• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site
vision

Machine Learning has Helped Forecast Global Hotspots of Unrest and Revolution

From Lebanon to Georgia. HSE cliometrists named the countries with the highest risk of civil unrest in the next five years

Wikimedia Commons

First — in brief

The problem: Both armed and peaceful protests are not uncommon in history. Since the middle of the last century, science has been actively studying and trying to predict these phenomena of social instability. It is difficult to develop predictive models of past revolutions because of the need to account for an enormous number of indicators. but it is even more challenging to try to discern future threats.

The solution: Now, AI — and machine learning in particular — has come to the rescue with its ability to analyse large amounts of statistical information.

Now, in greater detail

HSE scientists Andrey Korotayev and Ilya Medvedev used machine learning (ML) to build an index of instability in the world. The new method made it possible to use a large number of variables and distribute them in non-standard fashion. This enabled the researchers to propose and successfully test their own method of building an index for the next five years, as well as to compile lists of countries where protests are most likely to occur. The results of the work — that was carried out with the support of the Russian Science Foundation — are presented in the yearbook Systemic Monitoring of Global and Regional Risks (vol. 12. 2021).

The main idea

There are three traditional approaches to constructing an instability index. The first and most common is purely mathematical. It predicts events using a small number of broad variables that characterise a country’s development. The second is the ‘groupings and assessments’ method. This index is compiled on the basis of several sub-indices measuring security, political, economic and social factors, with experts adjusting the results. The third is a combined approach, with the final model incorporating the results of the mathematical modeling and a block from experts.

According to Ilya Medvedev and Andrey Korotayev, however, none of these methods reflects the interaction of different categories of indicators, takes all country-related factors into account, or combines data without losing some information.

To remedy this, the authors of this study developed their own approach. AI algorithms enabled them to take a large number of variables into account and use them simultaneously in the mathematical model — that is, without dividing them into sub-indices. They abandoned the use of expert assessments due to their variability and lack of universality among different countries.

How was the study performed?

The indices were constructed using the supervised learning method — that is, the algorithm learned and made a forecast based on the input data and a ready-made answer (the occurrence of revolutionary events).

The calculations employed the Yandex Catboost library, a machine learning model with a gradient boosting algorithm. A gradient is a vector that sets the direction of the fastest change in value. Together with the boosting function, the technology creates a predictive model in the form of a set of decision trees that make it possible to reach an answer. Here, the ‘boosting’ function means that several less well-trained algorithms are transformed into one strong one: each subsequent one taking into account the errors of the previous one and improving on its ability to make forecasts.

Researchers looked at unarmed and armed protests separately. Accordingly. there were two predictive models and, as a result, two indexes of revolutionary events. The first is for the world as a whole and the second is specifically for the Afro-Asian macrozone, one of the world’s most volatile regions. The model did not forecast revolutionary activity for a specific year, but the probably of it occurring over a five-year period.

The calculations were based on more than 240 indicators (economic, political, demographic and social) taken from several databases. These include the CNTS (Cross National Time Series Database of 200 countries from 1815 through 2020), the IMF report Global Manufacturing Downturn, Rising Trade Barriers–2019, World Development Indicators (2020), and others.

What was the result?

In all, 25 countries were shown to be at the greatest risk of experiencing unarmed protests, with Lebanon, Honduras, Algeria, Bolivia, and Montenegro foremost among them. They were closely followed by Colombia, Afghanistan, the Dominican Republic, Indonesia and Uzbekistan.

Forecasted risk index for the outbreak of unarmed revolutionary uprisings

(countries facing the greatest risk over the next five years — from the greatest forecasted instability to the least)

Lebanon

24.36

Honduras

18.64

Algeria

9.55

Bolivia

5.35

Montenegro

3.64

Albania

2.09

Iraq

1.9

Chile

1.34

Israel

1.31

Haiti

1.22

Nepal

0.78

Egypt

0.61

Sudan

0.6

France

0.56

Mali

0.52

Guatemala

0.5

Morocco

0.5

Iran

0.46

India

0.46

Myanmar

0.43

Columbia

0.43

Afghanistan

0.43

Domincan Republic

0.35

Indonesia

0.35

Uzbekistan

0.34

Source: I. Medvedev and A. Korotayev, ‘Building a forecast index of revolutions: the experience of using machine learning methods’

In Lebanon, the researchers commented, instability has, indeed, been observed for many years, and in Afghanistan, an unarmed revolution might follow the end of the current armed revolution.

In 14th place we find France, a country from the so-called world-system centre — that is, states with highly developed economies and political weight. These countries might see peaceful demonstrations akin to the ‘yellow vests’ movement.

The countries facing the highest risk of unrest are those with an average per capita income and that are located on the semi-periphery of the world-system list — that is, those that occupy an intermediate position between the above-mentioned centre and the periphery. These countries supply raw materials to the leading states and are dependent on them.

The result was primarily influenced by social indicators (the level of oppression of certain segments of the population, the religious freedom index, etc.). Economic and political indicators (per capita GDP, longevity of the political regime, etc.) are less significant, and demographic indicators (such as population density and size) were in last place.

With regard to armed revolutions, this picture is reversed and social indicators are less significant than political factors that reflect the nature of the country’s regime.

Foremost among the 19 states with the highest likelihood of armed conflicts erupting are India, Angola, Chad, Niger and Burundi, with Libya, Bangladesh, Mali, Myanmar and Egypt at the bottom of the list.

Forecasted risk index for the outbreak of armed revolutionary uprisings

(countries facing the greatest risk over the next five years — from the greatest predicted instability to the least)

India

3.22

Angola

2.31

Chad

2.31

Niger

2.15

Burundi

2.07

South Sudan

1.96

Mozambique

1.95

Afghanistan

1.92

Syria

1.89

Pakistan

1.83

Yemen

1.82

Central African Republic

1.79

Somalia

1.78

Iraq

1.73

Egypt

1.68

Myanmar

1.64

Mali

1.58

Bangladesh

1.56

Libya

1.50

Source: I. Medvedev and A. Korotayev, ‘Building a forecast index of revolutions: the experience of using machine learning methods’

‘India has long shown visible signs of instability. Niger, Chad and Mali are also not surprising: they have long been failed states’, explained Andrey Korotayev and Ilya Medvedev.

Unlike the forecast for unarmed uprisings, it is not the semi-periphery that is mainly subject to the risks of violent uprisings, but the periphery, those states with ‘low per capita incomes, high birth rates, and very young, poorly educated and poorly urbanised populations’.

The same result — that is, the readiness of countries primarily on the ‘periphery’ to rise up with weapons — was also demonstrated by the Afro-Asian zone: the Middle East (including North Africa), the Greater Middle East (including Central Asia and Pakistan) and the Sahel. Again, researchers calculated the forecast for this region separately.

The most economically and politically unstable territories in this zone face the greatest potential threats. They include: Chad, Niger, South Sudan, Afghanistan, Syria, Pakistan, Yemen, the Central African Republic, Somalia and Iraq. Those facing a less acute threat have problems that are more social in nature. These include Georgia, the United Arab Emirates (UAE), Armenia, Israel and Oman.

‘Very wealthy countries of the macrozone that have monarchical regimes (such as Qatar or the UAE) are characterised by low risks of armed revolutionary destabilisation’, the study noted. By the way, this also applies to unarmed protests: states with a per capita GDP that is average for the region and partially democratic rule (such as Lebanon or Iraq) are most at risk of experiencing‘non-violent’ demonstrations.

How is this useful?

The proposed machine learning-based method made it possible to evaluate a very large body of data and to formulate not only the current instability index, but also to predict destabilising events over the next several years. The new approach helps ‘eliminate many of the problems faced by previous researchers’ and, as a result, make forecasting more accurate.
IQ
 

Authors of the study:

 

Author: Svetlana Saltanova, April 08