Доклады на 2-й конференции
по вопросам стратегической стабильности

Dr. Sean P. O’Brien
Analyzing Complex Threats for Operations and Readiness (ACTOR):
A Methodology for Forecasting Country Instability

Center for Army Analysis

This study draws upon the state strength literature to identify relevant country macro-structural factors (e.g., GDP per capita, life expectancy, ethnic diversity, global trade patterns) that can contribute to different kinds and levels of conflict and country instabilities. Our database includes annual observations across a range of macro-structural factors for every major country in the world over the period 1975-1999. We use a pattern classification algorithm--Fuzzy Analysis of Statistical Evidence (FASE)-- developed by the Center for Army Analysis to analyze the relationships between country macro-structural factors and historical instances of country instability. A split-sample validation design is used to evaluate the ability of FASE to generate competent predictions, using the standard performance metrics overall accuracy, recall, and precision. The results demonstrate the potential for FASE to accurately forecast not just the occurrence, but also the level of intensity of country-specific instabilities five years in advance with about 80% overall accuracy. The forecasts generated through the year 2015 suggest that South Asia and East Africa will continue to harbor highly unstable states. However, most of the states expected to improve their prospects for greater stability are also located in these regions.

KEYWORDS: conflict and instability analysis, forecasting conflict, instability, political risk assessments, forecasts of future conflict environments


1.1 Introduction
1.2 Purpose of ACTOR Study
3.1 Independent variables
3.2 Dependent Variable
3.3 Including Internal and External Conflicts
3.4 A Preliminary Analysis


1.1 Introduction

Since the end of the Cold War, economic dislocations, civil war, famine, and ancient ethnic and religious animosities have contributed to conflict and political instability in states extending from Haiti to the vast archipelago of Indonesia. These conflicts and instabilities frequently challenge national security interests; at other times, the human rights atrocities that often accompany these dislocations offend the moral imperatives of individual states as well as the international community.

Increasingly, Western powers, acting alone or in concert with international organizations, have responded to these post-Cold War crises in myriad ways, including peacekeeping operations in the Balkans, Sierra Leone, and East Timor; enforcing sanctions in the Persian Gulf and no-fly zones over Iraq; conducting humanitarian operations and evacuating civilian non-combatants in Africa; and conducting maritime interdictions in the Caribbean. The uncertainty surrounding where and when these crises might erupt around the globe has frustrated crisis response planners and humanitarian relief officials who often must plan, prepare, and budget for these contingencies months, even years, in advance.

To deal with this uncertainty in an environment where humanitarian disasters have increasingly come to compete for scarce resources with threats to more vital national interests, military planners, intelligence analysts and policy makers need tools and models to help them anticipate when and where these crises are likely to emerge. Indeed, as Gurr and Moore (1997, 1080) have noted, “Those who make foreign and international policy seek more than explanation: they want better ‘early warnings’ of impending conflicts so that preventative diplomacy and other conflict management tools can be brought into play.”

In an apparent effort to appeal to this need, researchers have turned their attention to the concept of “crisis early warning” (Bueno de Mesquita 1996, 1998; Rupesinghe and Kuroda 1992; Esty et al., 1995, 1998; Bond et al., 1997; Gurr and Moore 1997; Schrodt and Gerner 1997; Schrodt 1997, 1998; Harff 1998; Davies et al., 1998; Scarborough 1998; Brecke 1998; Pevehouse and Goldstein 1999; Beck et al., 2000; King and Zeng 2001; Jenkins and Bond 2001). This focus on early warning research has coincided with increasing concern among some scholars for conducting policy-relevant research (see, for instance, Post and Ezekial 1988; Lepgold 1997; Richardson 1997).

Though each of these studies represents a unique contribution to the development of early warning insights and capabilities, it is the research conducted by the State Failure Task Force (hereafter, the SFTF), a commission of prominent scholars and contractors set up by former Vice President Al Gore’s office in 1994, that has received the most recent attention in academic and policy circles.

Stood up in the wake of the genocide in Rwanda, the U.S. government tasked and provided substantial funding to the SFTF to identify and examine key factors associated with serious state crises and to develop a methodology that could identify “critical thresholds” in these factors so that they might provide early warning of state failures up to 2 years in advance. State failures include four types of events: (1) genocides and politicides (“sustained policies by states or their agents and, in civil wars, by contending authorities that result in the deaths of a substantial portion of members of communal or political groups”); (2) ethnic wars (“secessionist civil wars, rebellions, protracted communal warfare, and sustained episodes of mass protest by politically organized communal groups”); (3) revolutionary wars (“sustained military conflicts between insurgents and central governments, aimed at displacing the regime”); and (4) adverse or disruptive regime transitions (“major, abrupt shifts in patterns of governance, including state collapse, periods of severe instability, and shifts toward authoritarian rule”).

The SFTF used logistic regression, neural networks, and genetic algorithms to identify patterns in the relationships between hundreds of explanatory factors and different types of state failures. The best global model included only three independent variables—measurements of a country’s level of democracy, trade openness, and infant mortality rate. Using only these factors in a logistic regression, the SFTF was able to discriminate historical state failures from stable countries about two-thirds of the time. King and Zeng (2001), among others, have criticized the research produced by the Task Force on a variety of methodological grounds. Using neural network models, a corrected version of the State Failure Project’s baseline model, and two additional variables (legislative effectiveness and fraction of population in the military), these authors offer a state failure forecasting model that out-performs those published by the SFTF to date.

1.2 Purpose of ACTOR Study

The Analyzing Complex Threats for Operations and Readiness (ACTOR) Study was commissioned by the Office of the Deputy Chief of Staff for Operations and Plans (ODCSOPS), War Plans Division, Headquarters, Department of the Army. To provide an assist to policymakers and strategic planners, ACTOR was explicitly designed to extend this line of work in several ways.

First, in this study we are interested in forecasting the likelihood of country instability or, more precisely, the conditions conducive to instability, for every major country of the world over each of the next 15 years. To do so, we identify, evaluate, and ultimately forecast those macro-structural factors at the nation-state level that, when combined with events or triggers such as assassinations, riots, or natural disasters, have historically (over the period 1975-1999) been associated with different kinds and levels of intensity of conflict. We forecast the conditions conducive to conflict (i.e., country instability) over the long term and not the occurrence of conflict per se, because we possess no special insight into the specific events that might trigger particular conflicts in the future. Nevertheless, we submit that, regardless of our ability to predict the spark that will ultimately set them ablaze, we can forecast the oiliness of the oily rags with a reasonable degree of expected accuracy.

We use a data set of state conflicts, from the KOSIMO database, to validate macro-structural factors as relevant contributors to country instability. The KOSIMO Data Project includes violent, internal (intra-state) and external (inter-state) wars and crises, including most of the genocides/politicides, ethnic wars and revolutionary wars examined by the SFTF. It also includes less violent (mostly non-violent), but still intense, forms of crises that may be characterized as “war in sight” crises (Pfetsch and Rohloff 2000). This would include, for instance, the crisis between Russia and Ukraine over the possession and ownership of Soviet-era strategic weapons in the near-term aftermath of the demise of the Soviet Union. Violent state crises rarely occur spontaneously, but rather nearly always evolve more gradually from an initially less violent series of events. Though these non-violent crises may not, and typically do not, escalate into shooting wars, often it is at least partly because they provide sufficient warning and opportunity to take diffusive actions.

Second, we are interested in forecasting not only the likelihood that an instability in some binary sense will occur in any given country in any given year, but also that the instability will occur within a certain range or level of intensity (e.g., low, moderate, or high levels of violence). Toward that end, including violent and non-violent conflict events in the data set allows us to construct an approximate index of country instability for use in validation analyses. The index of instability places stringent demands on the algorithm used for classification and forecasting. The increased fidelity, though, is more interesting than the binary alternative and may perhaps be more useful to policymakers as a means for conducting threat assessments and allocating scarce resources to prevent or deter conflict.

We use a pattern classification algorithm developed by Chen (2000) called Fuzzy Analysis of Statistical Evidence (FASE) to examine the patterns in the relationship between different configurations of country macro-structural variables and the likelihood that states will experience a given intensity level of instability. FASE is a hybrid method that incorporates theoretical elements from statistics, possibility theory, and fuzzy logic, a branch of fuzzy set theory that is used most often in the field of engineering to discriminate between vague concepts or when quantitative precision is lacking. FASE is based on the principle of inverse inference and possesses properties not unlike many Bayesian classifiers. This non-parametric technique is particularly well suited for pattern classification problems and was developed specifically for the ACTOR study.

Third, we use a split-sample validation design to validate a model that forecasts country instability out 15 years. One portion of the historical database (the training set) is used to train or fit a model of country instability. To determine how well the patterns in these relationships can be discerned by the algorithm and ultimately forecast, the macro- structural values only in the other portion of the data set (the test set) are used to estimate the likelihood that countries with a given configuration of macro-structural values will experience a certain level of intensity of instability. We then compare the model’s projections with actual occurrences in each country over the period covered in the test set and compute some performance metrics. The results of these out-of-sample validation analyses provide insight into how accurate our “true” forecasts (e.g., projections into the future) are likely to be.

Finally, we generate annual forecasts of the likelihood of country instabilities over the period 2001-2015. To do so, we use the entire historical data set as a training set, and, using the historical data as a baseline, apply a simple forecasting algorithm to project out to the year 2015 the trend exhibited by each macro-structural factor for each of the 171 countries contained in the data set. Based on the patterns exhibited in the training set over the period 1975-1999, and the forecast values of the macro-structural factors, we compute the likelihood that each country will experience some (generally specific) level of instability over each of the next 15 years. The forecasts suggest that East Africa and South Asia will continue to harbor highly unstable states through the indefinite future (or at least through 2015). However, these regions are also home to most of the states whose changing structural environments are expected to improve their prospects for greater stability.


Since 1945, most wars, crises, and armed conflicts have occurred not between the great powers but rather within and between underdeveloped states in Africa, the Middle East, South and Southeast Asia, and Central America (Wallersteen and Sollenberg, 1999; Pfesch and Rohloff, 2000). As Professor Kalevi Holsti (1996) points out, it is the anarchy within states in these regions rather than between them that accounts for most of the wars in the last half of the 20th century.

States that are experiencing wars and state crises are unstable; that is, they are confronted with internal or external incompatibilities, which they, or others, are attempting to resolve through conflict. Both the quantitative and qualitative conflict literatures suggest that a state’s macro-structural factors—its economic interactions, both internal (Gurr, 1970; Muller and Seligson, 1987; Boswell and Dixon, 1993; Schock, 1996) and external (Barberi, 1996; Oneal and Russett, 1997, 1999), its resource endowment (Maxwell and Reuveny, 2000), demographic pressures (Choucri and North, 1975; Wils et al., 1998), ethnic and religious diversity (Ellingsen, 2000; Vanhanan, 1999) and government performance (Linz, 1978; Holsti, 1996)—together, in some configuration, form the conditions that affect its susceptibility to different forms of conflict. We refer to these macro-structural conditions that, in certain configurations, increase the likelihood of a state’s involvement in conflict as conditions conducive to instability. In doing so, we agree with Blainey (1988:X) that:

“…the beginning of wars, the prolonging of wars, the ending of wars and the prolonging or shortening of periods of peace all share the same causal framework. The same explanatory framework and the same factors are vital in understanding each stage in the sequel of war and peace.”

We would go further, however, to suggest that the same set of factors that makes states susceptible to one form of conflict (e.g., internal violence) should also, in perhaps somewhat different configurations, make them susceptible to other forms of conflict (e.g., wars) or to periods of peace. These different forms of conflict, which vary by their intensity levels of violence, reflect the different levels of intensity of instability a state might experience. Macro-structural factors—population growth and Gross Domestic Product (GDP) per capita, for instance —increase and decrease gradually, generally in predictable directions. If one could identify and validate the macro-structural factors that make states more susceptible to conflict in one configuration and less susceptible in another, then forecasts on these generally predictable factors could, in turn, be used to predict the likelihood that states will experience a certain level of intensity of instability at some generally specific point in the future.