## Classification of Capital Account Crises (KAC) and Control Group (CG) Episode

## Source details

**Canonical URL:** [Classification of Capital Account Crises (KAC) and Control Group (CG) Episode](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0675.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0675.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0675.pdf.json)

---

### Overview and purpose
- Examines whether IMF-supported programs, conditional on episodes of intense market pressures, can help prevent a capital account crisis (avoidance of abrupt private capital outflows rather than crisis resolution).
- Uses cluster analysis to identify episodes of intense market pressures and classifies them into capital account crises (KAC) or control group (CG) episodes; applies a logit specification to test whether an episode of intense market pressure develops into a KAC or a CG after controlling for initial conditions, policy changes, and exogenous factors.
- Focuses on available IMF resources (disbursements or accumulated drawing rights under on-track precautionary programs) rather than program dummies or lending commitments.

### Data universe and sample
- Monthly data: 1994 to 2004.
- Countries (27): Algeria, Argentina, Brazil, Bulgaria, Chile, Colombia, Dominican Republic, Ecuador, Hungary, Indonesia, Korea, Malaysia, Mexico, Morocco, Pakistan, Panama, Peru, Philippines, Poland, Russia, South Africa, Thailand, Tunisia, Turkey, Ukraine, Uruguay, and Venezuela.
- High-intensity market pressure episodes identified: 32 (cluster 1).
- Classified into KAC and CG using net private capital flows (excluding FDI) over four quarters following pressure event:
  - KAC episodes: 11
  - CG episodes: 21
- Dataset for regressions: 32 pressure episodes × 4 quarters = 128 observations (results reported also for 32-observation averaging).

### Identification methodology — two-step cluster approach
- Step 1 — Index of market pressures (IMP) constructed from country-standardized components:
  - IMPit = (−ln(FXit/FXi,t−1)) + (−ln(REERit/REERi,t−1)) + ln(Sit/Sit−1)
  - Components standardized (mean = 0, standard deviation = 1) per country.
  - Interpretation: increases in FX reserves or currency appreciation indicate easing pressures; higher spreads indicate rising pressures.
- Step 2 — Private capital flows clustering (quarterly, net of FDI, percent of GDP) into five clusters: HI=1 through HO=5.
  - KAC defined: pressure event followed by at least two quarters of medium or high capital outflows (clusters 4 or 5) during the four quarters beginning in the quarter of the pressure event.
  - Persistence rule: at least two quarters of outflows within the t to t+3 window.
- Distinct events separated by at least 12 tranquil months.

### Stylized facts (medians) — market/financial and macro indicators
- Index of market pressures: peaks at time t in both groups; index higher for KAC events.
- Spreads and residual spreads: higher in level for KAC episodes and the gap widens after t.
- Interest rates:
  - Treasury Bill (percent a year) levels are higher during t+1 to t+4 for KAC cases.
- Real Effective Exchange Rate (t=100): REER index dips for both groups post-t, larger dip for KAC.
- Reserves (percent of GDP):
  - Pre-t: about 10 percent of GDP for both groups.
  - KAC cases experience a significant drop in period t and recover a few quarters after onset.
- Real GDP growth (percent a year): somewhat stronger in the control group; difference widens after t; both groups show a “V-shape” recovery.
- Inflation (percent a year): spikes in the KAC group after t.
- Current account balance (percent of GDP): broadly same before t; current account adjustment after market pressures much stronger among KAC events.
- Fiscal balance (percent of GDP): worse among KAC episodes prior to onset; fiscal balance in KAC group peaks at time t-1.
- Broad money velocity (annual percent change): stable prior to t; monetary policy appears to loosen in KAC aftermath (short-lived).

### External sector and solvency/liquidity indicators (medians)
- Exchange rate regime: control group slightly more flexible than crisis group (AREAER-based).
- Exchange rate overvaluation: CG less overvalued than KAC (measured as difference between actual REER and HP-filter trend).
- Private capital flows (percent of GDP): similar until t; by definition, KAC shows intense outflows after t.
- Total external debt (percent of GDP): about 55 percent in KAC cases, slightly over 10 percentage points higher than CG.
- Short-term debt ratios:
  - Short-term debt to reserves stable in CG events.
  - Rises sharply in KAC events from period t-1 to t+2, declines from t+3.

### Regressors and estimation strategy
- Dependent variable: logit Pr(KAC = 1) versus CG = 0 at period t.
- Regressor groups:
  - Initial Conditions: Total external debt-to-GDP; Short-term debt-to-reserves; Exchange rate regime prior to t; Exchange rate overvaluation; Political stability (ICRG index).
  - Policy Adjustment: Fiscal adjustment in pre-crisis period; Change in short-term real interest rates (monetary tightening).
  - Exogenous/Other Factors: Terms of trade movements; Size of economy; Regional dummies.
  - IMF Financing: ratio of available IMF resources (disbursed + accumulated drawing rights under precautionary arrangements) summed over t-4 to t-1 divided by short-term debt in t-1 (IMF financing defined as a share of short-term debt).
- Focus on available IMF resources to capture immediate liquidity and signaling/credibility.

### Sample balance and IMF financing descriptive statistics (Table 2 summary)
- IMF financing defined as share of short-term debt (STD).
- Full sample (If IMF financing exists):
  - Mean: 0.055
  - Median: 0.048
  - Std. Dev.: 0.038
  - Min. (Uruguay; 2001Q4): 0.008
  - Max. (Russia; 1997Q3): 0.120
  - Obs.: 1784
  - Share of the full sample of KACs: 39
- LAC region (If IMF financing exists):
  - Mean: 0.050
  - Median: 0.045
  - Std. Dev.: 0.042
  - Min.: 0.008
  - Max.: 0.118
  - Obs.: 416
  - Share of the full sample of KACs: 50
- Full sample when IMF financing may not exist (All episodes; second panel):
  - Mean: 0.021
  - Median: 0.000
  - Std. Dev.: 0.036
  - Min.: 0.000
  - Max.: 0.120
  - Obs.: 416
- KAC episodes (when IMF financing exists):
  - Mean: 0.085
  - Median: 0.052
  - Std. Dev.: 0.072
  - Min. (Argentina; 1998Q2): 0.008
  - Max. (Brazil; 2002Q2): 0.234
  - Obs.: 22
  - Share of the full sample of CGs: 26
- CG episodes (when IMF financing exists):
  - Mean: 0.022
  - Median: 0.000
  - Std. Dev.: 0.052
  - Min.: 0.000
  - Max.: 0.234
  - Obs.: 48
- Sample balance notes:
  - About one-third of observations are KACs and two-thirds are CG.
  - About 40 percent of KAC episodes have IMF resources available prior to t; slightly over 25 percent among CG episodes have IMF resources available prior to t.
  - Only two precautionary arrangements in sample; excluding them does not affect results.

### Main econometric findings (overview of specifications R1–R6)
- Regression designs:
  - R1: 32 observations (averaged t-4:t-1).
  - R2: 128 observations (quarterly).
  - R3: on-track IMF program dummy.
  - R4: on-track dummy + IMF financing regressor.
  - R5: IMF financing regressor only.
  - R6: interacts fiscal and monetary changes with on-track IMF dummy; IMF financing regressor based on available resources.
- Significant and robust determinants of higher crisis probability:
  - High external debt-to-GDP — statistically significant.
  - High short-term debt-to-reserves — statistically significant.
  - Exchange rate overvaluation — “quite important” determinant.
  - Size of the economy — contributes significantly to crisis propensity.
- Policy variables:
  - Monetary tightening (higher real interest rates) and fiscal tightening lower crisis probability; monetary variable significant at the 10 percent level in some specifications.
  - In R6, fiscal adjustment interactive with IMF dummy is negative and significant — greater fiscal adjustment in IMF-financed countries helps avert crises; combined fiscal adjustment term for program countries is negative and jointly significant.
- Political stability generally reduces crisis probability (coefficient negative), but not always significant.
- Regional dummies and terms of trade not consistently significant.

### Crisis prevention effects of IMF financing — key quantitative conclusions
- IMF financing coefficient (selected reporting from Table 3): 3/-25.36 **, -20.58 **, -37.23 **, -40.25 **, -40.06 ***
- IMF program dummy (selected reporting from Table 3): 4/-1.11, 0.46
- Main inferences:
  - The coefficient on the IMF financing to short-term debt regressor is negative and statistically significant across specifications. "IMF money matters."
  - IMF program dummy is consistently not significant — signaling appears related to the size of IMF financing rather than mere existence of an on-track program.
  - IMF financing effects persist after controlling for gross reserves, implying benefits beyond pure liquidity (policy/credibility channels).
- Model performance (selected):
  - No. of observations: 32, 128, 128, 128, 128, 128 across R1–R6.
  - LR Chi-square: 20.2 **, 30.6 ***, 50.7 ***, 25.7 ***, 30.6 ***, 65.8 ***.
  - Pseudo R-square: 0.51, 0.48, 0.53, 0.59, 0.59, 0.61.
  - Correctly classified (percent): 81, 80, 83, 87, 87, 87.
  - Type I errors (percent): 36, 34, 23, 16, 16, 18.
  - Type II errors (percent): 10, 13, 14, 12, 12, 11.
  - R6 has highest pseudo R-square and correctly classifies 87 percent of observations; R6 correctly classifies all KAC episodes at least one quarter prior to onset (no Type I errors in period t-1).

- Selected coefficient excerpts (exact values preserved):
  - Debt/GDP: 0.16 ***, 0.13 ***, 0.14 ***, 0.18 ***, 0.18 ***, 0.19 ***.
  - ST debt/Reserves: 0.87 ***, 0.75 ***, 1.06 ***, 1.06 ***, 1.03 **, 1.12 ***.
  - Exchange rate regime: -0.52, -0.40, -0.33, -0.56, -0.53, -0.67 *.
  - Political stability: -5.46, -5.26, -6.44 **, -6.14, -6.04 *, -6.05 *.
  - Exchange rate overvaluation: 7/17.65 ***, 25.22 ***, 24.74 ***, 26.29 ***.
  - Fiscal balance change: 8/-0.08 **, -0.03, -0.03, 0.23 **.
  - Fiscal balance interactive with Fund dummy: 9/-0.45 **.
  - Interest rate change (real terms): 10/-0.08 **, -0.07 *, -0.07 *, -0.03.
  - Size of the economy: 12/1.16 ***, 1.01 ***, 0.77 ***, 0.87 ***, 0.88 ***, 0.93 ***.
  - Latin American dummy: -1.43, -1.15, -0.69, -1.49, -1.51, -1.98.
  - Asian dummy: 2.71, 2.35, 2.41, 2.45, 2.34, 2.36.
- Model notes:
  - ***, **, and * indicate significance at the 1, 5, and 10 percent levels.
  - Standard errors adjusted for within-cluster correlation at pressure-episode level.
  - Random effects logit give similar results.

### Model classification performance (Table 4 statistics)
- Correctly classified KAC episodes (number of quarterly episodes): 37.
- Number of KAC episodes wrongly classified as CG in t-4 to t-1 (Type I error): 8.
- Number of KAC episodes wrongly classified as CG in t-1 (Type I error): 0.
- Correctly classified CG episodes (number of quarterly episodes): 74.
- Number of CG episodes wrongly classified as KAC in t-4 to t-1 (Type II error): 9.
- Number of CG episodes wrongly classified as KAC in t-1 (Type II error): 2.
- Correctly classified (percent): 87.
- Two instances where model incorrectly classifies an event as a KAC in period t-1: Mexico in 1994 and Turkey in 1998.

### Robustness tests and sample sensitivity
- Reclassifying Mexico 1994 and Turkey 1998 as KAC (model predicts KAC) does not alter IMF financing result.
- Adding Argentina (early 1995) as a 33rd KAC event does not affect IMF resources role.
- Dropping single-appearance KAC countries (Korea, Malaysia, Thailand) does not change IMF financing result; Indonesia and Philippines retained.
- Controlling for repeat-episode countries (dummy) leaves IMF financing effect unchanged.
- Dropping regressors one at a time or estimating parsimonious models leaves IMF financing role robust.
- dfbetas identification and dropping outliers does not change IMF financing result; dropping any single episode does not change the conclusion.
- Small sample noted: sometimes dropping outliers precludes maximum likelihood convergence; two strategies used to restore convergence produced unchanged IMF financing findings.
- Classification cutoff used: default probability = 0.5.

### Scenarios and marginal effects (Figure 7 conclusions)
- IMF disbursements (or accumulated drawing rights) over the 12 months prior to a pressure event reduce the probability a crisis will develop; program existence or commitments alone do not show same effect.
- Marginal effect depends on country fundamentals (external debt, exchange rate regime, short-term debt-to-reserves, policies).
- Significant drops in crisis probability can often be achieved by providing IMF financing, often requiring fewer resources than fully rolling over a country’s short-term debt.
- In some cases (example: Indonesia), plausible IMF resource levels had no meaningful effect.
- Resources needed to significantly reduce crisis probability generally exceed the average IMF financing levels typically provided to KAC cases prior to a crisis.
- Reducing crisis probability to a 5 percent cutoff requires exceptional access (i.e., more than 100 percent of quota) in 8 out of 11 KAC episodes.
- Average crisis probability among KAC episodes in period t-1 (with and without financing) was 87 percent.

### Policy-strengthening versus financing (parametric comparison)
- To achieve an average 20 percent reduction in crisis probability (observed average decline among CG events receiving IMF financing), R6 parametrically implies:
  - A combined increase in fiscal adjustment and tightening of monetary policy of close to 4½ percentage points each would be required (keeping IMF financing constant).
  - A reduction in exchange rate overvaluation by 6 percent produces a similar 20 percent reduction.
- These comparisons underscore that IMF financing operates through both liquidity and policy/credibility channels and that large policy adjustments or substantial resource provision may be substitutes in effect.

### Welfare gains from crisis prevention (back-of-the-envelope)
- Premise: expected decline in crisis probability from IMF financing equals average decline observed among CG events receiving IMF financing in t-4 to t-1 = 0.21 (21 percent).
- Uses discounted difference in output flows between KAC and CG over first three years following market pressure event; discount rate assumed = 5 percent.
- Table 5 headline result: estimated lower bound welfare gains = 5.6 percent of precrisis (t-4:t-1) GDP saved during the three years that follow a KAC.
- Method caveats:
  - Lower bound because differences in output flows between KAC and CG are large and avoiding sharp declines could yield substantially higher gains.
  - Assumes other covariates held constant; does not control for all differences between CG and KAC events.

### Policy implications and operational considerations
- Main policy-relevant conclusions:
  - IMF disbursements or available drawing rights under precautionary arrangements over the 12 months prior to intense market pressure materially reduce crisis probability; markets respond to actual or immediately available IMF money.
  - IMF resources are not fully substitutable by gross foreign exchange reserves; IMF resources provide both liquidity and a credibility signal reinforced by conditionality and the IMF “putting its own resources at risk.”
  - IMF involvement may be insufficient when fundamentals are weak.
- Operational considerations:
  - A more pro-active prevention role by the IMF could yield important welfare gains — making resources available before crises erupt.
  - Substantial front-loading of resources may be required where short-term debt coverage is low; front-loading raises risk of policy moral hazard.
  - IMF support should balance signaling to markets and ensuring that members benefit from IMF discipline and conditionality.

### Limitations and caution
- Small sample: 32 high market pressure episodes, only 11 full KACs; limits assessment of nonlinearities and contagion thresholds.
- Counterfactual establishment remains challenging despite balanced sample and robustness checks.
- IMF financing regressor may capture stronger policies associated with IMF programs that are not fully controlled for; channels not fully disentangled.

*Source: IMF Working Paper — "Classification of Capital Account Crises (KAC) and Control Group (CG) Episode" (content unit: _wp0675).*

### 1. Classification of Capital Account Crises (KAC) and Control Group (CG) Episode................. 9

### 1. Classification of Capital Account Crises (KAC) and Control Group (CG) Episode

### Overview and purpose
- Examines whether IMF-supported programs, conditional on episodes of intense market pressures, can help prevent a capital account crisis.
- Focuses on crisis prevention (whether abrupt private capital outflows are avoided), distinct from crisis resolution (resumption of private capital inflows).
- Uses cluster analysis to identify episodes of intense market pressures and then classifies them into capital account crises (KAC) or control group (CG) episodes to establish a counterfactual.
- Applies a logit specification to test whether an episode of intense market pressure develops into a KAC or a CG after controlling for initial conditions, policy changes, and exogenous factors.

### Key findings (as stated)
- Availability of IMF resources (either through disbursements or resources available under an on-track precautionary program) lowers the likelihood of a crisis.
- IMF support lowers the likelihood of a crisis even after controlling for (gross) foreign exchange reserves—i.e., “money matters” but signaling/credibility and stronger policies associated with IMF support are also important.
- IMF financing as a crisis prevention tool is most effective for countries with an intermediate range of economic fundamentals.

### Theoretical and empirical context
- Distinguishes catalytic effects studied in prior literature (where IMF financing multiplies official and private inflows) from the paper’s preventive focus (one dollar of IMF support resulting in more than one dollar of net inflows relative to the counterfactual of private exit).
- Reviews four signaling channels through which IMF-supported programs may reduce run likelihood: (i) providing liquidity; (ii) supporting stronger policies; (iii) signaling these policies to financial markets; (iv) enhancing credibility via program conditionality.
- Cites theoretical models (e.g., Morris and Shin (2006), Corsetti et al. (2004), Kim (2006), Penalver (2004), Zettelmeyer (2000)) that provide mixed results but include mechanisms where IMF financing can lower crisis likelihood, especially for intermediate fundamentals or where financing induces better policies.

### Data universe and scope
- Monthly data from 1994 to 2004 for 27 emerging market economies:
  - Algeria, Argentina, Brazil, Bulgaria, Chile, Colombia, Dominican Republic, Ecuador, Hungary, Indonesia, Korea, Malaysia, Mexico, Morocco, Pakistan, Panama, Peru, Philippines, Poland, Russia, South Africa, Thailand, Tunisia, Turkey, Ukraine, Uruguay, and Venezuela.

### Identification methodology — two-step cluster approach
- Uses cluster analysis to avoid ad-hoc crisis thresholds; classifies datasets into five clusters (1 = highest intensity of market pressures through 5) to span strengthening, neutral, and weakening pressures.
- Step 1 — Identify episodes of intense market pressures:
  - Constructs an index of market pressures (IMP) defined as:
    )S/Sln()REER/REERln()FX/FXln(IMP1,,1,,1,,,−−−+−−=tititititititi
  - Components: real exchange rates (REER), foreign exchange reserves (FX), secondary market spreads (S) on sovereign bonds (vis-à-vis the 10-year U.S. Treasury bill).
  - Each term is standardized (mean = 0, standard deviation = 1) for each country.
  - Interpretation: increases in FX reserves or currency appreciation indicate easing pressures; higher spreads indicate rising pressures.
  - From cluster 1 (highest intensity) a total of 32 market pressure episodes are identified (cluster 1) — these constitute the universe of KAC and CG events for 1994-2004. Duration measured in months; a “tranquil” period of 12 months is required between events to treat them as distinct.
- Step 2 — Segment high-pressure events into KAC and CG using private capital flows:
  - Applies cluster analysis to net private capital flows (excluding FDI, as a percent of GDP; quarterly data from 1994 to 2004) using five clusters: cluster 1 = high inflows (HI=1) through cluster 5 = high outflows (HO=5).
  - A KAC episode is defined when a pressure event is followed by at least two quarters (persistence) of medium or high capital outflows during the four quarters beginning in the quarter the market pressure event took place (clusters 4 or 5).
  - Resulting counts: 11 capital account crisis cases and 21 cases where severe market pressure did not result in a capital account crisis (the control group).

### Empirical strategy and estimation
- After classification, a logit model is estimated to examine the role of IMF financing in crisis prevention (probability that a pressure episode evolves into KAC vs CG), controlling for:
  - Initial conditions (including reserves),
  - Changes in economic policy,
  - Exogenous and other factors.
- Distinguishes between the existence of IMF-supported programs and the amount/availability of IMF resources; this paper focuses on available IMF resources (disbursements or drawing rights under on-track precautionary arrangements).

### Methodological notes and assumptions preserved
- Cluster analysis choice: five clusters chosen to provide interpretability; alternative tests for optimal cluster number are available but may reduce interpretability.
- Index construction rationale: uses REER, FX, and sovereign spreads as typical components of market pressure indices; standardization applied per country.
- Excludes FDI from net private capital flows when classifying outflows as crisis indicators.
- Persistence rule for KAC: at least two quarters of medium or high outflows within the four-quarter window starting in the quarter of the pressure event.

### Implications highlighted in the paper
- IMF financing reduces crisis probability both via liquidity and via strengthening policy credibility; effects persist after controlling for reserves.
- Crisis prevention effectiveness of IMF financing is nonlinear across fundamentals—most effective for intermediate fundamentals.

*Source: IMF Working Paper — "Classification of Capital Account Crises (KAC) and Control Group (CG) Episode", pages and sections as provided in the source content.*

### 1998. In each case the least severe market pressure was dropped (the first of the identified

### _wp0675 - 1998. In each case the least severe market pressure was dropped (the first of the identified

### Identification of sample groups and episodes
- Two sample groups constructed: capital account crises (KAC) and control group (CG).
- Market pressures identified by classifying monthly data into five clusters based on an index of market pressures that includes changes in REER, FX reserves, and spreads. Countries listed are in the cluster with the highest market pressures (cluster 1).
- Private capital flows (net of FDI) used to distinguish between KAC and CG episodes.
- Classification rule into KAC and CG:
  - A KAC event requires 2 quarters of either MO or HO in the 4 quarters that follow the build-up of market pressures.
  - All other episodes are in the control group (CG).
- A total of 11 KAC and 21 CG episodes are included in the analysis.

### Stylized facts: market/financial indicators (medians)
- Index of market pressures:
  - Peaks at time t in both groups; index is higher for KAC events.
- Spreads and residual spreads:
  - Spreads are higher in level for KAC episodes and the gap widens after t.
- Interest rates:
  - Treasury Bill (percent a year) levels are higher during t+1 to t+4 for KAC cases.
- Real Effective Exchange Rate (t=100):
  - REER index dips for both groups post-t, with a larger dip for KAC.
- Reserves (in percent of GDP):
  - Start at about the same level (about 10 percent of GDP) pre-t; KAC cases experience a significant drop in period t and recover a few quarters after the onset of the crisis.

### Stylized facts: macroeconomic indicators (medians)
- Real GDP growth (percent a year):
  - Somewhat stronger in the control group; difference widens after t. Both groups show a “V-shape” recovery.
- Inflation (percent a year):
  - Spikes in the KAC group after t, likely reflecting currency devaluation.
- Current account balance (percent of GDP):
  - Broadly same in both groups before t; current account adjustment after market pressures is much stronger among KAC events.
- Fiscal balance (percent of GDP):
  - Fiscal balance is worse among KAC episodes prior to onset; fiscal balance in KAC group peaks at time t-1.
- Broad money velocity (annual percent change):
  - Quite stable in both groups prior to t; monetary policy appears to loosen in KAC aftermath (short-lived).

### Stylized facts: external sector indicators (medians)
- Exchange rate regime:
  - AREAER-based data suggest the control group had a slightly more flexible exchange rate regime than the crisis group; no clear classification separation.
- Exchange rate overvaluation:
  - CG episodes appear to have a less overvalued exchange rate than crisis group.
  - Overvaluation measured as difference between actual REER and HP-filter trend of REER.
- Private capital flows (percent of GDP) 4/:
  - Similar in both groups until period t; by definition, KAC events show intense capital outflows after t.
- Reserves (in months of imports) and terms of trade:
  - Presented as medians for t-4 to t+4 (graphs in source).

### Stylized facts: solvency and liquidity indicators (medians)
- Total external debt (in percent of GDP):
  - Higher in KAC group: about 55 percent in KAC cases, slightly over 10 percentage points higher than CG.
- Short-term debt (in percent of GDP) and short-term debt (in percent of foreign reserves):
  - Ratio of short-term debt to reserves is stable among CG events but rises sharply among KAC events from period t-1 to t+2, then declines from t+3.

### Estimation approach and regressors
- Objective questions:
  - What determines whether intense market pressure develops into a capital account crisis?
  - Does IMF support play a role in the outcome?
- Econometric specification:
  - Logit model where dependent variable = 1 if episode is KAC at period t, 0 if CG.
  - Analysis focuses on the four quarters preceding period t (t-4 to t-1). Dataset: 32 pressure episodes × 4 quarters = 128 observations (results robust when using 32 observations).
- Regressor groups:
  - Initial Conditions:
    - Total external debt-to-GDP ratio (solvency).
    - Short-term debt-to-reserves ratio (liquidity).
    - Exchange rate regime prior to t (less flexible regime → higher crisis probability).
    - Exchange rate overvaluation (higher overvaluation → higher crisis probability).
    - Political stability (ICRG index; greater stability → lower crisis probability).
  - Policy Adjustment:
    - Fiscal adjustment in pre-crisis period and change in short-term real interest rates (monetary tightening).
  - Exogenous/Other Factors:
    - Terms of trade movements, size of economy, geographical (regional) location.
  - IMF Financing:
    - Uses ratio of available IMF resources to short-term debt in the four quarters up to each period (available = disbursed or accumulated drawing rights under precautionary arrangements).
    - IMF financing variable constructed as a share of short-term debt (STD); level in period t-1 calculated as sum of available IMF resources from t-4 to t-1 divided by short-term debt in t-1 (and similarly for earlier periods).

### Data and sample balance on IMF financing (Table 2 summary)
- Definition: IMF financing defined as a share of short-term debt (STD).
- Full sample (If IMF financing exists):
  - Mean: 0.055
  - Median: 0.048
  - Std. Dev.: 0.038
  - Min. (Uruguay; 2001Q4): 0.008
  - Max. (Russia; 1997Q3): 0.120
  - Obs.: 1784
  - Share of the full sample of KACs: 39
- LAC region (If IMF financing exists):
  - Mean: 0.050
  - Median: 0.045
  - Std. Dev.: 0.042
  - Min.: 0.008
  - Max.: 0.118
  - Obs.: 416
  - Share of the full sample of KACs: 50
- Full sample when IMF financing may not exist (All episodes; second panel):
  - Mean: 0.021
  - Median: 0.000
  - Std. Dev.: 0.036
  - Min.: 0.000
  - Max.: 0.120
  - Obs.: 416
- KAC episodes (when IMF financing exists):
  - Mean: 0.085
  - Median: 0.052
  - Std. Dev.: 0.072
  - Min. (Argentina; 1998Q2): 0.008
  - Max. (Brazil; 2002Q2): 0.234
  - Obs.: 22
  - Share of the full sample of CGs: 26
- CG episodes (when IMF financing exists):
  - Mean: 0.022
  - Median: 0.000
  - Std. Dev.: 0.052
  - Min.: 0.000
  - Max.: 0.234
  - Obs.: 48
- Sample balance and shares:
  - Sample of 32 episodes is balanced between KAC and CG episodes and between observations with and without IMF financing.
  - About one-third of all observations are KACs and two-thirds are in the control group.
  - About 40 percent of KAC episodes have IMF resources available prior to t; slightly over 25 percent among CG episodes have IMF resources available prior to t.
- Notes on IMF financing variable:
  - Sample has only two precautionary arrangements; distinguishing these econometrically from non-precautionary programs is not possible; excluding precautionary arrangements does not affect results.
  - IMF arrangements approved from period t onwards are excluded from the sample.

### Conceptual channels and empirical focus
- IMF support may prevent crises via:
  - Engendering sound policies and signaling them to markets.
  - Increasing liquidity available to the country.
- Empirical focus:
  - Uses IMF “available resources” (disbursed or accumulated drawing rights) rather than program dummies or lending commitments.
  - Controls implicitly for program on-track status by focusing on available resources.
- Note on alternative definitions:
  - Defining step 2 as private capital flows plus errors and omissions and applying cluster analysis changes classification for Mexico 1994 and Venezuela 2003 (would be KAC) but does not affect the paper’s results.

*Source: _wp0675 (excerpt).*

### Appendix Table 1 and Appendix Table 2 provide, respectively, descriptive statistics and

### _wp0675 - Appendix Table 1 and Appendix Table 2 provide, respectively, descriptive statistics and

### A. Results — overview of specifications and approach
- Regressions 1 and 2 (R1 and R2) examine the impact of IMF financing (without controlling for policy adjustment) using different sample sizes:
  - R1: estimation with 32 observations based on the simple average of quarterly data from t-4 to t-1.
  - R2: estimated using one observation for each of these periods (128 observations).
  - Motivation for using 128 observations: to examine how early a crisis can be predicted and to facilitate convergence when including additional regressors.
- Regressions 3 through 5 (R3, R4, R5) differ only on the definition of IMF involvement:
  - R3: dummy variable based on existence of an “on-track” IMF-supported program.
  - R4: adds IMF financing regressor to the on-track dummy.
  - R5: based solely on IMF financing regressor.
  - These three regressions include fiscal and monetary policy adjustments from t-4 to t-1 and exchange rate overvaluation relative to trend.
- Regression 6 (R6) controls for the interaction of changes in fiscal and monetary policy with the existence of an on-track IMF-supported program dummy.
- IMF financing regressor is based on available resources.

### Main econometric findings
- Control regressors generally have expected signs but are not always statistically significant; they are retained to reduce omitted variable bias.
- Significant and robust determinants of higher crisis probability:
  - High ratio of external debt to GDP (solvency threat) — statistically significant.
  - High ratio of short-term debt to FX reserves (liquidity threat) — statistically significant.
  - Exchange rate overvaluation — reported as “quite important” determinant.
  - Size of the economy — contributes significantly to crisis propensity.
- Political stability coefficient suggests higher political stability reduces crisis probability, but not always significant.
- Other controls:
  - More flexible exchange rate regime: small contribution to lower crisis probability; statistically significant only in one case.
  - Terms of trade: right sign but not statistically significant.
  - Regional dummies: neither is significant.
- Policy variables:
  - Monetary tightening (higher real interest rates) and fiscal tightening lower crisis probability; only the monetary variable is significant at the 10 percent level in some specifications.
  - In R6, fiscal adjustment interactive with IMF dummy is negative and significant — greater fiscal adjustment in countries receiving IMF financing helps avert crises, though the estimated coefficient is small.
  - Combined fiscal adjustment term for program countries is negative and jointly significant despite the non-interacted fiscal adjustment coefficient being positive and significant in R6.
  - Change in interest rates in R6 (with and without interaction) is not significant.

### Crisis prevention effects of IMF financing — key conclusions
- IMF money matters:
  - The coefficient on the IMF financing to short-term debt regressor is negative and statistically significant across all specifications.
  - IMF program dummy is consistently not significant — the signal appears related to the size of IMF financing rather than mere existence of an on-track program.
  - Conclusion: while an on-track program is important, IMF resources dominate the relationship.
- Channels and interpretation caveats:
  - Cannot fully disentangle all channels through which IMF-supported programs help prevent crises.
  - IMF financing regressor may capture stronger policies under IMF programs not controlled for by other regressors.
  - Because estimation controls for gross reserves, IMF financing effects go beyond liquidity — IMF financing appears more useful than other financing sources.
  - The credibility/signal of IMF support seems to depend on the IMF putting its own resources on the line.

### Summary bullets of main results (as stated)
- Stronger policies—tighter monetary policy (higher real interest rates) or greater fiscal adjustment (particularly in the context of a IMF-supported program)—are associated with a lower crisis likelihood and this association is statistically significant.
- IMF disbursements (or accumulated drawing rights) are a significant factor in crisis prevention: the larger are the disbursed IMF resources, the lower is the crisis likelihood.
- An important liquidity effect of IMF support on crisis prevention exists. IMF disbursements (or their availability under an on-track precautionary program) matters, rather than just an on-track program or possible future drawings under the arrangement.
- Benefits of IMF support go beyond liquidity effects: IMF financing is significant even controlling for foreign exchange reserves, implying effects from stronger policies bolstered by conditionality and the “seal of approval” implicit in IMF disbursements.

### Selected results from Table 3 (logit regression results) — key coefficients and statistics (exact values preserved)
- IMF involvement:
  - IMF financing (resource ratio): 3/-25.36 **, -20.58 **, -37.23 **, -40.25 **, -40.06 ***
  - IMF program dummy: 4/-1.11, 0.46
- Initial conditions:
  - Debt/GDP: 0.16 ***, 0.13 ***, 0.14 ***, 0.18 ***, 0.18 ***, 0.19 ***
  - ST debt/Reserves: 0.87 ***, 0.75 ***, 1.06 ***, 1.06 ***, 1.03 **, 1.12 ***
  - Exchange rate regime: -0.52, -0.40, -0.33, -0.56, -0.53, -0.67 *
  - Political stability: -5.46, -5.26, -6.44 **, -6.14, -6.04 *, -6.05 *
  - Exchange rate overvaluation: 7/17.65 ***, 25.22 ***, 24.74 ***, 26.29 ***
- Policy variables:
  - Fiscal balance change: 8/-0.08 **, -0.03, -0.03, 0.23 **
  - Fiscal balance interactive with Fund dummy: 9/-0.45 **
  - Interest rate change (real terms): 10/-0.08 **, -0.07 *, -0.07 *, -0.03
  - Interest rate interactive with Fund dummy: 11/-0.07
- Exogenous factors:
  - Terms of trade: 0.08, 0.04, -0.01, 0.02, 0.02, -0.00
- Other:
  - Size of the economy: 12/1.16 ***, 1.01 ***, 0.77 ***, 0.87 ***, 0.88 ***, 0.93 ***
  - Latin American dummy: -1.43, -1.15, -0.69, -1.49, -1.51, -1.98
  - Asian dummy: 2.71, 2.35, 2.41, 2.45, 2.34, 2.36
  - Constant: -5.49, -4.69, -5.92, -6.76, -6.98, -6.77
- Model fit and classification:
  - No. of observations: 32, 128, 128, 128, 128, 128
  - LR Chi-square: 20.2 **, 30.6 ***, 50.7 ***, 25.7 ***, 30.6 ***, 65.8 ***
  - Pseudo R-square: 0.51, 0.48, 0.53, 0.59, 0.59, 0.61
  - Correctly classified (in percent): 81, 80, 83, 87, 87, 87
  - Type I errors (in percent): 36, 34, 23, 16, 16, 18
  - Type II errors (in percent): 10, 13, 14, 12, 12, 11
- Notes:
  - ***, **, and * indicate significance at the 1, 5, and 10 percent levels of significance.
  - Standard errors are adjusted for within cluster correlation (i.e., correlation at the level of each pressure episode).
  - Logit regressions using random effects provide similar results.
  - Definitions and measurement notes for numbered regressors are provided in Table 3 footnotes (e.g., IMF financing definition, exchange rate regime scale, fiscal balance change definition).

### Model classification performance (Table 4) — selected statistics (exact values)
- Correctly classified KAC episodes (number of quarterly episodes)? 37
- Number of KAC episodes wrongly classified as CG in t-4 to t-1 (Type I error): 8
- Number of KAC episodes wrongly classified as CG in t-1 (Type I error): 0
- Correctly classified CG episodes (number of quarterly episodes)? 74
- Number of CG episodes wrongly classified as KAC in t-4 to t-1 (Type II error): 9
- Number of CG episodes wrongly classified as KAC in t-1 (Type II error): 2
- Correctly classified (in percent) 87
- R6 (regression 6) has the highest pseudo R-squared and correctly classifies 87 percent of the observations; R6 correctly classifies all KAC episodes at least one quarter prior to onset (no type I errors in period t-1).
- Two instances where model incorrectly classifies an event as a KAC in period t-1: Mexico in 1994 and Turkey in 1998.

### B. Robustness tests — sample and technique notes
- Sample: 27 countries during the period 1994-2004; data partition technique narrows sample to 32 intense market pressure episodes.
  - Advantages: all 32 episodes are intense market pressure events (helpful for establishing a counterfactual); balanced sample of KAC and CG episodes.
- The paper queries dependence of results on sample and potential influence of outliers (further robustness tests discussed subsequently in the original text).

*Source: _wp0675 - Appendix Table 1 and Appendix Table 2 provide, respectively, descriptive statistics and*

### conclusions regarding IMF resources? The robustness tests below are intended to test whether

### _wp0675 - conclusions regarding IMF resources? The robustness tests below are intended to test whether

### Robustness tests and sample sensitivity
- Reclassifying episodes:
  - Mexico 1994 and Turkey 1998 are predicted by the logit model as KACs but cluster analysis classifies them as CG; treating them as KACs does not alter results on IMF financing.
- Adding episodes:
  - Argentina (early 1995) was reclassified as a 33rd KAC event for testing; this does not affect the role of IMF resources in crisis prevention. 
- Dropping one-time KAC cases in Asia:
  - Korea, Malaysia, and Thailand were dropped to test potential bias from single-appearance KAC-only countries; Indonesia and Philippines were retained. Results on IMF resources do not change.
- Controlling for repeat episodes:
  - A dummy for countries with “repeat” episodes was added (complements windowing with a 12-month minimum separation); the importance of IMF resources for crisis prevention is unaffected.
- Sensitivity to inclusion/exclusion of regressors:
  - Dropping one regressor at a time and estimating a parsimonious model (only statistically significant regressors) both leave the role of IMF resources in crisis prevention robust.
- Controlling for data outliers:
  - dfbetas technique used to identify influential observations; dropping outliers and re-estimating does not change the IMF financing result.
  - Dropping any single episode does not change the IMF crisis prevention role.
- Sample size and convergence caveats:
  - The sample is small; at times dropping outliers precludes maximum likelihood convergence. Two approaches—dropping some regressors and setting different cutoff levels for the dfbeta technique—were used; once convergence is restored, results on IMF financing remain unchanged.
- Classification rule noted:
  - The model prediction uses default cutoff probability of 0.5: crisis probability > 0.5 → KAC; < 0.5 → control group.

### The role of IMF financing in crisis prevention (empirical findings and scenarios)
- Key empirical insight:
  - IMF disbursements (or accumulated drawing rights under precautionary arrangements) over the 12 months prior to an episode of market pressures appear to reduce the probability a crisis will develop; existence of an IMF-supported program or IMF resource commitments do not show the same effect.
- Illustration (period t-1 comparisons for countries receiving IMF financing):
  - Among countries that ultimately avoided a crisis, several had predicted crisis probability > 50 percent without IMF financing, lowered substantially with IMF financing.
  - Among countries that ultimately faced a capital account crisis, IMF financing sometimes lowered predicted crisis probabilities but they often remained high.
- Dependence on country fundamentals:
  - The marginal effect of IMF financing depends on the average level of other covariates (external debt, exchange rate regime, short-term debt-to-reserves, economic policies).
- Figure 7 conclusions (marginal impact of IMF financing given fundamentals):
  - When no IMF financing is available, crisis probability ranking reflects economic fundamentals: best, median, worst covariates shown as distinct curves.
  - Significant drops in crisis probability can frequently be achieved by providing IMF financing, often requiring fewer resources than fully rolling over a country’s short-term debt.
  - In some cases (example: Indonesia), plausible levels of IMF resources (maximum financing ever provided to a KAC country) have no meaningful effect on crisis probability.
  - In all cases, resources needed to significantly reduce crisis probability exceed the average levels of IMF financing typically provided to KAC cases prior to a crisis.
- Required financing to hit an example threshold:
  - Reducing crisis probability to a 5 percent cutoff requires exceptional access (i.e., more than 100 percent of quota) in 8 out of 11 capital account crises episodes.
  - In many cases, resources needed are less than what was ultimately provided once the crisis erupted.
- Policy-strengthening versus financing:
  - To achieve an average reduction in crisis probability of 20 percent (the observed average decline among control group events receiving IMF financing), parametrically varying policies (keeping IMF financing constant) in regression R6 implies:
    - A combined increase in fiscal adjustment and tightening of monetary policy of close to 4½ percentage points each would be required.
    - A reduction in exchange rate overvaluation by 6 percent produces a similar 20 percent reduction in crisis probability, underscoring the importance of exchange rate overvaluation among KAC cases.
  - The average crisis probability among KAC episodes, with and without financing, in period t-1 was 87 percent.

### Welfare gains from crisis prevention (quantitative illustration)
- Back-of-the-envelope calculation premise:
  - Assumes the expected decline in crisis probability from IMF financing equals the average decline observed among control group events receiving IMF financing in t-4 to t-1: 0.21 (i.e., 21 percent or "about 20 percent").
  - Uses discounted difference in output flows between KAC and CG episodes over the first three years following a market pressure event.
- Table 5 headline result:
  - Estimated lower bound welfare gains: 5.6 percent of precrisis (t-4:t-1) GDP saved during the three years that follow a KAC.
- Method and numbers used in Table 5 (summary):
  - Potential and actual output series and present values computed with assumed discount rate of 5 percent.
  - Difference in output flows between KAC and CG episodes over first three years: KAC higher loss (example line item: Output difference within each group shows -35.4 and -8.6; Output difference across groups (D) is -26.8).
  - Difference in crisis probability among CG countries (E) = 0.21.
  - Lower bound welfare gains (F = abs(D*E)) = 5.6 percent of t-4:t-1 GDP.
- Caveats on welfare calculation:
  - This is a lower bound for two reasons:
    - Differences in output flows between KAC and CG are large; avoiding sharp declines could yield substantially higher gains.
    - Declines in crisis probability among capital account crises countries with IMF support might be higher than observed for CG benchmark.
  - Calculation assumes other covariates remain constant and does not control for all differences between CG and KAC events.

### Concluding remarks and policy implications
- Main conclusions:
  - IMF disbursements or available drawing rights under precautionary arrangements over the 12 months prior to intense market pressure appear materially to reduce crisis probability; market participants respond to actual or immediately available IMF money.
  - IMF money is not fully substitutable by gross foreign exchange reserves; IMF resources likely provide both liquidity and a policy/credibility signal.
  - IMF involvement may be insufficient when fundamentals are weak (high contribution of covariates to crisis probability).
- Policy implications and operational considerations:
  - A more pro-active prevention role by the IMF could yield important welfare gains—this implies making resources available before crises erupt (“before the house burns down”).
  - Phasing and front-loading:
    - Substantial front-loading of resources may be required where short-term debt coverage is low.
    - Front-loading increases the risk of “policy moral hazard” because financing may be provided before adjustment is needed.
  - Signaling and conditionality:
    - IMF support should send a positive signal to markets while also conveying that the member will benefit from IMF discipline and credibility; the IMF’s willingness to put its own resources at risk strengthens the signal.
- Limitations and caution:
  - Small sample: 32 high market pressure episodes, only a handful turned into full blown capital account crises; limits ability to assess nonlinearities or contagion threshold effects.
  - Establishing counterfactuals is challenging; results are based on a carefully designed empirical methodology and are consistent across multiple robustness checks, but must be interpreted with caution given sample size and simultaneity concerns between policy responses and IMF financing.

*Source: IMF working paper (content unit: _wp0675 - conclusions regarding IMF resources? The robustness tests below are intended to test whether).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp0675.pdf_