## wpiea2021079-print-pdf

## Source details

**Canonical URL:** [wpiea2021079-print-pdf](https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021079-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2021/english/wpiea2021079-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2021/english/wpiea2021079-print-pdf.pdf.json)

---

### Identifying social unrest events from daily media reports (Section 2.1)
- Purpose and scope
  - Develop a high-frequency (daily) identification of major social unrest events using spikes in daily press coverage; daily dataset constructed for events between 2011 and 2020 in 72 countries.
  - Financial-market linkage: stock prices used as high-frequency, forward-looking indicators; abnormal decreases in stock returns around events indicate investors view unrest as bad news.

- Data sources and article selection
  - Press aggregation service: Factiva; same set of major English-language international news media used by Barrett et al. (2020).
  - Article inclusion criteria:
    - Must include: Country name AND (“protest*” OR “riot*” OR “revolution” OR ((“civil” or “domestic”) within 10 words of “unrest”))
    - Must exclude: Country-specific terms OR “vote of protest” OR “protest vote” OR “protestant*” OR “anniversary” OR “war” OR “memorial” OR “movie”
    - Location tag: Countryk; Subject tag: Domestic Politics Or Civil Unrest; Word count: 100+

- Construction of the daily Social Unrest (SU) index
  - SU_kt = x_kt / (1/60 * sum_{t=-60}^{-1} x_kt), with x_kt the number of articles about social unrest in country k on day t, t ∈ [−60,90], and t = 0 on the first day of the event month.
  - Denominator: two-month pre-event average of relevant articles; normalization abstracts from country-specific/time-invariant media coverage biases.

- Event dating, threshold, and timing adjustments
  - Day 0: first day in the event month when SU_kt > 15 (i.e., articles exceed pre-event average by a factor of 15).
  - Threshold rationale: fifteen-fold increase ≈ relatively high threshold (~6 standard deviations of pre-event index); reduces false positives but excludes less severe events.
  - Under SU_kt > 15, 133 of the 336 monthly events are excluded from the daily event list.
  - Timing adjustments for merging with local stock markets: reassign day 0 to one (two) days prior if event happens on a Saturday (Sunday). Validity checks use unadjusted day 0.

- Event duration definition and caveats
  - End of event: last day of a series starting on day 0 when SU_kt > 1; duration = days between beginning and end; capped at 30 calendar days.
  - Duration used as proxy for severity; caveat that episodic resurgence can underestimate true duration.

- Sample and summary statistics (daily events with SU_kt > 15; averages computed for daily events)
  - Africa: Monthly events 68; Daily events 52; Average peak SU_kt 276.9; Average duration 2.3
  - Asia-Pacific: Monthly events 38; Daily events 21; Average peak SU_kt 183.5; Average duration 3.4
  - Europe: Monthly events 91; Daily events 58; Average peak SU_kt 107.1; Average duration 5.5
  - Middle East & Central Asia: Monthly events 91; Daily events 39; Average peak SU_kt 135.5; Average duration 5.5
  - Western Hemisphere: Monthly events 48; Daily events 32; Average peak SU_kt 88.8; Average duration 3.8
  - World: Monthly events 336; Daily events 203; Average peak SU_kt 99.3; Average duration 4.2
  - Note: Some monthly events excluded due to extremely high pre-event averages or end-month timing.

- Anticipation checks
  - Median SU_kt is very flat until day 0; median equals pre-event average even on day before event; median preferred to mean due to upside skew and outliers.

- Key empirical linkage referenced
  - Main empirical finding: on average, social unrest events lower cumulative abnormal returns by 1.4 percentage points over two weeks.
  - Effect stronger for longer-duration events and for events in emerging markets; mitigated by stronger institutions (better governance, more democratic systems).
  - Sample size for event-stock analysis: daily stock price data collected for 156 social unrest events (2011–2020).

### Validation of SU_kt (Section 2.2)
- Validation strategy
  - Three exercises comparing SU_kt to external narrative dates: Powell and Thyne (2011) coups, Worth (2016) dates for Arab Spring (Egypt and Tunisia), and various external sources for 2019 South America protests.

- Powell and Thyne (2011) coups comparison (Table 3; 11 matched events)
  - Exact preserved entries (date, weekday, days after coup):
    - Burkina Faso: Coup 2014-10-30 Thursday; Unrest 2014-11-01 Saturday; Days after coup 2
    - Mali: Coup 2012-03-22 Thursday; Unrest 2012-03-23 Friday; Days after coup 1
    - Turkey: Coup 2016-07-15 Friday; Unrest 2016-07-16 Saturday; Days after coup 1
    - Mali: Coup 2020-08-18 Tuesday; Unrest 2020-08-19 Wednesday; Days after coup 1
    - Egypt: Coup 2013-07-03 Wednesday; Unrest 2013-07-03 Wednesday; Days after coup 0
    - Burkina Faso: Coup 2015-09-17 Thursday; Unrest 2015-09-17 Thursday; Days after coup 0
    - Zimbabwe: Coup 2017-11-15 Wednesday; Unrest 2017-11-15 Wednesday; Days after coup 0
    - Sudan: Coup 2019-04-11 Thursday; Unrest 2019-04-11 Thursday; Days after coup 0
    - Burundi: Coup 2015-05-13 Wednesday; Unrest 2015-05-08 Friday; Days after coup -5
    - Egypt: Coup 2011-02-11 Friday; Unrest 2011-02-02 Wednesday; Days after coup -9
    - Thailand: Coup 2014-05-22 Thursday; Unrest 2014-05-08 Thursday; Days after coup -14
  - Key validation points:
    - Modal difference is zero; other differences explained by discrete time subdivision, rounding, month-boundary dating, or SU_kt capturing earlier unrest escalation preceding coups.

- Arab Spring validation (Egypt and Tunisia)
  - SU_kt red lines (SU_kt > 15) vs Worth (2016) dashed lines: exact match for Egypt (rightmost charts); one-day difference for Tunisia.
  - For early-2011 events, SU_kt sometimes dates later due to gradual tension build-up and the pre-event average denominator (example: Tunisia mass demonstrations after January 4 death; SU_kt starts January 7).
  - Robustness: authors exclude Arab Spring events in some checks to ensure results not driven solely by these episodes.

- 2019 South America protests validation
  - SU_kt dates coincide almost perfectly with external-source dates for Venezuela, Chile, Ecuador, Colombia in plotted windows.
  - Conclusion: close alignment between SU_kt and external narrative dates for 2019 wave.

- Overall assessment
  - SU_kt identifies beginnings of major social unrest events with precision; calibrated to detect unrest although search criteria do not include the word “coup”.
  - Advantages: consistent, transparent criteria; can be constructed with limited time lag.
  - Limitations: discrete-time edge cases, slow-building protests affect day 0 precision.
  - Robustness checks and additional validity figures presented in Appendix.

### Control variable regression and event-study main results (Section 3.4)
- Event-study cumulative abnormal returns (Figure 5)
  - Cumulative abnormal returns drop by around 0.72 percentage points three days after the unrest event relative to day -1.
  - Returns continue to decline, reaching -1.4 percentage points two weeks after the event.
  - 1.4 percentage point decline corresponds to the ninety-fifth percentile of the two-week cumulative returns distribution during the pre-event sample (unconditional probability 5 percent).
  - Negative impact persists significantly at least 25 business days after the unrest event; no reversion to zero over a longer horizon.
  - No significant pre-event trend in cumulative abnormal returns.

- Regression estimates (estimating equation (6), Table 4; dependent variable: stock market index returns; Newey-West SEs, 2 lags)
  - Event window [0, +3]:
    - Dummy for social unrest = -0.00165 ∗∗ (standard error (0.001))
  - Event window [0, +7]:
    - Dummy for social unrest = -0.00109 ∗ (standard error (0.001))
  - Country FE: Yes; Weekday FE: Yes; Observations: 15075 (both columns); F-stat reported.
  - Interpretation:
    - Social unrest reduces stock market returns by 16.5 basis points on average in first four days (0.00165).
    - Effect attenuates to 10.9 basis points over 8 business days (0.00109).
  - Anticipation effects: including day -1 reduces magnitude; little anticipation effect observed.

- Comparison across empirical approaches
  - Adjusted beta coefficients closely track cumulative abnormal returns from event-study; results robust to empirical approach.

- Heterogeneities: duration and income
  - Duration classification:
    - Short-term: < 3 days (54% of events).
    - Medium-term: 3–7 days.
    - Long-term: > 7 days (16% of sample; 25 out of 156 events); average duration for long events = 16.5 days; average event duration in sample = 4.5 days.
  - Duration effects (Figure 7):
    - Long-term events produce much larger reactions: cumulative abnormal returns decrease by as much as 8 percentage points in persistent instability cases.
    - Medium-term events contribute to baseline effect; short-term events drive less of overall decline.
  - Country income groups (Figure 8):
    - Decrease in cumulative abnormal returns driven by events in EMDE and LIDC.
    - Advanced economies show on average no significant negative effect.

- The role of institutions
  - Polity Score analysis (median = 6; Polity5 imputed for 2019–2020):
    - Negative effect pronounced only in countries with below median Polity Score.
    - In below-median Polity Score countries, cumulative abnormal returns drop by more than 3 percentage points over the 14-day window.
    - Above-median Polity Score countries: cumulative abnormal returns not significantly different from zero.
  - World Governance Indicators (WGI) analysis:
    - Regulatory Quality and Voice and Accountability show statistically significant impacts: social unrest affects cumulative abnormal returns only in countries with weak (below median) governance on these dimensions.
  - Overall: democracy, civil freedoms, government accountability, and quality of private-sector regulation mitigate negative stock market consequences.

- Control variable regression (estimating equation (8); Table 5; dependent variable: average 4-day abnormal returns ˆω_i)
  - Column (1) — duration only (N = 156):
    - Medium event = -0.00178 (0.002)
    - Long event = -0.00736 ∗∗ (0.004)
  - Column (2) — income only:
    - LIDC = -0.00283 (0.002)
    - EMDE = -0.00335 ∗∗∗ (0.001)
  - Column (3) — duration + income:
    - Medium event = -0.00185 (0.001)
    - Long event = -0.00807 ∗∗ (0.004)
    - LIDC = -0.00442 ∗ (0.002)
    - EMDE = -0.00361 ∗∗∗ (0.001)
  - Column (4) — duration + income + Polity (N = 155):
    - Medium event = -0.00149 (0.001)
    - Long event = -0.00778 ∗ (0.004)
    - LIDC = -0.00276 (0.002)
    - EMDE = -0.00258 ∗ (0.001)
    - Polity score above median = 0.00257 ∗ (0.001)
  - Interpretation:
    - Duration, income group, and Polity Score independently influence average abnormal returns.
    - Long-duration unrest significantly reduces average abnormal returns.
    - Democratic institutions partly mitigate the effect; inclusion of Polity reduces income-group coefficients, suggesting income may proxy for democracy when Polity omitted.

### The role of elections (Section 4.5)
- Method and identification
  - CLEA data used to create Elections dummy = 1 if lower chamber legislative elections occur within [-50, +50] business days around a social unrest event.
  - Analysis on reduced-country sample: of 140 social unrest events in reduced set, 21 simultaneous with lower chamber elections.
  - Estimated equation (6) with interaction SU×Elections; country FE absorb time-invariant country characteristics.

- Key regression evidence (Table 6; Newey-West SEs, 2 lags)
  - Dummy for social unrest: -0.00191 ∗∗ ; -0.00166 ∗∗ (standard errors (0.001)(0.001))
  - SU*Elections: 0.00295 ∗∗ ; 0.00214 ∗ (standard errors (0.001)(0.001))
  - Country FE: Yes; Weekday FE: Yes; Observations reported.
  - Findings:
    - SU×Elections coefficient positive and statistically significant.
    - Quantified effects reported by authors:
      - On average, social unrest in democratic countries leads to a 19 b.p. decline in stock market returns.
      - Social unrest related to election outcomes improves stock market performance by 10 b.p.
    - Implication: lack of CAR decline in more democratic countries explained primarily by quality of democratic institutions rather than other governance dimensions.

- Data notes
  - CLEA records elections dating back to 1788; reduced-country analysis excludes 16 events in Egypt, Hong Kong SAR, Jordan, Kyrgyzstan, Morocco, Mali, Panama, Qatar and Vietnam from the final sample of 156 events.

### Robustness and additional results (References and Appendix summaries)
- Alternative market models and windows
  - Figure A4 (MSCI ACWI market returns): cumulative abnormal returns relative to day0 between -0.03 and 0.01 over -10 to 50 business days.
  - Figure A5 (S&P Global Index): cumulative abnormal returns relative to day0 between -0.03 and 0.01 over -10 to 30 business days.

- Alternative thresholds and placebo checks
  - Larger sample with selection threshold SU_it > 10:
    - [0,+3]: Dummy for social unrest = -0.00165 ∗∗ (0.001)
    - [0,+7]: Dummy = -0.000782 (0.000); Observations: 17954
  - Placebo samples (border and non-border) show near-zero and statistically insignificant coefficients.

- Heterogeneity and interaction robustness (selected appendix tables)
  - Interaction with duration (Table A4): SU*Long significant and negative across windows (e.g., SU*Long = -0.00750 ∗∗ for [0,+3]).
  - Interaction with income (Table A5): SU*EMDE = -0.00281 ∗ for [0,+3].
  - Clustered events, trading volume, and other robustness tests reported; long events associated with significant increases in trading volume (Table A14: Long*Unrest = 0.289*** for [0,+3]).

- Validation appendix (Table A1, A2, Figures A2, A3)
  - Media source lists and event listing structure provided.
  - Examples: Bahrain SU_kt increased by more than 3000 times compared to pre-event average (denominator close to zero); SU_kt time series plotted for Korea, Tunisia, Spain, Mali with external-source alignment shown.

_Italic: Content derived from wpiea2021079-print-pdf (selected sections 2.1, 2.2, 3.4, 4.5, References/Appendix) provided in the source PDF._

### 2.1    Identifying social unrest events from daily media reports   .  .  .  .  .  .  .  .  .  .5

### 2.1 Identifying social unrest events from daily media reports

### Overview and motivation
- Purpose: develop a high-frequency (daily) identification of major social unrest events using spikes in daily press coverage, building on Barrett et al. (2020).
- Sample frame: countries covered by Barrett et al. (2020); daily dataset constructed for events between 2011 and 2020 in 72 countries.
- Financial-market linkage: stock prices are used as high-frequency, forward-looking indicators. If investors view social unrest as bad news, an abnormal decrease in stock returns around the event is expected.

### Data sources and article selection criteria
- Press aggregation service: Factiva.
- Sources: same set of major English-language international news media used by Barrett et al. (2020).
- Article inclusion criteria (Table 1):
  - Must include: Country name AND (“protest*” OR “riot*” OR “revolution” OR ((“civil” or “domestic”) within 10 words of “unrest”))
  - Must exclude: Country-specific terms OR “vote of protest” OR “protest vote” OR “protestant*” OR “anniversary” OR “war” OR “memorial” OR “movie”
  - Location tag: Countryk
  - Subject tag: Domestic Politics Or Civil Unrest
  - Word count: 100+

### Construction of the daily Social Unrest (SU) index
- Definition:
  - SU_kt = x_kt / (1/60 * sum_{t=-60}^{-1} x_kt)
    - x_kt is the number of articles about social unrest in country k per day t, where t corresponds to the day within the 5-month window relative to the beginning of the event month, t ∈ [−60,90], and t = 0 on the first day of the event month.
    - The denominator is the two-month pre-event average number of relevant articles.
- Rationale: normalizing by the pre-event average abstracts from country-specific and time-invariant media coverage biases and from media-bias in the time series.

### Event dating and thresholds
- Day 0 definition: the first day in the event month when SU_kt > 15 (i.e., the number of unrest-related articles on day t exceeds pre-event average by a factor of 15).
- Threshold rationale and trade-offs:
  - A fifteen-fold increase is described as a relatively high threshold (~6 standard deviations of the pre-event index).
  - High threshold reduces false positives but excludes less severe events.
  - Under SU_kt > 15, 133 of the 336 monthly events are excluded from the daily event list (see Table 2 for summary statistics).
- Timing adjustments:
  - In Section 4 the authors adjust day 0 for timing differences between event occurrence and filing of news articles across time zones (e.g., shifting for New York filings referencing Korea).
  - Reassign day 0 to one (two) days prior if the event happens on a Saturday (Sunday) to facilitate merging with local stock market data.
  - Validity checks use unadjusted day 0 as comparison is based on events at local time.

### Event duration
- End of event: the last day of a series starting on day 0 when SU_kt > 1.
- Duration: number of days between beginning and end of an event.
- Duration capped at 30 calendar days.
- Duration is used as a proxy for event severity; longer duration interpreted as more severe unrest.
- Caveat: waves or episodic resurgence of coverage can lead to underestimation of true unrest duration.

### Sample and summary statistics (Table 2)
- Monthly events (Barrett et al. (2020)) and daily events (SU_kt > 15):
  - Africa: Monthly events 68, Daily events 52, Average peak SU_kt 276.9, Average duration 2.3
  - Asia-Pacific: Monthly events 38, Daily events 21, Average peak SU_kt 183.5, Average duration 3.4
  - Europe: Monthly events 91, Daily events 58, Average peak SU_kt 107.1, Average duration 5.5
  - Middle East & Central Asia: Monthly events 91, Daily events 39, Average peak SU_kt 135.5, Average duration 5.5
  - Western Hemisphere: Monthly events 48, Daily events 32, Average peak SU_kt 88.8, Average duration 3.8
  - World: Monthly events 336, Daily events 203, Average peak SU_kt 99.3, Average duration 4.2
- Notes:
  - Monthly events are from Barrett et al. (2020).
  - Daily events are those that satisfy SU_kt > 15.
  - Averages are computed for the sample of daily events.
  - Some monthly events are excluded due to extremely high pre-event averages or end-month timing (e.g., the 2020 US racial justice protests example).

### Anticipation checks and pre-event behavior
- The median SU_kt is very flat until day 0; even on the day before an event, the median index equals the pre-event average.
- The median is preferred to the mean due to upside skew and outliers; median value is less than one within 10 days of event beginning.
- Authors conclude the method picks a clear starting point with no obvious change in media coverage before an event; Section 4 further checks robustness to anticipation effects.

### Key empirical findings referenced in this section and introduction
- Main empirical finding (from the paper overview): on average, social unrest events lower cumulative abnormal returns by 1.4 percentage points over two weeks.
- The effect is:
  - More pronounced for events that last longer.
  - More pronounced for events that occurred in emerging markets.
  - Mitigated by stronger institutions, in particular, better governance and more democratic systems.
- Sample size for event-stock analysis: daily stock price data collected for 156 social unrest events in the sample period (2011 to 2020).

*Source: wpiea2021079-print-pdf - 2.1 Identifying social unrest events from daily media reports*

### 2.2    Validation

### 2.2    Validation

### Validation exercises and purpose
- Three exercises compare the daily media-based social unrest index (SU_kt) to narrative descriptions from external sources to check validity.
- Coup d’état dataset of Powell and Thyne (2011) used as a first testbed because coups are well-defined, well-recorded, and publicly announced.
- Second exercise tests day 0 identification against external dates for the Arab Spring in Egypt and Tunisia (using Worth (2016) as cited in Barrett et al. (2020)).
- Third exercise compares 2019 South America protest dates with external sources (Venezuela: Briceño-Ruiz (2019) as cited in Barrett et al. (2020); Colombia: Latin America Monitor (2020); Chile and Ecuador: Aljazeera news reports).

### Comparison with Powell and Thyne (2011) coups (Table 3)
- Sample: 11 social unrest events matched to Powell and Thyne (2011) coup dates.
- Table 3 entries preserved exactly as presented:
  - Burkina Faso: Coup 2014-10-30 Thursday; Unrest 2014-11-01 Saturday; Days after coup 2
  - Mali: Coup 2012-03-22 Thursday; Unrest 2012-03-23 Friday; Days after coup 1
  - Turkey: Coup 2016-07-15 Friday; Unrest 2016-07-16 Saturday; Days after coup 1
  - Mali: Coup 2020-08-18 Tuesday; Unrest 2020-08-19 Wednesday; Days after coup 1
  - Egypt: Coup 2013-07-03 Wednesday; Unrest 2013-07-03 Wednesday; Days after coup 0
  - Burkina Faso: Coup 2015-09-17 Thursday; Unrest 2015-09-17 Thursday; Days after coup 0
  - Zimbabwe: Coup 2017-11-15 Wednesday; Unrest 2017-11-15 Wednesday; Days after coup 0
  - Sudan: Coup 2019-04-11 Thursday; Unrest 2019-04-11 Thursday; Days after coup 0
  - Burundi: Coup 2015-05-13 Wednesday; Unrest 2015-05-08 Friday; Days after coup -5
  - Egypt: Coup 2011-02-11 Friday; Unrest 2011-02-02 Wednesday; Days after coup -9
  - Thailand: Coup 2014-05-22 Thursday; Unrest 2014-05-08 Thursday; Days after coup -14
- Key findings from Table 3:
  - The modal difference between the two sets of dates is zero.
  - For four cases where SU_kt lags Powell and Thyne (2011), differences are largely due to discrete time subdivision and rounding (examples: Turkey — media coverage mainly next day; Mali — similar rounding; Burkina Faso — event straddles month boundary leading to month-based dating).
  - For three cases where SU_kt predates Powell and Thyne (2011) — Burundi, Egypt (February 2011), Thailand — SU_kt captures earlier social unrest escalation preceding the coup (Thailand example: SU_kt identifies the start of political crisis after May 7, 2014 Constitutional Court removal of the prime minister; coup followed two weeks later).

### Arab Spring (Egypt and Tunisia) validation (Figure 2)
- Barrett et al. (2020) identify two event months for each country: early 2011 (January for Tunisia, February for Egypt) and later post-revolutionary turmoil (Tunisia around October 2011 elections; Egypt around July 2013 government overthrow).
- SU_kt plotted with:
  - Red solid lines: dates identified using criterion SU_kt > 15.
  - Blue dashed lines: dates from Worth (2016).
- Alignment:
  - Rightmost charts: exact match for Egypt; one-day difference for Tunisia.
  - Leftmost charts (early-2011 events): gaps reflect gradual tension build-up and reasonable differences in start-date identification. Example: Worth (2016) cites December 16 self-immolation of Mohammed Bouazizi; mass demonstrations followed Mr. Bouazizi’s death on January 4 evening; SU_kt dates start of unrest to January 7.
- Note on sensitivity:
  - Slow-building protests can affect precision of day 0 due to pre-event average in SU_kt denominator; robustness checks exclude Arab Spring events to ensure results are not driven solely by these episodes.

### 2019 South America protests validation (Figure 3)
- SU_kt dates coincide almost perfectly with external-source dates for:
  - Venezuela (01nov2018–01apr2019 window plotted)
  - Chile (01aug2019–01jan2020 window plotted)
  - Ecuador (01aug2019–01jan2020 window plotted; first SU_kt spike outside the event month is not marked as day 0)
  - Colombia (01sep2019–01feb2020 window plotted)
- Red line: SU_kt > 15 criterion; blue dashed line: external-source date.
- Conclusion: close alignment between SU_kt and external narrative dates for the 2019 wave.

### Overall assessment of the media-based SU_kt index
- SU_kt can identify the beginning of major social unrest events with precision:
  - In all 11 coup cases, SU_kt dates match Powell and Thyne (2011) either exactly or with a small, explainable lag.
  - SU_kt is calibrated to detect any instances of unrest; the search criteria do not include the word “coup”.
- Advantages over narrative approaches and alternative data sources:
  - Criteria are consistent and transparent.
  - Eliminates the need to search for and verify additional sources.
  - Can be constructed without substantial time lag, enabling high-frequency dating of unrest events.
- Limitations and edge cases:
  - Any methodology that divides time into discrete intervals will fail in some edge cases (examples: events straddling month boundaries, late-night events).
  - Slow-building protests can affect day 0 precision through the pre-event average in the denominator.
- Robustness:
  - Additional validity checks are presented in Appendix (Figures A2 and A3).
  - Robustness to excluding Arab Spring events is tested (see Section 5).

*Source: wpiea2021079-print-pdf — Section 2.2 Validation*

### 3.4    Control variable regression

### 3.4    Control variable regression

### Methodology: control variable regression approach
- Re-estimate market model over the period of [-50, +50] days around the event:
  - R_is = α_i + β_i R_ms + ω_i 1[Unrest_is] + μ_is  (equation (7))
  - 1[Unrest_is] is a dummy equal to 1 on event-window days (for s ∈ [0, +3]).
  - Use the estimated coefficient ˆω_i (i.e. 4-day average abnormal returns for event i) as the dependent variable in:
    - ˆω_i = F_i Ψ + η_i  (equation (8))
- Advantages:
  - Allows inclusion of cross-sectional control variables F_i that don’t vary daily.
  - Enables simultaneous consideration of multiple explanatory variables (duration, country characteristics, institutions).
- Limitations:
  - Number of observations limited to the number of cross-sectional units (events).

### Main results — Cumulative abnormal returns (event-study)
- Key event-study findings (Figure 5):
  - Cumulative abnormal returns drop by around 0.72 percentage points three days after the unrest event relative to day -1.
  - Returns continue to decline, reaching -1.4 percentage points two weeks after the event.
  - A 1.4 percentage point decline corresponds to the ninety-fifth percentile of the two-week cumulative returns distribution during the pre-event sample (unconditional probability of observing such change in the pre-event sample is 5 percent).
  - Negative impact persists: cumulative abnormal returns remain significantly negative at least 25 business days after the unrest event (end of event window); no reversion to zero over a longer horizon (Appendix Figure A4).
  - No significant pre-event trend: cumulative abnormal returns on days before the event are not significantly different from zero.
- Comparisons to literature:
  - Epstein and Schnietz (2002): 1.86 percent decline for Fortune 500 after WTO protests (Seattle, 1999).
  - Chen and Siems (2004): Dow Jones declines of 0.59% after World Trade Center Bombing (1993) and Embassy Bombing in Kenya (1998).
  - Acemoglu et al. (2017): 8-day cumulative abnormal returns drop by 14.5 percentage points for Egypt 2011 protests (focused on firms connected to political power).

### Regression results (estimating equation (6))
- Table 4 (dependent variable – stock market index returns):
  - Event window [0, +3]:
    - Dummy for social unrest = -0.00165 ∗∗  (standard error (0.001))
  - Event window [0, +7]:
    - Dummy for social unrest = -0.00109 ∗  (standard error (0.001))
  - Country FE: Yes; Weekday FE: Yes.
  - Observations: 15075 (both columns).
  - F-stat: 1.969 ∗∗∗ (column 1), 1.949 ∗∗∗ (column 2).
  - Newey-West standard errors with 2 lags reported.
  - Significance notation: ∗ p <0.10, ∗∗ p <0.05, ∗∗∗ p <0.01.
- Interpretation:
  - Social unrest reduces stock market returns by 16.5 basis points on average in the first four days after the event (0.00165 = 16.5 basis points).
  - Effect attenuates over longer window: 10.9 basis points over 8 business days (0.00109 = 10.9 basis points).
  - Results statistically significant at the 5% level across approaches.

- Anticipation effects:
  - Including day -1 into the event window reduces the magnitude of the drop; coefficient β_1 remains negative and significant but much smaller in absolute value — implies little anticipation effect the day before the event.
  - Consistent with absence of pre-event trend in Figure 5.

- Comparison across empirical approaches:
  - Estimate equation (6) for 25 event windows and multiply β_1 by number of days in the event window to align with cumulative abnormal returns.
  - Adjusted beta coefficients closely track cumulative abnormal returns from event-study (Figure 6) — results robust to empirical approach.

### Heterogeneities: duration and income
- Event duration classification:
  - Short-term: last for less than 3 days.
  - Medium-term: last between 3 and 7 days.
  - Long-term: last for more than 7 days.
  - Sample facts:
    - Majority (54%) of events are short-term.
    - Long-term events constitute 16% of the sample (25 out of 156 events), with average duration of 16.5 days.
    - Average event duration in sample: 4.5 days.
- Duration effects (Figure 7):
  - Longer events produce larger stock market reactions.
  - Long-term events: cumulative abnormal returns decrease by as much as 8 percentage points in persistent social instability cases.
  - Medium-term events also contribute to the baseline effect; short-term events drive less of the overall average decline.
- Country income groups (Figure 8):
  - Countries split into advanced, EMDE (other emerging and developing economies), and LIDC (low-income developing countries).
  - Findings:
    - Decrease in cumulative abnormal returns driven by events in EMDE and LIDC.
    - Stock markets in advanced economies on average show no significant negative effect — estimates for advanced economies never significantly different from zero.
  - Note: Although CARs for LIDC are significant at 5% starting from day 8, regression coefficients for LIDC are not statistically significant due to wider Newey-West confidence intervals.

### The role of institutions
- Democracy (Polity Score) analysis:
  - Polity Score ranges from -10 to +10; median Polity Score in sample = 6 (threshold for grouping).
  - Impute 2018 Polity Score values for events in 2019 and 2020 (Polity5 dataset covers 1800-2018).
  - Results (Figure 9):
    - Negative effect of social unrest pronounced only in countries with below median Polity Score.
    - In less democratic countries (below median), cumulative abnormal returns drop by more than 3 percentage points over the 14-day event window.
    - In countries above median Polity Score, cumulative abnormal returns are never significantly different from zero.
- World Governance Indicators (WGI) analysis (Figure 10):
  - WGI indicators used: Rule of Law, Control of Corruption, Government Effectiveness, Regulatory Quality, Political Stability, Voice and Accountability (percentile rankings; split by median).
  - WGI data 2011-2019; impute 2019 values for events in 2020.
  - Findings:
    - Regulatory Quality and Voice and Accountability show statistically significant impacts:
      - Social unrest affects cumulative abnormal returns only in countries with weak (below median) governance on these dimensions.
    - Interpretation:
      - Regulatory Quality relates to governance features affecting private sector development (price controls, investment freedom, discriminatory taxes).
      - Voice and Accountability aggregates democracy, freedom of press, and political rights — similar interpretation to Polity Score.
- Overall:
  - Democracy, civil freedoms, government accountability, and quality of private-sector regulation mitigate negative stock market consequences of social unrest.
  - Markets in countries with weak institutions are most vulnerable to social instability.

### Control variable regression results (estimating equation (8); Table 5)
- Dependent variable: average 4-day abnormal returns (ˆω_i).
- Table 5 summary (columns (1)–(4); N = 156 for columns 1–3, N = 155 for column 4):
  - Column (1) — duration only:
    - Medium event = -0.00178  (standard error (0.002))
    - Long event = -0.00736 ∗∗  (standard error (0.004))
  - Column (2) — income only:
    - LIDC = -0.00283  (standard error (0.002))
    - EMDE = -0.00335 ∗∗∗  (standard error (0.001))
  - Column (3) — duration + income:
    - Medium event = -0.00185  (0.001)
    - Long event = -0.00807 ∗∗  (0.004)
    - LIDC = -0.00442 ∗  (0.002)
    - EMDE = -0.00361 ∗∗∗  (0.001)
  - Column (4) — duration + income + Polity:
    - Medium event = -0.00149  (0.001)
    - Long event = -0.00778 ∗  (0.004)
    - LIDC = -0.00276  (0.002)
    - EMDE = -0.00258 ∗  (0.001)
    - Polity score above median = 0.00257 ∗  (0.001)
  - Standard errors in parentheses. Significance: ∗ p <0.10, ∗∗ p <0.05, ∗∗∗ p <0.01.
- Interpretation:
  - Duration, country income group, and Polity Score are simultaneously and independently important determinants of average abnormal returns.
  - Long-duration unrest implies starker decrease in average abnormal returns (long event coefficient significant across specifications).
  - Democratic institutions partly mitigate the effect (positive coefficient for Polity score above median in column (4)).
  - Inclusion of Polity Score reduces magnitude of income group coefficients, suggesting country income may proxy for democracy in specifications omitting Polity dummy.

*Source: wpiea2021079-print-pdf — 3.4    Control variable regression*

### 4.5    The role of elections

### 4.5    The role of elections

### Method and identification
- Used Polity Score as a measure of democratic institutions; concern that it may proxy for other governance quality dimensions.
- Employed Constituency-Level Elections Archive (CLEA) data to identify lower chamber legislative elections at constituency level and create an Elections dummy equal to one if elections occur within [-50, +50] business days around a social unrest event.
- Analysis conducted on a reduced set of countries (CLEA covers countries with regular elections); out of 140 social unrest events in the reduced set, 21 are simultaneous with lower chamber elections.
- Estimated equation (6) with an interaction term SU×Elections; country fixed effects absorb time-invariant country characteristics so the interaction coefficient captures difference between election-related unrest and other unrest events within countries.

### Key regression evidence (Table 6)
- Results include interaction terms between Social Unrest dummy and Election dummy.
- Reported coefficients (as presented):
  - Dummy for social unrest: -0.00191 ∗∗ ; -0.00166 ∗∗ (standard errors (0.001)(0.001))
  - SU*Elections: 0.00295 ∗∗ ; 0.00214 ∗ (standard errors (0.001)(0.001))
  - Country FE: Yes
  - Weekday FE: Yes
  - Observations1352913529
  - Newey-West standard errors with 2 lags in parentheses. Significance: ∗ p <0.10, ∗∗ p <0.05, ∗∗∗ p <0.01

### Findings on elections and stock markets
- The coefficient of SU×Elections is positive and statistically significant.
- Quantified effects:
  - On average, social unrest in democratic countries leads to a 19 b.p. decline in stock market returns.
  - Social unrest related to election outcomes improves stock market performance by 10 b.p.
- Implications:
  - Confirms absence of stock market reaction to unrest in countries with above median Polity Score and above median Voice and Accountability indicator.
  - Suggests that the lack of statistically significant decrease in CAR is explained primarily by the quality of democratic institutions rather than by differences along other institutional dimensions (such as governance).
- The paper indicates further mechanism discussion will appear in Section 6.

### Data notes and sample adjustments
- CLEA records elections dating from as early as 1788.
- The reduced-country analysis excludes 16 events in Egypt, Hong Kong SAR, Jordan, Kyrgyzstan, Morocco, Mali, Panama, Qatar and Vietnam from the final sample of 156 events.

*Source: wpiea2021079-print-pdf - 4.5    The role of elections*

### References

### wpiea2021079-print-pdf - References

### Main results
- Cumulative abnormal returns (market model, market returns based on MSCI ACWI; event window – 50 business days):
  - Figure A4 shows cumulative abnormal returns relative to day0 between -.03 and .01 over -10 to 50 business days since event.
- Cumulative abnormal returns (market model, market returns based on S&P Global Index):
  - Figure A5 shows cumulative abnormal returns relative to day0 between -.03 and .01 over -10 to 30 business days since event.
- Table A3: Estimating equation (6); dependent variable – stock market index returns. Alternative event windows to evaluate anticipation effects:
  - [-1, +3]: Dummy for social unrest = -0.00121* (0.001)
  - [-1, +7]: Dummy for social unrest = -0.000904* (0.001)
  - Country FE: Yes; Weekday FE: Yes; Observations: 15075
  - Newey-West standard errors with 2 lags in parentheses.
  - Significance notation: * p <0.10, ** p <0.05, *** p <0.01

### Heterogeneities: regression results
- Table A4: Interaction between Social Unrest dummy and duration groups; dependent variable – stock market index returns:
  - [0,+3]:
    - Dummy for social unrest = 0.000214 (0.001)
    - SU*Medium = -0.00225* (0.001)
    - SU*Long = -0.00750** (0.003)
    - Weekday FE: Yes; Event FE: Yes; Observations: 15075
  - [0,+7]:
    - Dummy for social unrest = 0.000146 (0.000)
    - SU*Medium = -0.000141 (0.001)
    - SU*Long = -0.00752*** (0.003)
    - Weekday FE: Yes; Event FE: Yes; Observations: 15075
  - Newey-West standard errors with 2 lags in parentheses.
- Table A5: Interaction between Social Unrest dummy and income groups; dependent variable – stock market index returns:
  - [0,+3]:
    - Dummy for social unrest = 0.000513 (0.001)
    - SU*LIDC = -0.00265 (0.002)
    - SU*EMDE = -0.00281* (0.001)
    - Country FE: Yes; Weekday FE: Yes; Observations: 15075
  - [0,+7]:
    - Dummy for social unrest = -0.000114 (0.001)
    - SU*LIDC = -0.00124 (0.001)
    - SU*EMDE = -0.00124 (0.001)
    - Country FE: Yes; Weekday FE: Yes; Observations: 15075
  - Newey-West standard errors with 2 lags in parentheses.

### Robustness checks
- Table A6: Additional significance tests (multiple test methods, CAR windows reported):
  - CAR[0,3], CAR[0,7], CAR[0,14], CAR[0,25], CAR[1,1] reported across Patell test, Boehmer-Musumeci-Poulsen test, and Generalised Rank test with Kolari and Pynnonen adjustment.
  - Portfolio CARs (Ptf CARs): -0.66%***, -0.85%**, -1.16%**, -1.38%**, -0.36%***
  - CAAR: -0.62%**, -0.76%, -0.99%, -1.12%, -0.36%**
  - Notes: Cumulative abnormal returns in this table are calculated using a Historical Mean Model (without controlling for global stock market returns). Tests account for cross-sectional correlation and autocorrelation; GRANK relaxes normality assumption.
- Table A7: Larger sample with selection threshold SU_it >10; dependent variable – stock market index returns:
  - [0,+3]: Dummy for social unrest = -0.00165** (0.001)
  - [0,+7]: Dummy for social unrest = -0.000782 (0.000)
  - Country FE: Yes; Weekday FE: Yes; Observations: 17954
  - Newey-West standard errors with 2 lags in parentheses.
- Table A8: Larger sample with duration interactions (SU_it >10):
  - [0,+3]: Dummy for social unrest = -0.000444 (0.001); SU*Long = -0.00553** (0.003)
  - [0,+7]: Dummy for social unrest = 0.0000172 (0.000); SU*Long = -0.00523*** (0.002)
  - Weekday FE: Yes; Event FE: Yes; Observations: 17954
- Table A9: Larger sample with income-group interactions (SU_it >10):
  - [0,+3]: Dummy for social unrest = -0.00102 (0.001); SU*LIDC = -0.00131 (0.002); SU*EMDE = -0.000501 (0.001)
  - [0,+7]: Dummy for social unrest = -0.0000143 (0.001); SU*LIDC = -0.00103 (0.001); SU*EMDE = -0.000932 (0.001)
  - Country FE: Yes; Weekday FE: Yes; Observations: 17954
- Placebo samples:
  - Table A10 (border placebo): [0,+3] Dummy for social unrest = 0.000470 (0.001); [0,+7] = 0.0000249 (0.000); Observations: 9755
  - Table A11 (non-border placebo): [0,+3] Dummy for social unrest = -0.0000475 (0.000); [0,+7] = -0.000283 (0.000); Observations: 11538
  - Country FE: Yes; Weekday FE: Yes in both.

### Additional results
- Table A12: Interaction between Social Unrest dummy and Cluster dummy; dependent variable – stock market index returns:
  - [0,+3]: Dummy for social unrest = -0.00159** (0.001); SU*Clustered event = -0.000570 (0.002); Observations: 15075
  - [0,+7]: Dummy for social unrest = -0.00107* (0.001); SU*Clustered event = -0.000141 (0.002); Observations: 15075
  - Country FE: Yes; Weekday FE: Yes
- Trading volume (Tables A13 and A14):
  - Table A13: Dependent variable – log trading volume:
    - [0,+3]: Social unrest = 0.0552 (0.060)
    - [0,+7]: Social unrest = 0.0406 (0.043)
    - Country FE: Yes; Weekday FE: Yes; Observations: 84378
  - Table A14: Dependent variable – log trading volume with duration interactions:
    - [0,+3]: Social unrest = -0.0144 (0.072); Medium*Unrest = -0.00128 (0.112); Long*Unrest = 0.289*** (0.100)
    - [0,+7]: Social unrest = 0.0105 (0.046); Medium*Unrest = -0.0667 (0.081); Long*Unrest = 0.195*** (0.067)
    - Event FE: Yes; Weekday FE: Yes; Observations: 84378
  - Newey-West standard errors with 2 lags in parentheses.

### Appendix — Measuring Social Unrest: sources, event listing, validation
- Table A1: Sources of media reports
  - US sources: The ABC Network, the CBS Network, the NBC Network, the Los Angeles Times, the Boston Globe, the New York Times, the Chicago Tribune, the Wall Street Journal, the Washington Post
  - UK sources: The BBC, the Financial Times, the Telegraph U.K., the Times U.K., the Telegraph, the Guardian U.K., the Economist
  - Canadian sources: The Canadian Broadcasting Corp, the Globe and Mail
- Table A2: Event listing structure and sample inclusion criteria
  - Column definitions: “Event” (brief description using SU_kt), “Daily identification” (N means day of beginning not found based on SU_kt >15), “Duration” (L for long, M for medium, S for short; see Section 4.3), “Financial data available” (Y/N). Only events with Y in both column 4 and column 6 are in main sample.
  - Example entries (date, country/region, event, daily identification, duration, fin data available): 
    - Jan 2011, Lebanon, Constitution protests start, Y, S, Y
    - Jan 2011, Tunisia, Tunisia Revolution, Y, L, Y
    - Feb 2011, Uganda, Election protest, Y, S, Y
    - Feb 2011, Egypt, Mubarak Resigns, Y, L, Y
    - ... (full listing continues through Jan 2020 and event classification codes Y/S/Y etc.)
- Counts and sample:
  - Number of events categories shown as 5-7 3-4 1-2 Not in the sample (as presented)
  - Number of Social Unrest events, 2011-2020: sample of 336 monthly events (Figure A1)
- Validation figures and notes:
  - Figure A2: SU_kt time series for Bahrain and Morocco. Note: media coverage of Bahrain’s unrest increased by more than 3000 times compared to pre-event average; before February 2011 media coverage of Bahrain was essentially zero (denominator close to zero).
  - Figure A3: SU_kt time series for Korea, Tunisia, Spain, Mali with red line date identified using criterion SU_kt >15 and blue dashed line dates from external sources. Sources: Kim (2017) as cited in Barrett et al. (2020) for Korea, Barrett et al. (2020) for Tunisia, BBC and Aljazeera news reports for Spain and Mali.

_Italic: Content derived from the References and Appendix sections of the provided PDF file._

---


_Source: https://www.imf.org/-/media/files/publications/wp/2021/english/wpiea2021079-print-pdf.pdf_