2.1 Identifying social unrest events from daily media reports

We start with the monthly event list from Barrett et al. (2020). The authors use the number of articles from major international news sources to create indices of social unrest for 130 countries. These indices exhibit very large spikes that are associated with major episodes of social unrest. The authors develop quantitative criteria to formalize these spikes and identify a list of social unrest events at the country-month frequency. They compare these events against consensus narratives for a number of case studies, showing that they align closely, and conclude that this method captures actual major social unrest events.³

We use Factiva, a press aggregation service, to search for articles in a short window around the 336 country-month unrest events identified by Barrett et al. (2020) between 2011 and 2020. We use the same set of sources (major English-language international news media – see Appendix Table A1 for a list of sources) and apply the same search criteria as Barrett et al. (2020) to identify unrest-related media reports. The relevant articles related to country k must include words such as “protest”, “riot” or “revolution”. They must also exclude certain terms to avoid counting reports about previous events or revolution-themed movies. The selected articles must be at least 100 words long and must mention the name of the country in question (see Table 1 for details).

^{Table 1:}

Article search criteria

^{Table 1:}

Article search 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	Country k
Subject tag	Domestic Politics Or Civil Unrest
Word count	100+

View Table

^{Table 1:}

Article search 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	Country k
Subject tag	Domestic Politics Or Civil Unrest
Word count	100+

While there is of course some degree of arbitrariness in the search terms and sources, we justify our approach on two grounds. First, as Barrett et al. (2020) show, the approach works well at monthly frequency: social unrest events are sufficiently obvious and interesting to international press, so that they are consistently identified regardless of the specifics of the search. The identification works particularly well for high-profile events, such as the Egyptian revolution of February 2011, the abrupt overthrow of the government of Burkina Faso in November 2014, the Latin American protests of Autumn 2019, and the Belarusian protests starting in Summer 2020. Events as newsworthy as these are relatively forgiving of the precise details of a media-based methodology for identifying events. Second, in Section 2.2 we test our identification strategy in several case studies. If our approach performs well against reasonable external descriptions of events, robustness to variation in languages, sources and search terms is ensured. We return to this point in Section 2.2.

We download all articles that match this set of search criteria for 2 months before and after each event month. Restrictions on the volume of daily-frequency queries in Factiva preclude downloading longer daily time series. This is why our starting point using Barrett et al. (2020) is essential: it helps us to narrow down the time window. We then count the number of articles filed on each day during this 5-month window. We use this article count to compute a daily Social Unrest (SU) index:

where 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, such that t = 0 on the first day of the event month, and t ∈ [—60, 90]), and $$\frac{1}{60}{\Sigma }_{t=-60}^{-1}{x}_{kt}$$ is the two-month pre-event average number of relevant articles. Normalizing the number of articles relative to the pre-event average ensures that the SU_kt abstracts from any factors that are constant during the country-specific event period. It excludes all forms of media bias which results in more or less media coverage in a given country. Because the numerator is calculated for each event, this normalization also eliminates potential media bias in the time series, for example, due to changing perceptions of a given country.

Timing is a crucial element of the event-study methodology. Our working hypothesis, which we verify in Section 2.2, is that major social unrest events occur suddenly and generate sizeable press coverage. It is therefore important that we identify the exact date that marks the beginning of each unrest event in our sample. This is what we refer to as “day 0”. We designate day 0 to be the first day in the event month when SU_kt > 15, i.e. when the number of unrest-related articles on day t exceeds pre-event average by a factor of 15.⁴ Hereafter, we use s to denote the number of days relative to day 0 (thus, s = 0 is the day when SU_kt > 15 for the first time).

There is a trade-off in choosing the threshold to identify event dates. A high threshold has the advantage of having a lower likelihood of false positive. At the same time, it is more likely to exclude less severe events. A fifteen-fold increase – about 6 standard deviations of the pre-event index – is a relatively high threshold. Under this threshold, 133 of the 336 monthly events are excluded from the daily event list. These include Arab Spring events in Jordan and Syria in 2011, Togolese protests in 2017, and the racial justice protests in the US in June 2020 (see Table 2 for details on sample attrition, and Table A2 in the Appendix for the full list of events).⁵ In Section 5, we test the robustness of our results to the choice of threshold and the alternative event identification criteria.

^{Table 2:}

Social unrest events 2011–2020: Summary Statistics

^a Monthly events are from Barrett et al. (2020), Daily events are those that satisfy SU_kt > 15. ^b Averages are computed for the sample of daily events.

^{Table 2:}

Social unrest events 2011–2020: Summary Statistics

Region	Monthly events	Daily events	Average peak SUkt	Average duration
Africa	68	52	76.9	2.3
Asia-Pacific	38	21	83.5	3.4
Europe	91	58	107.1	5.5
Mid. East & Cent. Asia	91	39	135	5.5
Western Hemisphere	48	33	88.8	3.8
World	336	203	99.3	4.2

^a Monthly events are from Barrett et al. (2020), Daily events are those that satisfy SU_kt > 15. ^b Averages are computed for the sample of daily events.

View Table

^{Table 2:}

Social unrest events 2011–2020: Summary Statistics

Region	Monthly events	Daily events	Average peak SUkt	Average duration
Africa	68	52	76.9	2.3
Asia-Pacific	38	21	83.5	3.4
Europe	91	58	107.1	5.5
Mid. East & Cent. Asia	91	39	135	5.5
Western Hemisphere	48	33	88.8	3.8
World	336	203	99.3	4.2

^a Monthly events are from Barrett et al. (2020), Daily events are those that satisfy SU_kt > 15. ^b Averages are computed for the sample of daily events.

An important issue for financial market event studies is whether the event could have been reasonably predicted. If so, the market would have reacted when information about the forthcoming event was revealed, rather than at the time of the event itself. To check potential anticipation effects, we plot the median index in Figure 1. The median of SU_kt is very flat until day 0. Even on the day before an event, the value of the median index is equal to the pre-event average.⁶ Overall, this suggests that our method for identifying the daily start of social unrest events picks a clear starting point with no obvious change in media coverage before an event. In Section 4 we further check the robustness of our results to anticipation effects.

View Full Size

Mean SU_kt (purple line) and median SU_kt (green line).

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

We also use SU_kt to determine the duration of an event. We define the end of an event as the last day of a series starting on day 0 when SU_kt > 1, and define the duration of an event as the number of days between the beginning and the end of an event.⁷ We think of the duration of the event as a proxy for event severity: the longer the duration, the more severe the unrest. We exploit this assumption in Section 4.3. Of course, this is not without drawbacks. In particular, it is possible that the media coverage of the event decreases for some time, and then recovers rapidly as unrest escalates again. This is more likely to be the case when there are waves of tightly bunched protest events. In this case, our measure of duration would only capture the length of the initial round of unrest, thus underestimating the true unrest duration.

An interesting implication of our approach is in regard to the definition of what constitutes social unrest. Drawing tight boundaries around the definition of social unrest is inherently a very challenging task, and we do not aim to do this. Instead, we suggest a systematic method to broadly reflect a consensus view of major social unrest events. Of course, if our sources do not accurately reflect this consensus view – perhaps due to bias, herding effects or challenges to reporting in particular countries – neither will our measure. This is exactly why we devote the next section to the validation of our event-dating methodology.

2.2 Validation

To check the validity of our daily event identification, we run three exercises to compare our data to narrative descriptions from external sources. In the first exercise, we compare our data with a peer-reviewed dataset of coups d’état compiled by Powell and Thyne (2011). One advantage of using coups as a testbed for our method is that coups are particularly well-defined and well-recorded events: those engaged in them have obvious incentives to be quick, unexpected, and to announce their actions publicly.

Table 3 lists the start dates of 11 social unrest events from our data and compares them with the dates of coups d’état or coup attempts from Powell and Thyne (2011). The modal difference between two sets of dates is zero. For the four cases where our measure lags that of Powell and Thyne (2011), the difference is essentially a rounding issue, arising because we subdivide time into discrete units. In Turkey, military activity began very late at night on July 15. As a result, most media coverage of this event occurred on the following day. Moreover, given that the event itself did not begin until after markets had closed on the 15^th, the 16^th of July is the right dating from a financial markets’ perspective. Rounding similarly explains the difference in the case of Mali. In Burkina Faso, the difference arises because the event straddles the start of a new month.⁸

^{Table 3:}

Comparing dates of social unrest with Coup d’etat Dataset

^{Table 3:}

Comparing dates of social unrest with Coup d’etat Dataset

Country	Coup	Weekday	Unrest	Weekday	Days after coup
Burkina Faso	2014–10-30	Thursday	2014–11-01	Saturday	2
Mali	2012–03-22	Thursday	2012–03-23	Friday	1
Turkey	2016–07-15	Friday	2016–07-16	Saturday	1
Mali	2020–08-18	Tuesday	2020–08-19	Wednesday	1
Egypt	2013–07-03	Wednesday	2013–07-03	Wednesday	0
Burkina Faso	2015–09-17	Thursday	2015–09-17	Thursday	0
Zimbabwe	2017–11-15	Wednesday	2017–11-15	Wednesday	0
Sudan	2019–04-11	Thursday	2019–04-11	Thursday	0
Burundi	2015–05-13	Wednesday	2015–05-08	Friday	-5
Egypt	2011–02-11	Friday	2011–02-02	Wednesday	-9
Thailand	2014–05-22	Thursday	2014–05-08	Thursday	-14

View Table

^{Table 3:}

Comparing dates of social unrest with Coup d’etat Dataset

Country	Coup	Weekday	Unrest	Weekday	Days after coup
Burkina Faso	2014–10-30	Thursday	2014–11-01	Saturday	2
Mali	2012–03-22	Thursday	2012–03-23	Friday	1
Turkey	2016–07-15	Friday	2016–07-16	Saturday	1
Mali	2020–08-18	Tuesday	2020–08-19	Wednesday	1
Egypt	2013–07-03	Wednesday	2013–07-03	Wednesday	0
Burkina Faso	2015–09-17	Thursday	2015–09-17	Thursday	0
Zimbabwe	2017–11-15	Wednesday	2017–11-15	Wednesday	0
Sudan	2019–04-11	Thursday	2019–04-11	Thursday	0
Burundi	2015–05-13	Wednesday	2015–05-08	Friday	-5
Egypt	2011–02-11	Friday	2011–02-02	Wednesday	-9
Thailand	2014–05-22	Thursday	2014–05-08	Thursday	-14

For the three remaining events in which we pre-date the coup identified by Powell and Thyne (2011) – Burundi, Egypt in February 2011 and Thailand – the coups were a consequence of previous social unrest, and thus the dates we identify are the days of the earlier unrest escalation, and not the coup itself. In Thailand, for instance, our approach identifies exactly the start of a political crisis resulting from the removal of the prime minister by the Constitutional Court on May 7^th 2014. A coup by the Army followed two weeks later.

This exercise clearly shows that media-based social unrest index can be used to identify the beginning of major social unrest events with precision. In all of the 11 coup cases the dates from SU_kt criterion match the Powell and Thyne (2011) dataset either exactly, or with a small reasonable lag. The precision of the results is even more striking considering that SU_kt is calibrated to search for any instances of unrest, and the word “coup” is not even a part of our article search criteria.

In our second verification exercise, we evaluate the validity of day 0 identification against external data sources, using the Arab Spring in Egypt and Tunisia as examples. These countries were two of the most affected by the events of 2011, with governments in both countries falling as a result of the protests. Barrett et al. (2020) identify two event months for each country around this time: one in early 2011 (January for Tunisia, February for Egypt), and one later as post-revolutionary political turmoil continued (around the October 2011 elections in Tunisia, and the overthrow of the government in July 2013 in Egypt). Figure 2 plots SU_kt for these periods, marking the dates identified in our data (red solid lines), as well as those from a reliable external source Worth (2016) (blue dashed lines).

View Full Size

SU_kt. Red line – date identified using criterion SU_kt > 15; blue dashed lines – dates from an external source.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

Source: Worth (2016) as cited in Barrett et al. (2020)

• Download Figure
• Download figure as PowerPoint slide

In the rightmost charts, the alignment of our method with the other source is clear, with no difference in Egypt and a difference of only one day in Tunisia. For the early-2011 events (leftmost charts), the gap between the dating of our data and Worth (2016) reflects reasonable differences in identifying the start of Arab Spring when the tension built up over a series of events. For example, in Tunisia the first event cited by Worth (2016) is the December 16^th self-immolation of Mohammed Bouazizi. Despite some protests in late December, it was not until the death of Mr. Bouazizi in the evening of January 4 that mass demonstrations began. We date the start of unrest to January 7, which is very close to this.⁹

Of course, the slow-building nature of some protests could affect the precision of day 0, identified in our data, since media coverage of unrest in the preceding months affects SU_kt through the pre-event average in the denominator. To ensure that these events do not introduce bias, and that our results are not driven by the Arab Spring events alone, we check the robustness of our results after excluding Arab Spring events from our sample (see Section 5).

In the third verification exercise, we compare our data on the 2019 wave of protests in South America with external sources. Figure 3 shows that the dates identified in our data coincide almost perfectly with those from external sources. Further validity checks are presented in Appendix (Figures A2 and A3).

View Full Size

SU_kt. Red line – date identified using criterion SU_kt > 15; blue dashed line – date from external sources.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

Sources: Briceño-Ruiz (2019) as cited in Barrett et al. (2020) for Venezuela, Latin America Monitor (2020) for Colombia, Aljazeera news reports for Chile and Ecuador. Since the first spike of the SU_kt in Ecuador is outside of the event month (October), it is not marked as day 0.

• Download Figure
• Download figure as PowerPoint slide

Overall, our high-frequency media-based index dates social unrest events with reasonable accuracy. More importantly, it has several advantages over narrative approaches and alternative data sources in identifying and dating social unrest events. Its criteria are consistent and transparent. It eliminates the need to search for and verify additional sources of information. And unlike other sources of unrest data, it can be constructed without the substantial time lag.

3.1 Data

We merge our data on social unrest events with stock market indices for each country. Since the latter is not available for some countries (in particular some of the Middle Eastern and African countries), we drop 47 of the 203 daily events. Given that stock market data is only available on business (rather than calendar) days, we gather stock index information for approximately 50 business days either side of day 0.¹⁰ To adjust for events that happen over the weekend, we reassign day 0 to one (two) day(s) prior if the event fell on a Saturday (Sunday). To take into account possible time difference between the event and media coverage, we move day 0 one day back in countries from the Asia and Pacific region.

Our final sample contains 156 social unrest events across 72 countries, with over 15,000 daily observations. Our country sample has a broad coverage by geographical region and income group (Figure 4). The average duration of social unrest event in our final sample is 4.5 days.¹¹ Despite losing some events, our sample includes some of the most well-known episodes of social unrest in the last decade (see Appendix Table A2 for a full list of events). As the country maps make clear, unrest often occurs in regional clusters, most notably the Arab Spring events in the Middle East in 2011 and the series of protests throughout Latin America in 2019.¹²

View Full Size

Countries with social unrest events in 2011–2020: final sample with both daily identification and stock market data (156 events). For a map illustrating the sample of monthly-identified events (336 events), see Figure A1.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

Data on stock indices, global and regional indices, as well as trading volume, exchange rate and bond yield come from Bloomberg and Datastream. Additional variables such as GDP per capita, market capitalization, inflation, interest rates, current account balance and government debt are from the IMF’s International Financial Statistics and World Economic Outlook, and the World Federation of Exchanges Database.

3.2 Event-study approach

We follow the standard event-study methodology (MacKinlay (1997)) to quantify the impact of social unrest events on stock returns. First, we use the market model approach to calculate normal returns for each event i:

where R_is are the returns of stock market index i on day s, and R_ms are the returns of the market portfolio on this day.¹³ Returns are measured as difference in log of stock market index between two business days. Given the cross-country nature of our study, in our headline results we use the returns of global stock market index MSCI ACWI as market returns.¹⁴ Equation (2) is estimated using OLS over the estimation period of [-50, -11] business days before the event. The estimation is performed for each event i separately – the estimated coefficients a_i and b_i are event-specific. We then compute abnormal returns (AR_is) as the difference between actual returns and the returns implied by the market model:

where â_i and $${\hat{b}}_i$$ are estimated in equation (2). Finally, we aggregate abnormal returns across events:

Hence, for each day s in the event window we obtain the estimate of average abnormal returns across all social unrest events in our sample.

The cumulative abnormal returns are the sum of average abnormal returns between day 0 (day of the event) and day τ:

The significance of both abnormal and cumulative abnormal returns is estimated using the Brown and Warner (1985) t-statistics.¹⁵

The event window is an important choice in the estimation. A short event window reduces the likelihood that confounding factors may contaminate the result and is suited for short events, such as an announcement of merger and acquisition or debt relief (see Campbell et al. (2010) for a literature summary). Yet a long event window is suited for big events that take time to unfold. For example, Forbes (2004) uses 2- and 12-weeks event windows to study the effect of the Asian financial crisis on stock returns. Social unrest events may be short or long. A protest may last for a day; coups may last for a few days; major events such as Arab Spring protests lasted for weeks. We address this nature of social unrest events in several ways. We estimate abnormal returns for up to 25 business days after the unrest event.¹⁶ This allows us to capture potential longer impacts of the events. For the regression analysis in Section 3.3, we focus on 4- and 8-day windows because, as we will show, the largest effects tend to occur in the immediate aftermath of an event.¹⁷ The crucial identification assumptions of the event study are that the timing of social unrest events is exogenous and correctly identified. As discussed earlier, we took careful steps to ensure the accuracy of the event timing in our dataset. This accuracy is also important for the exogeneity assumption. This is because even when the outburst of social discontent is expected to some degree (for instance, if unrest is caused by an economic shock, or by a controversial political decision), the exact timing of a social unrest event is rarely predictable and thus close to random.¹⁸ According to Figure 1, there is no evidence on anticipation effects in media coverage. As we will show, the unpredictability of the timing of events is further corroborated by the absence of pre-event trends in cumulative abnormal returns.

A common challenge to identification in event studies is that other shocks could affect the outcome variable simultaneously. This concern is most relevant in settings with only a few events. It is somewhat lessened in our setting as we have a large sample of crosscountry events over time. This makes it impossible for a single omitted factor to drive our results. Instead, the assumption we need to identify the conditional mean impact of unrest is that other shocks are not systematically correlated with the timing of unrest. We conduct a placebo test in Section 5, checking if stock market returns follow a similar pattern in countries not affected by social unrest.

3.3 Regression analysis

We complement the event-study approach with regression analysis. The regression approach has several advantages over the event-study approach. First, it allows for a multivariate setting in which characteristics of the event can be controlled for. Second, it relaxes the assumption on a linear relationship in the market model (2). Third, it improves the efficiency in the estimation of the market model (in case of low number of observations in the estimation period).¹⁹ Finally, it addresses the well-known clustering problem in the event study approach, which occurs when event windows of different cross-sectional units overlap in calendar time. Not addressing the clustering problem leads to a cross-sectional correlation of abnormal returns and can result in over-rejecting the null hypothesis (Kolari and Pynnonen (2010)). We estimate the following multivariate regression:

where R_is are the daily stock index returns, S and S² control for time trend (number of days since day 0, following Arteaga-Garavito et al. (2020)), and γ_k and δ_s are country and weekday fixed effects respectively. We estimate regression (6) over a period of [-50, +50] business days around the event. The variable of interest is 𝟙[Unrest_is] - a dummy variable equal to one on days within the event window. We focus on two event windows: [0, +3] and [0, +7] days after the event. We apply Newey-West variance estimator to account for heteroskedasticity and autocorrelation in standard errors (up to lag 2). The coefficient of interest is β₁ it measures the effect of social unrest on the average 4- or 8-day returns after the event. Under the assumption that the timing of social unrest episodes is exogenous conditional on covariates, β₁ can be interpreted as the average causal effect of social unrest on stock market performance.

The event-study and regression analysis differ in three important regards. First, the event-study approach computes event-specific counterfactuals, estimating different loadings (the a_i and b_i) on the global portfolio for different events. In contrast, the estimated coefficients for the regression approach represent the average impact of unrest conditional on covariates (β₁), which is by construction identical for all events. Second, the counterfactuals computed across the two methods have different functional forms. In the event-study case, only returns on the global portfolio matter, whereas the regression setting allows for additional controls (country and weekday fixed effects, quadratic time trend). Finally, the data used to inform the counterfactuals differs across the methods. In the event-study case, only pre-event data is used to estimate counterfactual returns, whereas in the regression setting post-event returns also matter. The advantage of the regression approach is that it uses more data. The disadvantage is that if the unrest has an effect on daily returns at horizons longer than the event window, the post-event data will be wrongly labeled as untreated and will bias the coefficient estimates towards zero.²⁰ One final advantage of the regression methodology is that it allows us to look at the effect of social unrest on variables other than stock returns for which there is no obvious benchmark – trading volume, bond yield and exchange rate. We discuss some of these additional results in Section 6.

3.4 Control variable regression

High-frequency data is important for our purpose – it guarantees that our results are not driven by unobserved simultaneous factors. However, most, if not all, country-specific indicators are unavailable in daily frequency. Therefore, they cannot be controlled for in the regression model (6). We use two approaches to account for heterogeneity across countries. First, we split the sample by country characteristics and include the interactions of country group dummies into the regression. Second, we use a control variable regression model in which cumulative abnormal returns are regressed on a vector of control variables of interest. Following Forbes (2004), we re-estimate our market model over the period of [-50, +50] days around the event:

where 𝟙[Unresti_S] is a dummy equal to 1 on event-window days (for s ∈ [0, +3]). We then use the estimated coefficient $${\hat{\omega }}_i$$ (i.e. 4-day average abnormal returns for event i) as the dependent variable in equation (8):

Note the absence of time subscript s in equation (8) – the regression is no longer time-dependent, and the vector of control variables F_i can include regressors that don’t vary daily. An advantage of this control variable approach over interacting a sample split dummy is that we can consider multiple variables at the same time and examine whether they are simultaneously important.²¹

4.1 Cumulative abnormal returns

Figure 5 presents our main results from the event-study approach.

View Full Size

Cumulative abnormal returns based on market model from equation (2). Market returns based on MSCI ACWI. Blue dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

We find that social unrest leads to a significant reduction in stock market returns. Cumulative abnormal returns drop by around 0.72 percentage points three days after the unrest event relative to day -1. The returns continue to decline afterwards, reaching -1.4 percentage points two weeks after the event. The effect is not only statistically significant over the whole event window, but also economically meaningful: a 1.4 percentage point decline corresponds to the ninety-fifth percentile of the two-week cumulative returns distribution during the pre-event sample. In other words, the unconditional probability of observing the change of such magnitude in two-weeks cumulative returns before the event is only 5 percent.²²

The estimated effect of social unrest is of a similar magnitude to that in the literature. Epstein and Schnietz (2002), for instance, find that Fortune 500 companies experienced a 1.86 percent decline in their cumulative abnormal returns after WTO protests in Seattle in 1999. The magnitude of a decline in cumulative abnormal returns that we observe following a social unrest event is larger than the decline in the Dow Jones Industrial Average index following the terrorist events such as World Trade Center Bombing in 1993 (0.59%) or Embassy Bombing in Kenya in 1998 (0.59%, see Chen and Siems (2004)).²³ Several additional results in Figure 5 are worth noting. First, the negative impact of social unrest on stock return is not short-lived. Cumulative abnormal returns remain significantly negative at least 25 business days after the unrest event (i.e. the end of our event window). We also observe no reversion to zero over a longer time horizon, as shown in Appendix Figure A4. The long duration of the impact implies that social unrest may have a level effect on stock market performance, leading to permanently lower valuation. Second, the results presented in Figure 5 are robust to the choice of market portfolio. They are qualitatively and quantitatively similar if we use an alternative global stock index in the market model (e.g. S&P Global, see Appendix Figure A5). Last but not least, we see no trend in the period leading up to the event. Figure 5 shows that cumulative abnormal returns on days before the event are not significantly different from zero. This result implies that events are not priced in by market participants in the pre-event period. It further corroborates our identification assumption on the unpredictability of the event timing. Overall, given that abnormal returns before the event are not different from zero, and that the probability of observing such a drop in buy-and-hold returns in the pre-event sample is very low, we can reasonably interpret the observed decline as a causal effect of social unrest on stock market performance.

4.2 Regression results

Table 4 reports the main results of the regression model (6). The two columns correspond to two different regressions with 4- and 8-day event windows respectively.

^{Table 4:}

Estimating equation (6); dependent variable – stock market index returns

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

^{Table 4:}

Estimating equation (6); dependent variable – stock market index returns

	[0,+3]	[0,+7]
Dummy for social unrest	-0.00165** (0.001)	-0.00109* (0.001)
Country FE	Yes	Yes
Weekday FE	Yes	Yes
Observations	15075	15075
F-stat	1.969***	1.949***

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

View Table

^{Table 4:}

Estimating equation (6); dependent variable – stock market index returns

	[0,+3]	[0,+7]
Dummy for social unrest	-0.00165** (0.001)	-0.00109* (0.001)
Country FE	Yes	Yes
Weekday FE	Yes	Yes
Observations	15075	15075
F-stat	1.969***	1.949***

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

We find that social unrest reduces stock market returns by 16.5 basis points on average in the first four days after the event. This effect attenuates somewhat over the longer event window, reaching 10.9 basis points over 8 business days after the event. We report the Newey-West (heteroskedasticity- and autocorrelation adjusted) standard errors with lag 2 in the table parenthesis.²⁴ Our error assumptions are thus less restrictive compared to the Brown and Warner (1985) t-statistics reported in the event-study approach. The fact that our estimates are statistically significant at the 5% level in both approaches gives us confidence that we do not falsely reject the null hypothesis of zero cumulative abnormal returns.

We also use regression analysis to check the existence of anticipation effects. To do this we change the definition of event window by including day -1 into the interval. Appendix Table A3 shows that although the coefficient β₁ over both event windows is still statistically significant and negative, it is much smaller in absolute value. In other words, including day -1 before the event reduces the magnitude of a drop in returns during the event window, implying little anticipation effect the day before the event. This result is also consistent with the absence of pre-event trend in cumulative abnormal returns in Figure 5.

To provide further comparison between results from the two approaches, we estimate equation (6) for 25 different event windows, and plot adjusted coefficients β₁ against cumulative abnormal returns from Figure 5. We adjust the coefficients by multiplying them with the number of days in the event window (thus, coefficient from the first column of Table 4 is multiplied by 4, the coefficient from the second column – by 8 and so on). This adjustment corrects for the difference in magnitude between cumulative (event-study) and average (regression) returns. The results of comparison are summarized in Figure 6.

View Full Size

Cumulative abnormal returns. Based on market model from equation (2), market returns based on MSCI ACWI. Blue dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics. Orange line connects beta coefficients from equation (6) multiplied by the number of days in the event window.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

This comparison shows that our results are not sensitive to the choice of empirical approach. Regardless of whether we use event-study or regression, the estimates are quantitatively and qualitatively unchanged. We will come back to the robustness of our results in Section 5.

4.3 Heterogeneities: duration and income

We have established that social unrest leads, on average, to a sharp and persistent reduction in stock market returns in the country where unrest occurs. We next investigate potential heterogeneities in this effect across events and countries. We first ask whether long- and short-duration events differ. We plot the results on Figure 7.

View Full Size

Cumulative abnormal returns by duration groups. Based on market model from equation (2), market returns based on MSCI ACWI. Dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

To investigate the role of event duration, we classify events as short-term (last for less than 3 days), medium-term (last between 3 and 7 days), or long-term (last for more than 7 days).²⁵ The majority (54%) of the events in our sample are short-term. Long-term events constitute 16% of the sample (25 out of 156 events, with average duration of 16.5 days). They include, among others, Arab Spring protests in Tunisia, Egypt and Iran in 2011, Ukrainian civil unrest in 2013 (“Euromaidan”), Greek anti-austerity protests of 2012 and French “Yellow vests” movement of 2018. Our list of long-term events seems reasonable when compared to external sources, as it includes many instances of social instability that are commonly known to have lasted for days and even weeks. Appendix Table A2 shows the duration of each event in our sample.

We find that the duration of the event matters: the longer the event, the larger are the stock market reactions. Figure 7 clearly indicates that the average effect we found in our baseline results is driven by long-term events and, to a lesser extent, by medium-term events (see Appendix Table A4 for the results of regression estimation). The impact of long-term events is statistically significant and substantial in magnitude: cumulative abnormal returns decrease by as much as 8 percentage points in countries that experience persistent social instability²⁶ The magnitude of the response is closer to the 14.5 percentage point drop in cumulative abnormal returns that Acemoglu et al. (2017) report for 2011 Arab Spring protests in Egypt.

Next, we investigate the role of economic development. We split our sample of events into three groups based on country income level (advanced economies, low-income developing countries – LIDC, and other emerging and developing economies – EMDE) and repeat the procedure. Figure 8 and Table A5 show the results. We find that the decrease in the cumulative abnormal returns of stock market indices is driven by events in EMDE and LIDC.²⁷ Stock markets in advanced economies on average are not subject to the negative consequences of social unrest – the estimate for advanced economies is never significantly different from zero. This result is in line with the anecdotal evidence that US stock market reacts very mildly to protests on police brutality (Sraders (2020)). Of course, country characteristics such as income may be the result of other factors, such as a country’s political and social institutions. We turn to this issue in the next section.

View Full Size

Cumulative abnormal returns by country income groups. Based on market model from equation (2), market returns based on MSCI ACWI. Dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

4.4 The role of institutions

The literature on the effects of social and political instability highlights the importance of institutions in mitigating its negative consequences. Eldor and Melnick (2004), for instance, find that market liberalization reforms contribute to coping with terrorist attacks and help markets function efficiently. According to Bernal-Verdugo et al. (2013), the recovery of output after social instability incidents depends on whether the country implements reforms that improve the level of governance. In this section, we examine whether institutions matter for stock market reactions to social unrest.

We first look at the role of democratic institutions. We split our sample of events into two groups based on the Polity Score developed by the Center for Systemic Peace.²⁸ The Polity Score is a widely used measure of the degree of regime authority. It ranges from -10 (hereditary monarchy) to +10 (consolidated democracy). We use the median Polity Score in our sample (score of 6) as a threshold.²⁹ We use the Polity5 dataset which covers the period of 1800–2018, and, assuming that Polity Score doesn’t vary much over time, impute 2018 Polity Score dummy values for the events that happened in 2019 and 2020. The results are presented in Figure 9.

View Full Size

Cumulative abnormal returns by Polity Score. Based on market model from equation (2), market returns based on MSCI ACWI. Dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

Figure 9 has a clear message: the negative effect of social unrest on stock index returns is pronounced only in countries with below median Polity Score. The impact of social unrest is more negative in less democratic countries: cumulative abnormal returns in these countries drop by more than 3 percentage points over the 14-day event window. The impact on the economies with above median Polity Score seems to be negligible: the cumulative abnormal returns are never significantly different from zero. We further look at additional indicators of institutional quality – World Governance Indicators.³⁰ WGI Project compiles data on traditions and institutions from various sources into six governance indicators: Rule of Law, Control of Corruption, Government Effectiveness, Regulatory Quality, Political Stability and Voice and Accountability. We use the percentile ranking of countries for each of these indicators to split our sample into two groups based on the median. Since WGI indicators are only available for 2011–2019, we impute 2019 values for the events in 2020. The results are summarized in Figure 10.

View Full Size

Cumulative abnormal returns by groups based on World Governance Indicators, 2011–2020. CARs based on market model from equation (2), market returns based on MSCI ACWI. Dashed lines – 95% confidence intervals, computed using Brown and Warner (1985) t-statistics.

Citation: IMF Working Papers 2021, 079; 10.5089/9781513572765.001.A001

• Download Figure
• Download figure as PowerPoint slide

Governance indicators along two dimensions – Regulatory Quality and Voice and Accountability – have a statistically significant impact on stock market performance after social unrest. In both cases social unrest events affect cumulative abnormal returns only in countries with weak (below median) governance. Regulatory Quality indicator captures governance features which promote or prevent the development of the private sector (e.g. price controls, investment freedom, discriminatory taxes etc.). The measure of Voice and Accountability is similar in interpretation to Polity Score: this indicator aggregates information on democracy, freedom of press and political rights. The results in Figures 9 and 10 point to the importance of governance institutions. Democracy, civil freedoms and government accountability, as well as quality of private sector regulation play a crucial part in shaping the consequences of social unrest events. The aftermath of social unrest is more severe in countries with weak institutions: the stock markets in these economies seem to be the most vulnerable to social instability.

The heterogeneity analysis conducted in previous sections suggests that the influence of social unrest on stock market behavior across the world depends on several event-and country-characteristics: duration of social unrest, country income group and quality of institutions. One potential concern is that some of these parameters are correlated. For instance, it is possible that the strong negative effect of unrest in LIDC and EMDEs is driven by the fact, that social unrest in these countries tends to be more long-term. To check whether the three factors are simultaneously and independently important, we use the control variable approach summarized in Section 3.4. The results of estimating equation (8) are presented in Table 5.

^{Table 5:}

Estimating equation (8) – control variable regression; dependent variable – average 4-day abnormal returns.

Standard errors in parentheses ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

^{Table 5:}

Estimating equation (8) – control variable regression; dependent variable – average 4-day abnormal returns.

	(1)	(2)	(3)	(4)
Medium event	-0.00178 (0.002)		-0.00185 (0.001)	-0.00149 (0.001)
Long event	-0.00736** (0.004)		-0.00807** (0.004)	-0.00778* (0.004)
LIDC		-0.00283 (0.002)	-0.00442* (0.002)	-0.00276 (0.002)
EMDE		-0.00335*** (0.001)	-0.00361*** (0.001)	-0.00258* (0.001)
Polity score above median				0.00257* (0.001)
N	156	156	156	155

Standard errors in parentheses ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

View Table

^{Table 5:}

Estimating equation (8) – control variable regression; dependent variable – average 4-day abnormal returns.

	(1)	(2)	(3)	(4)
Medium event	-0.00178 (0.002)		-0.00185 (0.001)	-0.00149 (0.001)
Long event	-0.00736** (0.004)		-0.00807** (0.004)	-0.00778* (0.004)
LIDC		-0.00283 (0.002)	-0.00442* (0.002)	-0.00276 (0.002)
EMDE		-0.00335*** (0.001)	-0.00361*** (0.001)	-0.00258* (0.001)
Polity score above median				0.00257* (0.001)
N	156	156	156	155

Standard errors in parentheses ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

The first two columns of Table 5 show the results that consider only event duration (column 1) or only country income group (column 2). Column 3 shows results that consider both country income (with advanced economies as reference category) and duration (with short-term events as reference category) groups. Although the two variables are negatively correlated (since the coefficients of income decrease), both income and duration coefficients are statistically significant, at least for long events in EMDEs. Column 4 further includes a dummy variable equal to 1 if country is above median Polity Score. Despite a small number of observations, the coefficients of duration, income and Polity Score are statistically significant at the 10% level. As before, long duration of social unrest implies starker decrease in average abnormal returns, while democratic institutions partly mitigate the effect. The coefficient of EMDE dummy, however, is smaller in magnitude. This result seems to suggest that country income group (at least in case of LIDC) serves as a proxy for democracy in regressions with omitted Polity Score dummy.

4.5 The role of elections

According to Table 5, the quality of democratic institutions – as measured by Polity Score – is an important predictor of the stock market reaction to social unrest even after controlling for income and duration. It is possible, however, that the democracy indicator reflects some underlying country-specific features, not captured in our control variable regression. For instance, one could argue that the countries with above median Polity Score also exhibit high regulatory quality and effective governance (our analysis of World Governance Indicators in Figure 10 provides some indirect evidence of this possibility). If this is the case, the democracy indicator in Table 5 serves as a proxy for unobserved quality of governance.

In order to disentangle the two channels, we use data on lower chamber legislative elections at a constituency level from Constituency-Level Elections Archive (CLEA).³¹ CLEA records the dates and the outcomes of elections from as early as 1788. We generate a dummy variable equal to one if the elections occur within [-50, +50] business days around the social unrest event. Thus, we can distinguish between election-related social unrest events and other types of social unrest. By the nature of collected information, CLEA provides election data for a selected pool of countries that have regular elections. Thus, we conduct our heterogeneity analysis on a reduced set of countries.³² Out of 140 social unrest events in our reduced set, only 21 happen simultaneously with lower chamber elections, and can be classified as election-related. We then estimate equation (6), including interaction term SU ×Elections. Since our regression equation controls for country fixed effects, the coefficient of the interaction term captures the difference between election-related unrest and all the other unrest events within countries. The results are presented in Table 6.

^{Table 6:}

Estimating equation (6); dependent variable – stock market index returns. Includes interaction terms between Social Unrest dummy and Election dummy.

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

^{Table 6:}

Estimating equation (6); dependent variable – stock market index returns. Includes interaction terms between Social Unrest dummy and Election dummy.

	[0,+3]	[0,+7]
Dummy for social unrest	-0.00191** (0.001)	-0.00166** (0.001)
SU*Elections	0.00295** (0.001)	0.00214* (0.001)
Country FE	Yes	Yes
Weekday FE	Yes	Yes
Observations	13529	13529

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

View Table

^{Table 6:}

Estimating equation (6); dependent variable – stock market index returns. Includes interaction terms between Social Unrest dummy and Election dummy.

	[0,+3]	[0,+7]
Dummy for social unrest	-0.00191** (0.001)	-0.00166** (0.001)
SU*Elections	0.00295** (0.001)	0.00214* (0.001)
Country FE	Yes	Yes
Weekday FE	Yes	Yes
Observations	13529	13529

Newey-West standard errors with 2 lags in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.01

The coefficient of the interaction term SU × Elections is positive and statistically significant. Although on average social unrest in democratic countries leads to a 19 b.p. decline in stock market returns, social unrest related to election outcomes actually improves stock market performance by 10 b.p. Thus, politically-conditioned social unrest leads to an increase in stock market returns. This finding has several important implications. First, it confirms the previously documented absence of stock market reaction to unrest in countries with above median Polity Score (Figure 9) and above median Voice and Accountability indicator (Figure 10). Second, it shows 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). We will return to a more comprehensive discussion of the mechanisms behind these results in Section 6.