1.0
- Colombia: 1999–00 — 2.0
- Croatia: 1996 — 1.0
- Czech Republic: 1991–95 — 5.0
- Dominican Republic: 2003–05 — 2.0
- Ecuador: 1996–02 — 7.0
- Guyana: 1993–95 — 3.0
- Indonesia: 1992,1997–02 — 3.5
- Jamaica: 1994–00 — 7.0
- Jordan: 1990 — 1.0
- Kenya: 1992–95 — 4.0
- Korea: 1997–02 — 6.0
- Lebanon: 1990 — 1.0
- Malaysia: 1997–01 — 5.0
- Mexico: 1992–97 — 6.0
- Nicaragua: 1990–96, 2000–01 — 4.5
- Nigeria: 1993–95 — 2.0
- Paraguay: 1995–00 — 5.0
- Peru: 1993–94 — 2.0
- Philippines: 1998–03 — 6.0
- Poland: 1991–95 — 5.0
- Sri Lanka: 1990–93 — 1.0
- Tanzania: 1990–91 — 2.0
- Thailand: 1997–02 — 6.0
- Turkey: 1991, 1994, 2000–02 — 1.7
- Uruguay: 2002–03 — 2.0
- Venezuela: 1993–95 — 3.0
- Countries with no crisis during the sample period include Antigua and Barbuda, The Bahamas, Barbados, Belize, Chile, Costa Rica, Dominica, Egypt, El Salvador, Grenada, Guatemala, India, Israel, Papua New Guinea, South Africa, St. Kitts and Nevis, St. Lucia, St. Vincent and the Grenadines, Trinidad and Tobago.

### Key averages and comparisons (Table 2) — selected indicators (sample averages in percent unless noted)
- GDP growth: 3.4 (sample average)
  - No Crisis in t / No Crisis in t+1: 3.8
  - No Crisis in t / Crisis in t+1: 1.6
  - Crisis in t / Crisis in t+1: 1.2
- Inflation: 73.9 (sample average)
  - No Crisis in t / No Crisis in t+1: 39.4
  - No Crisis in t / Crisis in t+1: 30.2
  - Crisis in t / Crisis in t+1: 55.3
- Nominal depreciation: 100.3 (sample average)
  - No Crisis in t / No Crisis in t+1: 77.6
  - No Crisis in t / Crisis in t+1: 80.2
  - Crisis in t / Crisis in t+1: 248.0
- Official FX reserves (to broad money): 30.9 (sample average)
  - No Crisis in t / No Crisis in t+1: 31.7
  - No Crisis in t / Crisis in t+1: 26.9
  - Crisis in t / Crisis in t+1: 28.4
- Export growth: 9.4 (sample average)
  - No Crisis in t / No Crisis in t+1: 9.9
  - No Crisis in t / Crisis in t+1: 7.0
  - Crisis in t / Crisis in t+1: 8.1
- Growth of real private credit: 19.3 (sample average)
  - No Crisis in t / No Crisis in t+1: -1.1
  - No Crisis in t / Crisis in t+1: 187.4
  - Crisis in t / Crisis in t+1: 268.5
- Real deposit rate: 6.1 (sample average)
  - No Crisis in t / No Crisis in t+1: -13.2
  - No Crisis in t / Crisis in t+1: 438.3
  - Crisis in t / Crisis in t+1: -152.3
- Existence of deposit insurance: 44.8 (sample average)
  - No Crisis in t / No Crisis in t+1: 40.3
  - No Crisis in t / Crisis in t+1: 47.4
  - Crisis in t / Crisis in t+1: 55.2
- Average rigidity of de facto exchange rate regime: 2/ = 2.0 (sample average)
  - Definition: 1 (floating), 2 (soft pegs), 3 (fixed or hard pegs).
- Liquidity: Private sector credit (to deposits): 88.0 (sample average)
  - No Crisis in t / No Crisis in t+1: 85.3
  - No Crisis in t / Crisis in t+1: 97.5
  - Crisis in t / Crisis in t+1: 100.0
- Liability dollarization: FX deposits (to official FX reserves): 92.3 (sample average)
  - No Crisis in t / No Crisis in t+1: 87.4
  - No Crisis in t / Crisis in t+1: 126.5
  - Crisis in t / Crisis in t+1: 164.0
- FX open position: Net foreign assets (to GDP): 6.1 (sample average)
  - No Crisis in t / No Crisis in t+1: 9.4
  - No Crisis in t / Crisis in t+1: -14.3
  - Crisis in t / Crisis in t+1: -24.2
- Capital: Equity (to assets): 8.7 (sample average)
  - No Crisis in t / No Crisis in t+1: 9.1
  - No Crisis in t / Crisis in t+1: 8.1
  - Crisis in t / Crisis in t+1: 7.4
- Profitability: Return over average asset: 1.1 (sample average)
  - No Crisis in t / No Crisis in t+1: 1.4
  - No Crisis in t / Crisis in t+1: 0.6
  - Crisis in t / Crisis in t+1: 0.5
- Profitability: Interest rate spread: 15.3 (sample average)
  - No Crisis in t / No Crisis in t+1: 17.0
  - No Crisis in t / Crisis in t+1: -0.8
  - Crisis in t / Crisis in t+1: 3.0
- Asset quality: NPLs (to total loans): 8.0 (sample average)
  - No Crisis in t / No Crisis in t+1: 7.4
  - No Crisis in t / Crisis in t+1: 5.0
  - Crisis in t / Crisis in t+1: 8.6

### Variable importance (Table 3) — ranking and scores (selected)
- Baseline Rank 1: Nominal depreciation — Baseline Score 100
- Baseline Rank 2: Profitability: Interest rate spread — Baseline Score 92
- Baseline Rank 3: Inflation — Baseline Score 73
- Baseline Rank 4: Liability dollarization (FX deposits to official reserves) — Baseline Score 45
- Baseline Rank 5: Liquidity — Baseline Score 55
- De facto exchange rate regime: Baseline Rank 10 — Baseline Score 17
- Severe Crises Rank 1: Nominal depreciation — Severe Crises Score 100
- Severe Crises Rank 2: Profitability: Interest rate spread — Severe Crises Score 64
- Severe Crises Rank 3: Inflation — Severe Crises Score 26
- Severe Crises Rank 4: Liability dollarization — Severe Crises Score 36
- Severe Crises Rank 5: Liquidity — Severe Crises Score 47

### Binary classification tree highlights (Figures 2–7)
- Baseline splitting rules frequently include thresholds:
  - Inflation ≤ 18.7% or > 18.7%.
  - Terms of trade (TOT) growth ≤ 3.3% or > 3.3%.
  - Interest rate margin ≤ 3.1% or > 3.1%.
  - FX deposits ≤ 140% (or 139%/179% in regional variants) of official FX reserves.
  - Nominal depreciation ≤ 9.1% (or 9.8%).
  - Illiquidity (credit-to-deposits) thresholds used: 150%, 127.8%, 79%, 178%.
  - Export growth thresholds: 11.8% and 12.1%.
- Example terminal node conditional probabilities (Figure 2 / Table 4):
  - Terminal node with Inflation > 18.7% & TOT Growth <= 3.3%: Conditional probability of crisis 21.4 percent.
    - Median nominal depreciation: 35.2
    - Median interest rate spread: 13.0
    - Median inflation: 44.7
    - Median liability dollarization index: 67.5
    - Median illiquidity index: 74.7
  - Terminal node with Inflation < 18.7% & Interest Rate Margin <= 3.1% & Export Growth <= 11.8%: Conditional probability of crisis 20.9 percent.
    - Median nominal depreciation: 0.0
    - Median interest rate spread: 2.1
    - Median inflation: 5.4
    - Median liability dollarization index: 20.0
    - Median illiquidity index: 103.3
  - Terminal node with Inflation < 18.7% & Interest Rate Margin > 3.1% & Liability Dollarization > 140% & Nominal Depreciation > 9%: Conditional probability of crisis 25 percent.
    - Median nominal depreciation: 11.8
    - Median interest rate spread: 10.6
    - Median inflation: 10.8
    - Median liability dollarization index: 201.8
    - Median illiquidity index: 96.0
  - Terminal node with Inflation < 18.7% & Interest Rate Margin > 3.1% & Liability Dollarization > 140% & Nominal Depreciation <= 9% & Illiquidity Index > 150%: Conditional probability of crisis 100 percent.
    - Median nominal depreciation: -12.8
    - Median interest rate spread: 14.7
    - Median inflation: 4.0
    - Median liability dollarization index: 140.5
    - Median illiquidity index: 158.2

### Logit analysis results (Table 5) — selected coefficient signs, significance, and predictive performance
- Significant positive predictors (selected; p-values in parentheses):
  - Dummy (inflation higher > 18.7%): Coefficient 3.992 (0.03)**.
  - Dummy (nominal depreciation > 9.1%): Coefficient 5.130 (0.01)**.
  - Dummy (dollarization of liabilities > 140% of official FX reserves): Coefficient 4.462 (0.00)**.
  - Profitability: return over average asset: Coefficients -0.695 and -0.640 (0.00)** and (0.03)** (negative sign — higher profitability lowers crisis probability).
  - Asset quality: NPLs to total loans: Coefficient -0.156 (0.02)** (negative sign).
  - Illiquidity dummy (private sector credit to deposits>150%): Coefficient 1.574 (0.35) [not significant at 5/10 percent].
  - Interest profitability dummy (Interest rate spread > 3.1 percent): Coefficient -6.083 (0.00)**.
- Examples of coefficient magnitudes and significance:
  - Nominal depreciation (continuous): coefficient reported as -0.144 with (0.01)** in one specification, but negative/insignificant in others; dummy for large depreciation (>9.1%) is strongly positive and significant.
  - Growth of real private credit: coefficient 0.081 (0.01)** in specification (iii).
  - Net FX open position: NFA to GDP: coefficient -0.033 (0.06)* in one specification.
- Model fit and prediction:
  - Number of observations: 452 and 595 across specifications.
  - Number of crises: 17 and 24 across specifications.
  - Psuedo R-squared: 0.31 and 0.54 and 0.40 across reported specifications.
  - Significance (Chi-square): 239.3**, 3601.9**, 54.14**.
  - Prediction performance (selected):
    - % of crises predicted correctly: 82.4 percent; 94.1 percent; 91.7 percent (across specifications).
    - % of non-crises predicted correctly: 78.1 percent; 89.1 percent; 84.6 percent (across specifications).

### Diagnostic groupings and thematic vulnerabilities (Table 4 labels)
- Four main outcomes/diagnostic groupings identified:
  - Macroeconomic Instability (e.g., high inflation, weak TOT growth).
  - Low Bank Profitability.
  - High FX Risk with Nominal Depreciation (high liability dollarization combined with large nominal depreciation).
  - High FX Risk with Low Liquidity (high liability dollarization combined with high illiquidity).

*Source: Authors’ calculations, as reported in the IMF working paper chapter on de facto exchange rate regime and the Annual Report on Exchange Arrangements and Exchange Restrictions (figures and tables reproduced in source content).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp0893.pdf_