Forecasting the risk of terrorist attacks based on machine learning algorithms

Novikov Andrey Vadimovich Assistant, Plekhanov Russian University of Economics 117997, Russia, Moskva, g. Moscow, per. Stremyannyi, 36, kab. 339 Camouflage@yandex.ru Other publications by this author

Introduction

Terrorist attacks, as one of the most urgent types of conflicts, usually occur unexpectedly, lead to human casualties and eventually sow chaos ^[18]. According to the Global Terrorism Database (GTD), the number of terrorist attacks in the world has grown from an average of 1,000 cases per year at the beginning of the XXI century to more than 7,000 cases in 2020 ^[25]. Although the Governments of many countries have spent huge sums on the fight against terrorism during this period, the results have not been as impressive as they could have been. Counter-terrorism operations, such as pre-empting and suppressing terrorist attacks, are widely used, but sometimes they can further fuel attacks rather than prevent them ^[12]. Forecasting the risk of terrorist acts is useful for developing preventive measures, as well as for ensuring more targeted long-term policy formation. However, the factors that affect the risk of terrorist attacks are numerous and complexly interrelated. This ambiguity complicates forecasting and can lead to rash political decisions that endlessly delay resources and create panic among the population. Therefore, it is very important to identify the most important indicators affecting the risk of terrorist acts, on the basis of which forecasting will become clearer and more reliable, and decisions on countering terrorism will become more active ^[9].

In recent years, thanks to the development of machine learning technologies and their ability to effectively identify multifactorial relationships, relevant scientific papers are gradually overcoming the seemingly "unpredictable" problem of predicting terrorist attacks ^[5]. These works can be divided into two categories. In the first category, indicators of the level of individual terrorist incidents are mainly used to predict the future nature of terrorism. In the work of N. Mo, 56 initial attributes of a terrorist attack were selected from the GTD database, including time, location, type of attack, etc., and the types of terrorism were classified using the support vector machine (SVM), naive Bayesian classifier (NB) and logit regression (LR) ^[31]. K. Minu's research uses wavelet neural networks and GARCH models to predict the future time of terrorist attacks based on historical data on the number of attacks per month ^[30]. H. M. Ismail and H. Kazi applied 16 attributes of the level of a specific incident (such as year, type of attack, number of perpetrators, etc.) to assess the nature of terrorist attacks based on an ensemble classifier combining Bayesian models and decision tree (DT) models ^[26]. In the work of R. T. Brandt, the Markov switching Bayesian model was used to predict the intensity of future conflicts based on previous terrorist incidents ^[13]. Also, a number of Russian authors used Bayesian models and random forests to model the probability of terrorism ^{[3, 5]}.

However, in the works of the first category, only micro-level indicators related to specific attacks are considered, and the influence of the "root factors" of terrorism, which are otherwise associated with macro-conditions or the external environment, is ignored ^[6]. The attacks take place in the broader context of terrorism and these complex conditions encompass political, economic, religious and other factors. An extensive group of works focused on the root causes of terrorism, starting with M. Crenshaw as the most representative early work ^[19]. In this literature, the methodologies used include qualitative analysis based on theoretical propositions and quantitative empirical studies using statistical tools. The main conclusions concerning the external factors of terrorism include: 1) poverty cannot directly lead to terrorism ^[10]; 2) democracy reduces the direct costs associated with terrorist attacks, but also increases the relative costs ^[22]; 3) urbanization is a breeding ground for terrorism ^[19]; 4) countries with growing populations seem to suffer less terrorism ^[21]; 5) terrorism is closely connected with other types of political violence, conflicts and wars ^[14]. In general, a consensus can be reached that the origin and spread of terrorism are caused by broad and far-reaching political, economic, ethnic and other problems ^[2]. However, these factors are insufficient to explain specific cases of attacks ^[4]. As part of the forecasting efforts in time series, it is necessary to take into account both the factors of incidents as internal causes of terrorist attacks and external conditions.

Based on this, studies of the second category on the prediction of attacks take into account macro-level factors when making forecasts. S. Perry considers 30 economic indicators that measure factors such as unemployment, income and predicts mortality from terrorism using neural networks of back propagation (BP) ^[33]. N. V. Weidman and M. D. Ward also use four indicators of external conditions containing population size, ethnic composition and landscape to predict conflicts at the municipal level based on a spatial-temporal logistic model ^[37]. In the work of M. Hao uses a random estimate of the density of forests and cores to predict the potential risk of terrorist attacks on the Indochina peninsula ^[24]. Mainly in this work with the help of 15 indicators of external conditions containing three social indicators (fragility of the state, population density, the spread of drug trafficking) and eleven geographical indicators, such as average temperature and topography. In contrast to this approach, which considers only indicators of root causes, F. Ding considers both characteristics of incidents and macro-level indicators ^[20]. It uses a reverse propagation neural network (BPNN), SVM and random forest (RF) to predict the risk of terrorist attacks in various countries, mainly analyzing three incident indicators (such as latitude and longitude) and 10 macro-causes containing five geographical indicators (for example, average rainfall), as well as four demographic indicator. In the second category of studies on the prediction of terrorist attacks, the status of environmental factors was assessed.

Nevertheless, existing research in the field of risk assessment is mainly aimed at improving the accuracy of forecasting based on pre-selected indicators representing various factors related to terrorist acts, and less often return to the validity of the choice of indicators ^[7]. This is partly due to the fact that the "black box" of machine learning models with numerous parameters and complex conclusions makes it difficult to clearly analyze the causal relationships between target dependent and independent variables ^[34]. Thus, the internal mechanism of the learning model is not clear enough and the relationship between the indicators cannot be understood intuitively. At the same time, after changing the input indicators, the initial conclusions cease to be reliable ^[15].

Methodology

The general structure of the proposed approach consists of four stages. Forecasting the risk of terrorist attacks in different countries for a particular year is carried out using a sliding window with several inputs and outputs. The risk of terrorist attacks as a predicted target consists of four main sub - components: 1) Will there be more than one major terrorist attack next year? 2) What is the maximum level of property damage caused by terrorist acts next year? 3) What is the maximum level of human casualties as a result of terrorist attacks in the coming year? 4) What is the average success rate of terrorist attacks next year? Since it is impractical to consider each terrorist attack equally important and pay equal attention to them if no distinction is made between the severity of the attacks ^[18]. In general, this study is more concerned with the risk of attacks with severe consequences.

As for the input data (independent variables) for forecasting, both external and internal factors are taken into account. External factors play the role of an invisible hand and represent various aspects of the state of social anomie, directly leading to the spread of terrorism and indirectly to terrorist attacks ^[11]. Internal factors are both triggers and consequences of terrorist attacks, fueling the terrorist phenomenon through a self-replicating feedback loop ^[27]. In total, 28 indicators were selected as predictors, of which 17 are macro–level indicators, and the remaining 11 are related to the internal characteristics of attacks.

The initial data used in this study are collected from three well-known databases, such as the Global Terrorism Database (GTD), the International Country Risk Guide (ICRG) and the World Bank database ^{[35, 36, 38]}. GTD is one of the largest databases on terrorist attacks with open access, in which the target, type of attack, location and other indicators of terrorist attacks have been recorded since 1970. The ICRG registers the indicators of each country for each year in three subcategories of risks: political, financial and economic. The World Bank database contains reliable statistical socio-economic data at the country level.

Selection of indicators and data preprocessing

Terrorist attacks stem from the abstract context of terrorism, in which various factors contribute to the process of radicalization, and indicators are a quantitative interpretation of factors. As shown in Figure 1, the process of transition from radicalism to terrorist attacks consists of several stages ^[12]. First, terrorism is spreading because of deep-rooted structural causes. Secondly, some citizens become terrorists under the influence of external circumstances and subjective individual causal factors. Finally, due to specific triggers or reproducible feedback from previous acts, terrorists commit more and more attacks ^[8].

Fig. 1. The process of transition to terrorism under the influence of external factors.

Source: compiled by the author.

Structural, accelerating and motivational factors are considered as three subcategories of the causes of terrorism based on T. Bjorgo's research ^[12]. Structural causes are factors affecting people's lives at the macro level that people may or may not be aware of. Accelerating factors make terrorism "attractive" and encourage people to become more prone to political violence or joining terrorist organizations. Motivational factors are people's personal experiences that encourage them to use terrorist tactics, including joining radical organizations or preparing for terrorist attacks. Factors in these three subcategories primarily contribute to the spread of terrorism and the recruitment of militants. Moreover, motivational reasons can also potentially lead to attacks, but only at the preparatory stage. Provoking and feedback factors are considered as two components of causes at the level of specific incidents. Provoking factors are direct precursors of specific attacks. For example, specific conflicts, controversial events, etc. Feedback factors basically represent the entire previous cycle of terrorist attacks. After the commission of a terrorist act, many audiences participate in the subsequent process, including allies, enemies, members of a terrorist organization, innocent civilians. The consequences, intentions and details of the attacks can be disclosed to the media, which can lead to an escalation effect, further increasing the risk of the next attack.

This work focuses on identifying the most important indicators that affect the risk of terrorist attacks from the point of view of forecasting. The factors of the level of root causes and the level of incidents of terrorist acts are comprehensively considered. Internal factors play the role of an invisible hand and represent various aspects of the conditions of social disorder. The factors associated with terrorist attacks are both consequences and triggers of terrorist attacks. To quantify these abstract factors, a total of 28 indicators were selected as the initial set of input predictors of the proposed RF forecasting model. Based on this, the RF-RFE method is proposed to determine the most important indicators by recursively reducing the number of variables. The results of the study show that the minimum set of input indicators before the effectiveness of forecasting deteriorates significantly includes: the number of victims of attacks in the previous year, GDP growth, military spending, population growth rates, population size, unemployment, urban population growth, internal conflicts. The identified factors should be considered important and necessary for understanding the risk of terrorist attacks. The most important indicators identified indicate that the terrorist risk is due to both the root causes of terrorism and previous incidents of political violence.

It is known that efforts to reduce the risk of terrorist attacks can be undertaken in two directions. Regarding the first aspect, it is necessary to constantly eliminate the root causes of terrorism by increasing the stability and legitimacy of the Government, reducing intra-State conflicts and contradictions, increasing freedom of entrepreneurship, increasing the dynamics of economic growth and reducing unemployment. As for the second aspect, measures are needed to prevent the growth of terrorism by minimizing the positive impact of previous terrorist attacks on possible future attacks. Key actions include reducing the likelihood of success of terrorist attacks, reducing the number of human casualties and the material and economic losses they cause. In addition, it is necessary to minimize the dissemination of negative information, such as statements about the intentions of terrorist organizations. The main limitation of this article is that due to the complexity of data collection and the relative sensitivity of the sample, only 28 indicators were considered. Future studies should analyze additional factors of terrorism development and focus on the primary impact of political risks, such as government stability.