## wpiea2024034-print-pdf

## Source details

**Canonical URL:** [wpiea2024034-print-pdf](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024034-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024034-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024034-print-pdf.pdf.json)

---

### Key findings
- Main empirical conclusions:
  - (i) The main stressors that increase coup probabilities are a destabilized economic situation (weaker growth or external position, high overall and food inflation) and a destabilized political and security situation.
  - (ii) The main sources of political fragility are usually demographic pressures (a younger population) and weak structural fundamentals that point to poverty, lower economic, social or political inclusion (higher inequality, lower literacy rates, higher ethnic fractionalization, weaker democratization), weak governance, and a more recent and higher incidence of coups.
  - (iii) These structural sources of fragility act as double-sided amplifiers of stressors: they increase the probability of coups when stressors are present but also reduce the probability of coups faster when the stressors recede.
  - (iv) Policy improvements yield stronger dividends in fragile states: even moderate improvements in the policy environment (fiscal position and governance) help reduce coup probabilities, especially when structural fundamentals are weaker.
  - (v) In countries where structural fundamentals are strong, coup probabilities are not very responsive to economic, political or policy shocks.
  - (vi) Weaknesses on multiple structural fundamentals compound each other and make countries more prone to coups.
  - (vii) Stressors can also compound each other, with the overlapping crises facing most countries in 2020–23 likely prone to fragilizing political systems, especially in countries with weaker fundamentals.
- Overall implication:
  - Political and economic instability are easier to exploit where structural fragility is present, but strengthened policies and macroeconomic outcomes have higher returns in such environments. This supports continued engagement by international financial institutions through financial assistance and policy support.

### Methodology overview
- Event study:
  - Ten-year window around coups to identify pre-coup stressors as fast-moving variables with significantly different dynamics in the years leading up to coups.
  - Specification follows Gourinchas & Maurice (2012) and Catão and Milesi-Ferretti (2014).
  - Results reported as regression coefficients measuring predictor values in the ten-year window relative to tranquil periods outside the window.
  - Sample for event study: 86 variables, 192 countries, 1970–2019.
- Machine learning:
  - Training sample: 192 countries, period 1970–2019; 2020-2022 kept for out-of-sample testing.
  - Features: consolidated set of 55 predictors (first-difference transformations included).
  - Best-performing model: random forest; variable importance measured by Shapley values.
  - Nonlinearities and interactions examined via tree-based ensembles and Shapley decomposition.
- Causal inference:
  - Neither the event study nor the machine learning exercise is claimed to establish causal relationships between variables and coup events.

### Definition and data coverage
- Coup definition (Powell and Thyne (2011)): “overt attempts by the military or other elites within the state to unseat the sitting head of state using unconstitutional means”.
  - If perpetrators hold power for at least seven days, the coup is considered successful; otherwise it is a coup attempt.
- Dataset coverage:
  - Based on Powell and Thyne (2011) database covering 1950–September 8, 2023.
  - Total coups: 491 in 97 countries since 1950.
  - Of the 491 coups, 49.8 percent were successful in changing the regime and 50.2 percent were unsuccessful (coup attempts).

### Stylized facts about coups and empirical patterns
- Historical frequency and trends:
  - Number of coups peaked at 18 in 1966.
  - Average of 12 per year during the 1960s.
  - Fell to 3 during the 2010–19 decade.
  - The pandemic and post-pandemic period 2020-2023 saw a resurgence with 15 coups and coup attempts in these years, all but one in sub-Saharan Africa.
- Regional distribution:
  - Sub-Saharan Africa: 46½ percent of all coups historically.
  - Latin America and the Caribbean: 24.1 percent.
  - Middle Eastern region: 11.2 percent.
  - Over 2000-2023: Sub-Saharan Africa accounted for 70.7 percent of coups, East Asia and the Pacific 10.7 percent, Latin America and the Caribbean 8 percent.
- Development and regime type correlations:
  - Probability of experiencing at least one coup: 28 percent for countries in the first (poorest) income quintile, but less than 5 percent for those in the fifth (wealthiest) income quintile.
  - Regime type: Close to 92 percent of coups occurred in either closed or electoral autocracies; of these coups, 53 percent succeeded.
- Post-coup transitions to elections:
  - Restoration of electoral powers happens faster in high income countries (after two years, on average), whereas in low-income countries a transition government stays on relatively longer (4 years on average).
  - In low-income countries, approximately 50 percent of cases experience a transition duration of around 2 years between the coup year and the subsequent election, and about a quarter have transitions that last five years or more.
  - In the higher income group, the median transition duration is about 1 year, although in a quarter of the countries transitions last three years or more.
- Spatial clustering and spillovers:
  - Of the 491 coups, 65.4 percent occurred against the backdrop of other coups in adjacent countries within the previous five years.
  - 45.7 percent of coups occurred within 2 years of other coups in adjacent countries.
  - Example: wave of coups during 2020–23 in the Sahel region; all countries within the Sahel (narrowly defined), except for Mauritania, had at least one instance of a coup since 2020.

### Event study — dynamics around coups
- Pre-coup stressors (statistically significant deterioration prior to coups):
  - Growth collapses: growth is on average more than 1 percentage point lower one year before coups than in tranquil times.
  - In LICs and SSA: economic growth is on average 2 percentage points lower one year before coups.
  - Heightened political instability: government crises, major cabinet changes, social unrest increase in the year prior to coups and intensify further in the coup year.
  - Army recruitment increases significantly at least two years prior to coup events.
  - Energy and food inflation increase significantly a year before coups in LICs and ME&CA respectively.
- Concurrent dynamics (movements during coup years):
  - Macroeconomic deterioration in coup year: deeper declines in growth, foreign reserves, and drops in tax revenue.
  - Major constitutional changes often occur in the coup year.
  - Heightened social unrest, strikes, or demonstrations may precede or follow coups (examples: 2020 Mali; Sudan, 2021–22).
  - Spike in fatalities associated with bloodier coups in earlier periods and/or background insecurity (Sahel 2020-22).
- Post-coup dynamics (successful coups reported):
  - Governance and policymaking deteriorates: regulatory quality, control of corruption, rule of law, government and legislative effectiveness, voice and accountability decline.
  - Some governance indicators broadly revert to normality 4–5 years after the coup, but many remain depressed over the medium-term.
  - Political instability and violence persist into the medium-term and are most pronounced the year after the coup.
  - Macroeconomic effects:
    - Growth declines a year prior to coups, dips further during the coup year, and seems to recover to tranquil-time levels very soon following the coup; for successful coups, growth underperformance persists one more year after the coup.
    - Official development assistance clearly declines in the first three years after a coup, recovering to normal levels only afterward.
    - Fiscal position weakens in the first three years after successful coups: fiscal deficits widen—statistically significant in the year following the coup—in part due to decline in revenues; public debt increases and strengthens only afterward.
- Note: post-coup dynamics are deviations relative to tranquil times and not claimed as causal impacts of coups.

### Machine learning — top predictors and nonlinearities
- Top drivers (full sample, 1970–2019):
  - Younger demographics.
  - More recent coup history.
  - Lower level of development.
  - Economic and political instability (conjunctural stressors).
  - Conflict.
  - Weaker degree of democratization.
- Regional differences:
  - ME&CA: younger population, higher inflation, lower current account balance are oversized predictors.
  - SSA: inflation (including food inflation) is important; higher natural resource rents are important.
- Time evolution:
  - Shift toward social and political instability as more important in recent years; macro instability and weak policies were stronger drivers in earlier decades.
- Nonlinear thresholds and ranges (preserved verbatim examples from source):
  - "Share of elderly population: probability does not decrease further once the share of elderly population reaches 5-6 percent."
  - "GDP per capita: probability does not decrease further once GDP per capita exceeds approximately 2,500 PPP dollars."
  - "Growth: probability does not decrease further when growth is above 5-6 percent; lower growth increases coup probabilities especially below 2-4 percent (depending on sample)."
  - "Younger population: countries with a high share of younger population (younger than 65 years of age)—above 95 percent or so—are more prone to coups."
  - "Coup history: probability drops significantly only after some 23 years without coups; having 3 or more coups significantly increases probability of another coup."
  - "Income: most coup incidence associated with low income levels—below some 2,500 PPP dollars per capita."
  - "Inequality: in 2000-19 period, Gini above some 40 out of 100 associated with much higher coup probability."
  - "Natural resource rents: in full sample coup probability particularly high when rents range from 5-25 percent of GDP; in the most recent 20 years rents above 10 percent of GDP associated with high propensity for coups."
  - "Rule of law and political stability: weak rule of law and low political stability increase coup probability; lower literacy rates increase coup probability when literacy rates below 60-70 percent (and in SSA below some 40 percent)."
  - "Inflation: higher inflation associated with higher likelihood of coups; food inflation above some 10 percent is another coup driver (most significant in SSA)."
  - "External position: in most subsamples (e.g., 2000-19 and ME&CA) a weak external position is a significant driver (higher current account deficit and lower FX reserves increase coup probabilities); in SSA the opposite often holds—higher current account balances (usually higher than about -6 percent of GDP) increase coup probabilities due to resource rents and positive terms-of-trade shocks increasing incentives to seize power."
  - "Fiscal and policy indicators: improvements/deteriorations in the fiscal balance reduce/increase coup probabilities (2000-19 sample shown); government revenues as a share of GDP not a top contributor (ranked 21st) but revenues below 10 percent of GDP for the full sample and below 18 percent for the last twenty years increase coup probability."
  - "Public debt: when below about 50 percent of GDP, higher debt associated with higher coup probability."
  - "Military variables: increase in military expenditure associated with lower probability of coups (especially in 2000-19 and SSA); level of military spending has a U-shaped effect—very low (less than 1 percent of GDP) or very high (above 2-5 percent of GDP) levels of spending reduce coup probabilities; a higher share of armed forces increases probability of coups when they exceed about 1 percent of the labor force (recent years)."

### Interaction effects (stressors × structural features)
- Key interaction findings:
  - (i) Weak structural fundamentals amplify stressors: negative growth shocks or high inflation increase coup probability more in countries with weaker fundamentals (lower income, higher inequality, weaker inclusion, weaker governance); positive shocks reduce coup probability more in such countries.
  - (ii) Policies interact with sources of fragility: deteriorations in fiscal position or rule of law increase coup probability more in countries with weaker structural fundamentals; conversely, moderate policy improvements deliver stronger reductions in coup probability when fragility is present.
  - (iii) Stressors amplify each other: higher natural resource rents have a larger impact on coup probability in politically unstable countries; higher Gini increases coup probability more in countries with weaker voice and accountability; weak growth combined with high inflation increases coup likelihood more; faster population growth exacerbates weak growth’s effect.
- Policy interpretation:
  - A coup is easier to exploit where structural fragility exists; overlapping shocks (e.g., pandemic, war in Ukraine, food/energy price increases) make fragile countries especially prone to political breakdowns.

### Predictive performance and recent out-of-sample results (2020–22)
- Model design:
  - Random forest selected after block time-series cross-validation and Bayesian hyperparameter optimization.
  - Year t-1 predictors used to predict coup in year t.
- Out-of-sample performance (2020–22):
  - Model’s predicted coup probabilities increased since 2020, broadly tracking observed coups.
  - Of the 10 countries that had coups in the last 3 years, the model places 9 in the top 25 countries with highest coup probabilities (above the 80th percentile).
  - The model captures all recent coups in sub-Saharan Africa, leaving out only Myanmar from the top probability countries.
  - Out-of-sample AUC for 2020-22: 0.878.
- Main drivers of high predicted probabilities since 2020:
  - Recent coup incidence.
  - Weak growth.
  - Young demographics.
- Illustrative country-level driver patterns (predictions for 2022 as indicative):
  - Niger: main driver category—conflict and broader weak socio-political stability; spillover effects from neighboring coups.
  - Gabon: main driver categories—weak inclusion (high inequality, weak democratization) and weak governance.
- Uses and limitations:
  - Emphasizes cross-country structural fundamentals more than time-bound stressors.
  - Performs relatively well predicting coups in SSA during 2020–23.
  - Predictive results for countries with high predicted coup probability but without coups are not reported in the paper due to sensitivities.

### Policy implications and recommendations
- Strengthen structural fundamentals to reduce baseline fragility and increase resilience:
  - Economic development, inclusion, governance, literacy, decentralization prioritized.
- Policy improvements yield higher returns in fragile environments:
  - Even moderate improvements in fiscal position, governance, and rule of law can materially reduce coup probabilities, especially where fundamentals are weak.
- Manage macroeconomic stressors:
  - Support growth, contain inflation, maintain external buffers—particularly critical in fragile contexts.
- Address interaction risks:
  - Policy responses should account for amplifying effects (e.g., resource rents in politically unstable contexts, inequality in low-voice environments).
  - Prioritize combined structural and conjunctural reforms and continued engagement by multilateral institutions and donors: financial support and program designs targeting stronger policies can help stabilize economies and mitigate coup risks.
- Specific note on resource rents:
  - Higher resource rents increase the benefits of being in power and thereby raise the likelihood of coups in many lower income countries.

### Annexes and case studies (high-level)
- Annex Table 1a: comprehensive list of predictors used in event studies across categories: Development & Demographics; Inclusion & Governance; Macro Stability; Policy; Sociopolitical Stability.
- Annex summaries:
  - Annex II: Event Study Results — charts for economic, socio-political, and governance indicators across subsamples (Baseline, LICs, SSA, 1970-1999, 2000-2019).
  - Annex III: Machine Learning Results — Shapley Values for top-20 predictors in multiple samples (full sample, 2000-2019, SSA 1970-2019, ME&CA 1970-2019).
  - Annex IV: Selected case studies and stylized findings — Sahel Region 2020–23; Mali; Burkina Faso; Niger; Sudan; Venezuela.
- Stylized case highlights:
  - Sahel Region, 2020–23: post-2011 dynamics, rise of armed conflicts, seven coups or attempts in Sahel, deterioration in security and wellbeing despite international support.
  - Venezuela: long history of political fragility with episodes of coup attempts clustering with macroeconomic stress periods and severe economic underperformance during 2012–20.

*Source: IMF Working Paper No. WP/24/34 — Political Fragility: Coups d’État and Their Drivers (event study, machine learning, predictive analysis, and annex summaries).*

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

### wpiea2024034-print-pdf - References

### Key findings
- The paper finds that:
  - (i) the main stressors that increase coup probabilities are a destabilized economic situation (weaker growth or external position, high overall and food inflation) and a destabilized political and security situation.
  - (ii) the main sources of political fragility are usually demographic pressures (a younger population) and weak structural fundamentals that point to poverty, lower economic, social or political inclusion (higher inequality, lower literacy rates, higher ethnic fractionalization, weaker democratization), weak governance, and a more recent and higher incidence of coups.
  - (iii) these structural sources of fragility act as double-sided amplifiers of stressors: they increase the probability of coups when stressors are present but also reduce the probability of coups faster when the stressors recede.
  - (iv) policy improvements yield stronger dividends in fragile states: even moderate improvements in the policy environment (fiscal position and governance) help reduce coup probabilities, especially when structural fundamentals are weaker.
  - (v) in countries where structural fundamentals are strong, coup probabilities are not very responsive to economic, political or policy shocks.
  - (vi) weaknesses on multiple structural fundamentals compound each other and make countries more prone to coups.
  - (vii) stressors can also compound each other, with the overlapping crises facing most countries in 2020–23 likely prone to fragilizing political systems, especially in countries with weaker fundamentals.
- Overall implication: political and economic instability are easier to exploit if structural sources of fragility are present, but strengthened policies and macroeconomic outcomes have higher returns in such environments. This supports continued engagement by international financial institutions through financial assistance and policy support.

### Methodology overview
- Event study:
  - Uses a ten-year window around coups to identify pre-coup stressors defined as fast-moving variables with significantly different dynamics in the years leading up to coups.
- Machine learning:
  - Employs flexible nonparametric machine learning models to accommodate nonlinearities and a wide range of predictors.
  - Aims to identify both stressors (conjunctural factors) and sources of fragility (more structural characteristics), and to examine interactions between them.
- Causal inference:
  - Neither the event study nor the machine learning exercise is claimed to establish causal relationships between variables and coup events.

### Definition and data coverage
- Coup definition (Powell and Thyne (2011)): “overt attempts by the military or other elites within the state to unseat the sitting head of state using unconstitutional means”.
  - If perpetrators hold power for at least seven days, the coup is considered successful; otherwise it is a coup attempt.
- Dataset coverage:
  - Based on Powell and Thyne (2011) database covering 1950–September 8, 2023.
  - Total coups: 491 in 97 countries since 1950.
  - Of the 491 coups, 49.8 percent were successful in changing the regime and 50.2 percent were unsuccessful (coup attempts).

### Stylized facts about coups and empirical patterns
- Historical frequency and trends:
  - Number of coups peaked at 18 in 1966.
  - Average of 12 per year during the 1960s.
  - Fell to 3 during the 2010–19 decade.
  - The pandemic and post-pandemic period 2020-2023 saw a resurgence with 15 coups and coup attempts in these years, all but one in sub-Saharan Africa.
- Regional distribution:
  - Sub-Saharan Africa: 46½ percent of all coups historically.
  - Latin America and the Caribbean: 24.1 percent.
  - Middle Eastern region: 11.2 percent.
  - Over 2000-2023: Sub-Saharan Africa accounted for 70.7 percent of coups, East Asia and the Pacific 10.7 percent, Latin America and the Caribbean 8 percent.
- Development and regime type correlations:
  - Probability of experiencing at least one coup: 28 percent for countries in the first (poorest) income quintile, but less than 5 percent for those in the fifth (wealthiest) income quintile.
  - Regime type: Close to 92 percent of coups occurred in either closed or electoral autocracies; of these coups, 53 percent succeeded.
- Post-coup transitions to elections:
  - Restoration of electoral powers happens faster in high income countries (after two years, on average), whereas in low-income countries a transition government stays on relatively longer (4 years on average).
  - In low-income countries, approximately 50 percent of cases experience a transition duration of around 2 years between the coup year and the subsequent election, and about a quarter have transitions that last five years or more.
  - In the higher income group, the median transition duration is about 1 year, although in a quarter of the countries transitions last three years or more.
- Spatial clustering and spillovers:
  - Of the 491 coups, 65.4 percent occurred against the backdrop of other coups in adjacent countries within the previous five years.
  - 45.7 percent of coups occurred within 2 years of other coups in adjacent countries.
  - Example noted: wave of coups during 2020–23 in the Sahel region (Chad, Central African Republic, Mali, Burkina Faso, Niger, Guinea, Sudan) amid common challenges (escalation in terrorist incidents, social discontent, weak economic environment, geopolitical dynamics, social media influence). All countries within the Sahel (narrowly defined), except for Mauritania, had at least one instance of a coup since 2020.

### Position in literature and contribution
- Builds on literature linking low income, low growth, weak institutions, and political instability to coups and conflict.
- Novel contributions claimed:
  - First paper to examine empirically the predictive power of a wide range of macroeconomic and sociopolitical variables for coups with global evidence from more than 190 countries and more than 50 years.
  - First paper to apply machine learning to predicting coups (noting other conflict-related ML applications exist).

### Structure of the paper (as provided)
- Event study results examine dynamics of variables around coup events and identify stressors.
- Machine learning results identify main drivers (structural and conjunctural), examine nonlinearities and interactions, and explore regional and temporal differences.
- Additional analysis uses machine learning to assess likelihood of coups during 2020-2022, comparing predictions to actual outcomes.
- Annexes include Data and Methodology, Event Study Results, Machine Learning Results, and Selected Case studies (The Sahel Region, 2020–23; Sudan, 2021; Venezuela).

*International Monetary Fund. IMF WORKING PAPERS Political Fragility: Coups d’État and Their Drivers.*

### 3. Dynamics Around Coups d’État: Event Study

### Dynamics Around Coups d’État: Event Study

### Methodology and scope
- Event study specification follows Gourinchas & Maurice (2012) and Catão and Milesi-Ferretti (2014).
- Results reported as regression coefficients measuring predictor values in the ten-year window around the coup year relative to the tranquil period outside this window.
- Sample and predictors:
  - 86 variables capturing demographic, development, inclusion, governance, macro stability, and sociopolitical stability characteristics.
  - 192 countries during the period 1970–2019.
- Overlapping coup windows: coefficients are additive, capturing combined effects of relevant coups.
- Sample limited to 2019 because predictor data are scarcer in more recent years and pandemic years 2020–22 may introduce different dynamics.
- Political instability and security proxied by World Governance Indicators’ “political stability and absence of violence/terrorism”; results similar using ICRG political risk rating.
- Results discussed only for variables that show statistically significant dynamics relative to tranquil periods in the full sample or in regional, time, or income subsamples.

### Pre-coup dynamics: stressors
- Main indicators with statistically significant deterioration prior to coups:
  - Growth collapses.
  - Heightened political instability, including government crises, major cabinet changes, and social unrest.
- Key quantified findings:
  - Growth is on average more than 1 percentage point lower one year before coups than in tranquil times.
  - Army recruitment increases significantly at least two years prior to coup events.
  - In low-income countries (LICs) and in sub-Saharan African countries, economic growth is on average 2 percentage points lower one year before coups.
  - Energy and food inflation increase significantly a year before coups in LICs and Middle East and Central Asia (ME&CA) countries, respectively.
- Temporal pattern:
  - Political instability deterioration is driven particularly by the last twenty years, during which political stability seems to have gained in importance.
  - Government crises, major cabinet changes, and social unrest increase in the year prior to coups and intensify further in the year of the coup.

### Concurrent dynamics: co-movements with coups
- Some variables show significant movements during coup years but data frequency does not establish lead/lag:
  - Macroeconomic deterioration in coup year: deeper declines in growth, foreign reserves, and drops in tax revenue—these can be both precursors and consequences of coups.
  - Major constitutional changes often occur in the coup year; they may precede coups or follow them as consolidation measures.
  - Heightened social unrest, strikes, or anti-government demonstrations in the coup year may precede coups (example: 2020 Mali) or follow coups as protests against putschists (example: Sudan, 2021–22).
  - Spike in fatalities associated with bloodier coups in the earlier period and/or background insecurity (examples: Sahel coups in 2020-22).
  - Political instability and violence show significant deterioration during coup years.
  - In subsamples, energy inflation increase is statistically significant during the coup year in LICs and over the period 1970-1999.
- Recommendation for further analysis:
  - Leverage higher-frequency data to examine whether within-year movements precede or follow coups.

### Post-coup dynamics: implications of political fragility
- Results reported for successful coups only (irregular transitions that actually happen); full sample (coup and coup attempts) shows broadly similar qualitative effects.
- Governance and policymaking:
  - Quality of policymaking and governance deteriorates significantly after coups: regulatory quality, control of corruption, rule of law, government and legislative effectiveness, voice and accountability decline.
  - Some governance indicators broadly revert to normality 4–5 years after the coup, possibly coinciding with return to democratic elections, but many remain depressed over the medium-term.
- Political instability:
  - Political instability and violence/terrorism persist into the medium-term and are most pronounced the year after the coup.
  - Social unrest, government crises, and constitutional changes flare up the year after coups with almost equal force as pre-coup.
  - Post-coup government crises often reflect fragmentation within transition governments leading to frequent changes in government composition.
- Macroeconomic effects:
  - Growth declines a year prior to coups, dips further during the coup year, and seems to recover to tranquil-time levels very soon following the coup; for successful coups, growth underperformance persists one more year after the coup.
  - Official development assistance clearly declines in the first three years after a coup, recovering to normal levels only afterward.
  - Fiscal position weakens in the first three years after successful coups:
    - Fiscal deficits widen—statistically significant in the year following the coup—in part due to decline in revenues.
    - Public debt increases and strengthens only afterward, possibly when an elected government comes to power or when financing constraints bind.
- Note on causality:
  - Post-coup dynamics reported are deviations relative to tranquil times and are not claimed as causal impacts of coups; formal causal examination (e.g., local projections models) is left for future research.

*Source: IMF Working Paper — Political Fragility: Coups d’État and Their Drivers (event study results).*

### 4. Understanding the Drivers of Coups: Machine

### 4. Understanding the Drivers of Coups: Machine Learning

### Data, model, and approach
- Training sample: 192 countries during the period 1970–2019; period 2020-2022 kept for out-of-sample testing.  
- Prediction framework: year t-1 data used to predict coup in year t.  
- Features: consolidated set of 55 predictors (first-difference transformations included to capture within-country dynamics).  
- Best-performing model: random forest.  
- Variable importance and contributions to predicted coup probability measured by Shapley values.

### Top predictors and their evolution through time
- Overall top drivers (past fifty years):  
  - Younger demographics.  
  - More recent coup history.  
  - Lower level of development.  
  - Economic and political instability (conjunctural stressors).  
  - Conflict.  
  - Weaker degree of democratization.  
- Figure methodology: charts report drivers that together account for 50 percent of the contributions to the estimated coup probability for each sample (contribution measured by Shapley value).  
- Full sample (1970–2019) conjunctural drivers that increase coup probability: weaker growth, increased conflict, higher political instability, deterioration in governance. Structural drivers: higher share of young population, more recent/higher incidence of coups, lower income per capita, lower degree of democratization.  
- Regional subsamples: drivers are more homogeneous and fewer predictors account for 50 percent of coup probability:  
  - Middle East and Central Asia (ME&CA): younger population, higher inflation, lower current account balance are oversized predictors.  
  - Sub-Saharan Africa (SSA): inflation (including food inflation) is important; higher natural resource rents are one of the important drivers.  
- Time evolution: shift toward social and political instability as more important in recent years. In early years (1970s–1990s) macro instability, weak policies, weak inclusion and governance, poverty and demographics were main drivers; in recent decades economic policy and macro stability have reduced coup probabilities relative to the sample average, while political instability’s relative contribution increased.

### Nonlinear effects of individual predictors (how predictors affect coup probabilities)
- General statement: Most predictors affect coup probabilities nonlinearly; effect often occurs over specific ranges of predictor values.  
- Examples of nonlinear thresholds and ranges (preserved verbatim from source):  
  - Share of elderly population: probability does not decrease further once the share of elderly population reaches 5-6 percent.  
  - GDP per capita: probability does not decrease further once GDP per capita exceeds approximately 2,500 PPP dollars.  
  - Growth: probability does not decrease further when growth is above 5-6 percent; lower growth increases coup probabilities especially below 2-4 percent (depending on sample).  
  - Younger population: countries with a high share of younger population (younger than 65 years of age)—above 95 percent or so—are more prone to coups.  
  - Coup history: probability drops significantly only after some 23 years without coups; having 3 or more coups significantly increases probability of another coup.  
  - Income: most coup incidence associated with low income levels—below some 2,500 PPP dollars per capita.  
  - Inclusion and governance: low political inclusion (high autocracy index) associated with higher coup probability, though exception in full sample for highly autocratic or highly liberal governments which have low coup probabilities; in SSA coups generally occurred under the most autocratic governments.  
  - Inequality: in 2000-19 period, Gini above some 40 out of 100 associated with much higher coup probability.  
  - Natural resource rents: in full sample coup probability particularly high when rents range from 5-25 percent of GDP; in the most recent 20 years rents above 10 percent of GDP associated with high propensity for coups.  
  - Rule of law and political stability: weak rule of law and low political stability increase coup probability; lower literacy rates increase coup probability when literacy rates below 60-70 percent (and in SSA below some 40 percent).  
  - Centralization: higher degree of centralization (lower share of population with regional autonomy) associated with higher coup probability in recent years (2000-2019).  
  - Inflation: higher inflation associated with higher likelihood of coups; food inflation above some 10 percent is another coup driver (most significant in SSA).  
  - External position: in most subsamples (e.g., 2000-19 and ME&CA) a weak external position is a significant driver (higher current account deficit and lower FX reserves increase coup probabilities); in SSA the opposite often holds—higher current account balances (usually higher than about -6 percent of GDP) increase coup probabilities due to resource rents and positive terms-of-trade shocks increasing incentives to seize power.  
  - Fiscal and policy indicators: improvements/deteriorations in the fiscal balance reduce/increase coup probabilities (2000-19 sample shown); government revenues as a share of GDP not a top contributor (ranked 21st) but revenues below 10 percent of GDP for the full sample and below 18 percent for the last twenty years increase coup probability. Public debt: when below about 50 percent of GDP, higher debt associated with higher coup probability.  
  - Military variables: increase in military expenditure associated with lower probability of coups (especially in 2000-19 and SSA); level of military spending has a U-shaped effect—very low (less than 1 percent of GDP) or very high (above 2-5 percent of GDP) levels of spending reduce coup probabilities (level not shown in detail); a higher share of armed forces increases probability of coups when they exceed about 1 percent of the labor force (recent years).

### Interaction effects (how stressors and structural features combine)
- Method: tree structure of random forest uncovers interactions; plots compare Shapley contributions for quartiles of interacting predictors.  
- Key interaction findings:  
  - (i) Weak structural fundamentals amplify stressors: negative growth shocks or high inflation increase coup probability more in countries with weaker fundamentals (lower income, higher inequality, weaker inclusion, weaker governance); positive shocks reduce coup probability more in such countries as well. Example: weak growth increases coup probability more in low-income countries, while strong growth reduces it more in those countries. A coup is easier to exploit where structural fragility exists.  
  - (ii) Policies interact with sources of fragility: deteriorations in fiscal position or rule of law increase coup probability more in countries with weaker structural fundamentals (higher inequality, weak governance, low literacy, higher ethnic fractionalization, lower political stability); conversely, moderate policy improvements deliver stronger reductions in coup probability when fragility is present. Example: deterioration in fiscal balance in weak-fundamentals countries produces larger increase in coup probability; improvement in fiscal balance delivers stronger reductions (2000-19 sample shown).  
  - (iii) Stressors amplify each other: higher natural resource rents have a larger impact on coup probability in politically unstable countries; higher Gini increases coup probability more in countries with weaker voice and accountability; weak growth combined with high inflation increases coup likelihood more; faster population growth exacerbates weak growth’s effect on coup probability. Overlapping shocks (e.g., 2020–22 pandemic, war in Ukraine, and food/energy price increases) make countries with weaker fundamentals especially prone to fragilizing political systems.

### Policy implications drawn from predictive patterns
- Strengthening structural fundamentals—economic development, inclusion, governance, literacy, decentralization—reduces baseline political fragility and increases resilience to conjunctural stressors.  
- Even moderate improvements in policy variables (fiscal position, governance, rule of law) can materially reduce coup probabilities, especially in fragile countries.  
- Managing macroeconomic stressors (supporting growth, containing inflation, maintaining external buffers) matters for reducing coup risk, and is particularly important where structural fragility is present.  
- Addressing interaction risks: policy responses should account for amplifying effects (e.g., resource rents in politically unstable contexts, inequality in low-voice environments) and prioritize combined structural and conjunctural reforms to reduce vulnerability to coups.

*Source: IMF authors’ calculations (machine learning analysis reported in the chapter “Understanding the Drivers of Coups: Machine Learning”).*

### 5. Predicting Coups with the Machine Learning

### 5. Predicting Coups with the Machine Learning Model

### Model performance: out-of-sample prediction (2020–22)
- Training sample: data from 1970-2019.
- Out-of-sample period used to gauge performance: 2020–22.
- The model’s predicted coup probabilities increased since 2020, broadly tracking the rise in observed coups during the same period.
- Of the 10 countries that had coups in the last 3 years, the model places 9 in the top 25 countries with highest coup probabilities (above the 80th percentile of the coup probability distribution).
- The model captures all recent coups in sub-Saharan Africa, leaving out only Myanmar from the top probability countries.
- Out-of-sample area under the curve (AUC) for 2020-22: 0.878 out of 1, indicating overall good predictive performance.

### Main factors driving high predicted coup probabilities
- Common top drivers across recent cases include:
  - Recent coup incidence.
  - Weak growth.
  - Young demographics.
- Distinctive driver patterns in two illustrative country cases (predictions for 2022 used as indicative of pre-coup drivers):
  - Niger (Sahel region):
    - Main category driving high coup probability: conflict and broader weak socio-political stability.
    - Spillover effects from coups in neighboring countries emerge as an important driver (consistent with concentration of recent coups in Mali, Burkina Faso, and Chad).
  - Gabon (Central Africa):
    - Main categories driving high coup probability: weak inclusion (high inequality, weak democratization) and weak governance.
- These driver groups align with perceived contributors to coups in the discussed country cases.

### Model design and data inputs (methodological notes)
- Machine learning exercise uses a total of 55 variables (see Annex Table 1b).
- Key methodological choices:
  - Drop several potentially collinear variables from the event-study list to improve Shapley value interpretability; replace them with aggregated indicators (e.g., weighted conflict index; political stability and absence of violence/terrorism).
  - Use nominal GDP per capita (PPP terms) as a proxy for economic development to minimize predictors given high correlation (above 0.5) among development-related variables.
  - Include short-term dynamics via one-year change or percentage change transformations for many variables.
- Sample coverage: 192 countries/economies over fifty years from 1970 to 2022; machine learning training sample uses 1970-2019 to avoid pandemic years and account for data lags.
- For coups d’état the paper uses the database by Powell and Thyne (2011) covering 1950-2023 September 8; for 1970-2019 the database includes a total 293 coups, or 2.6 percent of all country-year observations during this period. Exercises include both coups and coup attempts.

### Uses, limitations, and monitoring implications
- The model emphasizes cross-country structural fundamentals more than time-bound stressors.
- It performs relatively well at predicting coups d’état in SSA during 2020–23 (including Mali, Burkina Faso, Niger, Gabon), and can be used to monitor countries in fragile political situations.
- Predictive results for countries with high predicted coup probability but that have not experienced coups are not reported in the paper due to sensitivities.

### Policy conclusions and recommendations
- Fragility drivers and shock interactions:
  - Destabilization of a country’s economic, political, or security environment raises coup likelihood.
  - Overlapping stressors amplify each other’s effects, increasing the likelihood of political breakdowns.
  - Global shocks during 2020–22 produced: lower growth, higher food, energy and overall inflation, and weaker external positions for food or oil price importers—all identified as factors that drive up coup probabilities.
- Structural vulnerabilities that increase breakdown risk:
  - Poverty, poor inclusion, weak governance, and high political and security risks.
  - These factors explain higher coup frequency in lower income countries and the recent wave of coups in sub-Saharan Africa (including the Sahel).
- Policy priorities with higher returns in fragile environments:
  - Strengthen fiscal policies (important given that weaker fiscal positions make breakdowns more likely).
  - Strengthen structural fundamentals: inclusion, governance, education.
  - Continued engagement by multilateral institutions and donors in fragile situations, because:
    - Financial support can help stabilize economies and mitigate coup risks.
    - Stronger policies targeted in program designs reduce political risks.
- Resource rents:
  - Higher resource rents (e.g., during times of higher oil prices) increase the benefits of being in power and thereby raise the likelihood of coups in many lower income countries.
- Overall implication:
  - Strengthened structural fundamentals and policy outcomes have higher returns in structurally fragile environments in terms of staving off political breakdowns than in more robust environments.

*Source: IMF Working Paper — chapter "5. Predicting Coups with the Machine Learning Model" (includes related conclusions).*

### Annex Table 1a. List of Predictors and their Categories Included in the Event Studies

### Annex Table 1a. List of Predictors and their Categories Included in the Event Studies

### Predictors by category (variable — source)
- Development & Demographics
  - Income per capita — World Economic Outlook Database
  - Individuals using the Internet (% of population) — World Development Indicators
  - Literacy rate — World Development Indicators
  - Natural Resources Rents — World Development Indicators
  - Population Density — Cross-National Time-Series Data Archive (CNTS)
  - Population in urban agglomerations of more than 1 million (% of total population) — World Development Indicators
  - Population Size — World Development Indicators
  - Share of Elder Population — World Development Indicators
  - Share of Young Population — World Development Indicators
- Inclusion & Governance
  - Competitiveness of Chief Executive Recruitment — Polity V
  - Competitiveness of Nominating Process — Cross-National Time-Series Data Archive (CNTS)
  - Competitiveness of participation — Polity V
  - Control of Corruption — World Governance Indicators
  - Degree of Democratization — Polity V
  - Degree of Parliamentary Responsibility — Cross-National Time-Series Data Archive (CNTS)
  - Economic Complexity Index — Atlas of Economic Complexity
  - Economic Freedom — Heritage Economic Freedom Index
  - Estimate of Gini index of inequality in equivalized household disposable income — SWIID
  - Ethnic fractionalization Index — Havard university
  - Executive constraints — Polity V
  - Head of State — Cross-National Time-Series Data Archive (CNTS)
  - Legislative Effectiveness — Cross-National Time-Series Data Archive (CNTS)
  - Legislative Selection — Cross-National Time-Series Data Archive (CNTS)
  - Number of active groups in this country — EPR (Ethnic Power Relations Dataverse)
  - Number of groups with regional autonomy in this country — EPR (Ethnic Power Relations Dataverse)
  - Number of relevant groups in this country — EPR (Ethnic Power Relations Dataverse)
  - Openness of Chief Executive Recruitment — Polity V
  - Party Coalitions — Cross-National Time-Series Data Archive (CNTS)
  - Party Fractionalization Index (Scaling: 0.0001) — Cross-National Time-Series Data Archive (CNTS)
  - Party Legitimacy — Cross-National Time-Series Data Archive (CNTS)
  - Regime Durability — Polity V
  - Regulation of Chief Executive Recruitment — Polity V
  - Regulation of participation — Polity V
  - Regulatory Quality — World Governance Indicators
  - Rule of Law — World Governance Indicators
  - Size of Legislature/Number of Seats, Largest Party — Cross-National Time-Series Data Archive (CNTS)
  - Sum of discriminated population as a fraction of ethnically relevant population — EPR (Ethnic Power Relations Dataverse)
  - Sum of discriminated population in this country (as a fraction of total population). — EPR (Ethnic Power Relations Dataverse)
  - Sum of the ethnically relevant population in this country (as a fraction of total population) — EPR (Ethnic Power Relations Dataverse)
  - Total population of all minority and excluded groups in this country (as a fraction of total population) — EPR (Ethnic Power Relations Dataverse)
  - Total population with regional autonomy in this country (as a fraction of total population). — EPR (Ethnic Power Relations Dataverse)
  - Voice and Accountability — World Governance Indicators
- Macro Stability
  - Agriculture, forestry, and fishing, value added (% GDP) — World Development Indicators
  - CPI Inflation — World Economic Outlook Database
  - Current Account Balance (% GDP) — World Economic Outlook Database
  - Domestic credit to private sector by banks (% GDP) — World Development Indicators
  - Energy inflation — World Economic Outlook Database
  - Food Inflation — World Economic Outlook Database
  - Gross domestic product, constant prices, National Currency, percent change — World Economic Outlook Database
  - Gross fixed capital formation, current prices, (% GDP) — World Economic Outlook Database
  - Labor force participation rate, total (% of total population ages 15-64) — World Development Indicators
  - Manufacturing, value added (% GDP) — World Development Indicators
  - National currency units per U.S. dollar, end of period — World Economic Outlook Database
  - Net ODA received (% of GNI) — World Development Indicators
  - Purchasing Power Parity per capita — World Economic Outlook Database
  - Terms of trade, goods, US Dollars, percent change — World Economic Outlook Database
  - Total external debt, percent of GDP — World Economic Outlook Database
  - Total Reserve Assets — World Economic Outlook Database
  - Unemployment, total (% of total labor force) — World Development Indicators
  - Wage and salaried workers, total (% of total employment) — World Development Indicators
- Policy
  - Fiscal Balance — World Economic Outlook Database
  - Public Debt (% of GDP) — World Economic Outlook Database
  - Government Effectiveness — World Governance Indicators
  - Government Tax Revenue (%GDP) — World Economic Outlook Database
- Sociopolitical Stability
  - Anti-Government Demonstrations — Cross-National Time-Series Data Archive (CNTS)
  - Armed forces personnel (% of total labor force) — World Development Indicators
  - Assassinations — Cross-National Time-Series Data Archive (CNTS)
  - Change in political regimes from last year — Political Regime
  - General Strikes — Cross-National Time-Series Data Archive (CNTS)
  - Government Crises — Cross-National Time-Series Data Archive (CNTS)
  - Guerrilla Warfare — Cross-National Time-Series Data Archive (CNTS)
  - Internally displaced persons, total displaced by conflict and violence (number of people) — World Development Indicators
  - Military expenditure (% GDP) — World Development Indicators
  - Number of conflict incidence — UCDP/PRIO
  - Number of Legislative Elections — Cross-National Time-Series Data Archive (CNTS)
  - Number of Major Cabinet Changes — Cross-National Time-Series Data Archive (CNTS)
  - Number of Major Constitutional Changes — Cross-National Time-Series Data Archive (CNTS)
  - Political Stability and Absence of Violence — World Governance Indicators
  - Purges — Cross-National Time-Series Data Archive (CNTS)
  - Refugee population by country or territory of asylum — World Development Indicators
  - Reported Social Unrest Index — Barrett et al. (2020)
  - Revolutions — Cross-National Time-Series Data Archive (CNTS)
  - Riots — Cross-National Time-Series Data Archive (CNTS)
  - Total fatalities due to conflict — UCDP/PRIO
  - Total Number of People Affected by Natural Disaster — EMDAT
  - Weighted Conflict Index — Cross-National Time-Series Data Archive (CNTS)

### Methodology for Event Study — key elements
- Estimation specification follows Gourinchas & Maurice (2012) and Catão and Milesi-Ferretti (2014).
- Treatment indicator:
  - Dummies D_{i c t+p} take value 1 when a coup or a coup attempt occurs in country i of continent c at time t, and 0 otherwise.
  - There are a total of 11 dummies per coup event spanning the 11-year window centered around the year t when there is a coup or a coup attempt.
- Controls and fixed effects:
  - Controls include the number of prior coups in country i of continent c, denoted n..._... (as specified in the source).
  - Country-continent fixed effect ⍺_{i c} and continent-specific time fixed effects λ_{c t} are included.
- Interpretation:
  - Coefficients β_p measure the difference between the level of the variable when around a coup (within the 10-year window around it) and its average level during “tranquil” periods (outside the 10-year window).
- Overlapping events:
  - When coup windows overlap, coefficients for overlapping years are summed to measure the treatment effect for those years.
  - Example: coups in 2001 and 2002 produce first coup window 1996–2006 and second 1997–2007, resulting in overlapping years 1997–2006; the treatment effect for overlapping years is the sum of effects from both coups.
- Annex Table 2 illustrative timeline (excerpted conceptually):
  - Coup in 1996 (t-5) marked as 1 in 1996
  - Coup in 2001 and 2002 lead to overlapping 1s from 1997 through 2006 (as shown in the source table)

### Methodology for Machine Learning — models, explainability, and evaluation
- Algorithms used:
  - Random Forests (Breiman, 2001)
  - XGBoost (Chen & Guestrin, 2016)
- Rationale:
  - Tree-based ensemble methods accommodate nonlinearities and interactions and reduce overfitting by combining many trees and introducing sampling/randomness.
- Explainability:
  - Predictor contributions reported via Shapley values (Strumbelj and Kononenko, 2010; Lundberg and Lee, 2017), which measure additive contribution of each predictor to the likelihood of a coup relative to the sample-average predicted probability.
- Model details:
  - Binary classification tree (BCT) is the base learner.
  - Random Forest: uses bootstrap aggregating (bagging) and random feature sampling; class predictions by majority vote; scores by averaging individual tree scores.
  - XGBoost: uses gradient boosting; trees trained sequentially to learn residuals; adds penalty term to loss function to control complexity measured by sum of the depth of all trees and the number of trees.
- Performance metric:
  - Area under the ROC curve (AUC) used to measure model performance; 0.5 = random guessing, 1 = perfect model.
- Hyperparameter tuning and validation:
  - Block time-series cross-validation method based on Burman et al. (1994) and Racine (2000) is employed to account for cross-sectional dependence in the panel sample.
  - Step 1: Construct 5 {training set, validation set} pairs with years:
    - {1970 - 2009, 2010 - 2011}
    - {1970 - 2011, 2012 - 2013}
    - {1970 - 2013, 2014 - 2015}
    - {1970 - 2015, 2016 - 2017}
    - {1970 - 2017, 2018- 2019}
    - Each training set contains observations from 1970 up to the cutoff year (2009, 2011, 2013, 2015, and 2017); each validation set contains the subsequent two years (2010-11, 2012-13, 2014-15, 2016-17, and 2018-19).
  - Step 2: For each ML algorithm, starting with a random set of hyperparameters, train the model on the training sets and apply to validation sets to obtain five AUCs.
  - In each tree-based algorithm, 1000 trees are constructed; hyperparameters to be tuned are specified in the source.

*Annex Table 1a. List of Predictors and their Categories Included in the Event Studies — source content from the provided PDF.*

### 1.    Random Forest: maximum tree depth for base learners and subsample ratio of columns when

### 1.    Random Forest: maximum tree depth for base learners and subsample ratio of columns when

### Hyperparameter tuning and model selection
- Random Forest hyperparameters tuned:
  - maximum tree depth for base learners
  - subsample ratio of columns when constructing each tree
- XGBoost hyperparameters tuned:
  - L2 regularization term on weights
  - boosting learning rate
- Optimization procedure:
  - Step 2 is repeated with different sets of hyperparameters which are chosen by Bayesian optimization over 100 iterations.
  - For each ML algorithm, the set of hyperparameters that yield the highest average AUC across the five validation sets are chosen as the optimal set of hyperparameters.
  - Step 3. Finally, the ML algorithm with its optimal set of hyperparameters that yields the highest average AUC is chosen as our final selected ML algorithm and the corresponding hyperparameters.

### Imputation
- Given the wide coverage of countries and long span of years of our sample, there is missing data especially for some earlier years or some individual countries.
- Approach:
  - Let the machine learning algorithms impute the missing data as the models are trained.

### Shapley Values — linear models
- For linear models, Shapley value interpretation:
  - Beginning with linear models, Shapley value of a predictor for an observation is simply the estimated coefficient multiplied by the observation’s value of the predictor.
- Equations (as presented):
  - 푏푏̂(푥푥푖)=휙휙0�푏푏̂�+�휙휙푘푛푘푘=1�푥푥푖;푏푏̂�=훽훽̂0+�훽훽̂푘푛푘푘=1푥푥푖,푘
  - Shapley value of the 푘푘th predictor for observation 푥푥푖 is calculated as 훽훽̂푘푥푥푖,푘, and the sum of Shapley values of all predictors for observation 푥푥푖 is the difference between its model prediction and the average prediction in the training sample.

### Shapley Values — non-linear, model-agnostic
- Conceptual framing:
  - Shapley values for non-linear models draw on cooperative game theory.
  - A model prediction for an observation is treated as a cooperative game: each predictor value is a “player” and the prediction is the “payout”.
  - The “gain” is the difference between the payout and the average prediction for all observations.
  - The Shapley value of a predictor is the average marginal contribution of that predictor across all possible coalitions of predictors.
- Formal decomposition (as presented):
  - 푏푏̂(푥푥푖)=휙휙0푆푆�푏푏̂�+�휙휙푘푆푆푛푘푘=1�푥푥푖;푏푏̂�
  - 휙휙푘푆푆�푥푥푖;푏푏̂�=
    �|
    푥푥′
    |
    !
    (
    푛푛−
    |
    푥푥′
    |
    −1
    )
    !
    푛푛!
    �
    �푏푏̂(푥푥푖|푥푥′∪{푥푥푘})−푏푏̂(푥푥푖|푥푥′)�
    푥푥′
    ⊆{푥푥1,푥푥2,...,푥푥푛}\{푥푥푘}
  - Definitions within the formula:
    - 푥푥′⊆{푥푥1,푥푥2, ...,푥푥푛}\{푥푥푘} is the set of coalitions used in the model prediction consisting of all predictors but the 푘푘th predictor.
    - �푥푥′� is the number of predictors included in each coalition except the 푘푘th predictor.
    - �푥푥′�! �푛푛−�푥푥′�−1�! / 푛푛! is the weighting factor.
    - 푏푏̂(푥푥푖|푥푥′∪{푥푥푘}) and 푏푏̂(푥푥푖|푥푥′) are the model predictions with and without the 푘푘th predictor conditional on the set of coalition consisting of all other predictors; their difference is the pay-off for including the 푘푘th predictor.

### Annex II — Event Study Results (summary)
- Purpose:
  - Reports results of the event study analysis for variables that show significant deviations around coups (including coups and attempted coups) relative to tranquil times.
- Indicators and subsamples highlighted (described as charts):
  - Economic Indicators: Growth Aid (% of GNI), Investment, Debt, Private credit, Energy inflation, Foreign Reserves, Tax Revenue, Fiscal Balance.
  - Socio-Political Stability Indicators: Reported Social Unrest, General Strikes, Anti Government Demonstration, Political Stability, Number of Major Cabinet Changes, Number of Major Constitutional Changes.
  - Governance Indicators: Regulatory Quality, Control of Corruption, Rule of Law, Government Effectiveness, Legislative Effectiveness, Voice and Accountability.
- Subsamples used in charts:
  - Baseline
  - LICs
  - SSA
  - 1970-1999
  - 2000-2019

### Annex III — Machine Learning Results (summary)
- Content:
  - Reports Shapley Values for the top-20 predictors of coups for four samples:
    - (i) the full sample (Annex Figure 7)
    - (ii) the subsample covering 2000-2019 (Annex Figure 8)
    - (iii) the sub-Saharan African sample for 1970-2019 (Annex Figure 9)
    - (iv) the Middle East and Central Asia sample for 1970-2019 (Annex Figure 10)
- Reading guidance (note repeated for figures):
  - The chart shows the contribution to coup probability (as measured by the Shapley value) for the top 20 drivers of the coup probability for the sample.
  - Each dot represents one observation (one country-year pair), and its color indicates the predictor value, with orange being low and green high.
  - The position of dots (left/right) indicates negative/positive contribution relative to the sample average; color and position together illustrate how the predictor value affects coup probability.
  - Examples provided in notes:
    - Full sample (1970-2019): share of elder population is the top predictor; most green dots (higher share of elder population) are to the left (reduce coup probability) and most orange dots (lower share) are to the right (increase coup probability).
    - 2000-19 sample: years since the last coup is the top predictor; green dots (higher years since last coup) are to the left (reduce coup probability) and orange dots (lower years since last coup) are to the right (increase coup probability).
    - Sub-Saharan African sample (1970-2019): income per capita is the top predictor; green dots (higher income per capita) are to the left (reduce coup probability) and orange dots (lower income per capita) are to the right (increase coup probability).
    - Middle East and Central Asia sample (1970-2019): share of elder population is the top predictor with the same left/right color interpretation as in the global sample.

### Annex IV — Selected case studies and stylized findings
- Sahel Region, 2020–23:
  - Context:
    - Post-2011 outflow of armed groups from Libya → heightened political fragility in Sahel.
    - Rise of armed conflicts and jihadist movements escalating political instability in Mali, Burkina Faso, Niger.
    - Developmental challenges: extreme poverty, low education, weak governance and institutions, dearth of public services.
  - Outcomes:
    - By 2020–23, deterioration in security despite international support; decline in wellbeing, large population displacements, corruption weakening security and public financial management.
    - Coup chronology noted: coup in Mali in 2020 followed by a string of seven coups or coup attempts in the Sahel region (two in Burkina Faso in 2022; one in Chad in 2021; another in Mali in 2021; two in Niger in 2021 and 2023).
    - Post-coup conditions have not demonstrated subsequent amelioration (Annex Figure 11).
- Mali (2020, 2021):
  - Eight coups since independence in 1960 (five successful); four coups since 2012.
  - 2012 coup led to election of President Keïta in 2013 (re-elected 2018); persistent security deterioration and corruption led to spiraling violence and socio-political tensions around 2020.
  - Pre-2020 coup factors: perceived rigging by Constitutional Court, corruption scandals, nepotism, frequent turnover of high-level officials.
- Burkina Faso (2022):
  - History: first coup in 1966; intense instability 1980–1987 with six successful coups.
  - Post-2012 surge in conflict and terrorist attacks spreading from Mali; 2014 ousting of President Blaise Compaoré.
  - By 2022, security challenges and failure to support military with equipment contributed to January 2022 military coup; continued deterioration led to another coup months later.
  - Since 2015: over 2,000 lives lost and 1.5 million displaced; intercommunity conflicts escalated.
- Niger (2021, 2023):
  - Since 1960, six coups, five successful.
  - 2021 election of President Mohamed Bazoum marked first conventional democratic transition; failed coup attempt occurred months after election.
  - July 2023 coup background: widespread conflicts, high poverty, insecurity along borders, Boko Haram activity; National Council for the Safeguarding of the Homeland (CNSP) cited deterioration of security and inadequate economic and social governance as motivations; proximate causes speculated (planned dismissal of head of Presidential Guard and/or forced retirement of five army generals).
- Sudan (2021):
  - Seventeen coups since 1956, six successful; multiple civil wars (1955–1972, 1983–2005) with over 1 million casualties and approximately 2 million displaced.
  - 2019 overthrow of President Omar al-Bashir after extensive demonstrations; subsequent civilian-military power-sharing Sovereignty Council planned transfer to civilians in November 2021.
  - Military staged successful coup in October 2021 after escalating political crisis and an unsuccessful coup one month earlier; government dissolved and state of emergency declared.
  - Since Q2 2023, armed conflict between two rival fractions of the military government; coup further adversely impacted economic activity, inflation, and fiscal position.
- Venezuela: long history of political fragility
  - Post-1914 oil discovery and end of Gómez rule in 1935 led to protracted political fragility.
  - From 1945 for twenty years: 1 successful political assassination and 14 coup attempts (three successful).
  - 1966–early 1990s: relative stability with coup-proofing strategies.
  - Recent history characterized by severe economic underperformance:
    - long recession over 1978-1989
    - extended growth stagnation over 1990-2003
    - growth collapse (or depression) over 2012–20
  - Chronic fragility features: low or negative growth, high levels of debt, high commodity dependency, governance weaknesses, lower social inclusion, rising poverty, political polarization.
  - Recent shocks and political events contributing to depression:
    - death of President Hugo Chavez in early March 2013
    - large negative 2014–16 terms-of-trade shock
    - imposition of escalating sanctions in August of 2017 by a block of western countries and government duality (from January 2019-February 2024)
  - Human consequences: reduced public caloric intake, increased disease and mortality, and millions displaced (as discussed with reference to Sachs and Weisbrot, 2019).
  - Coup attempts clustered in periods of macroeconomic stress: February and November 1992 (recession), 2002 (stagnation), and 2019 (depression).

*IMF Working Paper No. WP/24/34 — Political Fragility: Coups d’État and Their Drivers*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024034-print-pdf.pdf_