Forecasting Civil Wars: Theory and Structure in an Age of “Big Data” and Machine Learning

• Date: 01 November 2020
• Author: Robert A. Blair,Nicholas Sambanis
• DOI: 10.1177/0022002720918923
• Published date: 01 November 2020
• Subject Matter: Articles

Article

Forecasting Civil Wars:

Theory and Structure

in an Age of “Big Data”

and Machine Learning

Robert A. Blair

1

and Nicholas Sambanis

2

Abstract

Does theory contribute to forecasting accuracy? We use event data to show that a

parsimonious model grounded in prominent theories of conflict escalation can

forecast civil war onset with high accuracy and over shorter temporal windows than

has generally been possible. Our forecasting model draws on “procedural” variables,

building on insights from the contentious politics literature. We show that a pro-

cedural model outperforms more inductive, atheoretical alternatives and also out-

performs models based on countries’ structural characteristics, which previously

dominated models of civil war onset. We find that process can substitute for

structure over short forecasting windows. We also find a more direct connection

between theory and forecasting than is sometimes assumed, though we suggest that

future researchers treat the value-added of theory for prediction not as an

assumption but rather as a hypothesis to test.

Keywords

forecasting, civil wars, event data, machine learning

1

Department of Political Science, Watson Institute for International and Public Affairs, Brown University,

Providence, RI, USA

2

Department of Political Science, University of Pennsylvania, Philadelphia, PA, USA

Corresponding Author:

Robert A. Blair, Department of Political Science, Watson Institute for International and Public Affairs,

Brown University, 111 Thayer St., Providence, RI 02912, USA.

Email: robert_blair@brown.edu

Journal of Conflict Resolution

2020, Vol. 64(10) 1885-1915

ªThe Author(s) 2020

Article reuse guidelines:

sagepub.com/journals-permissions

DOI: 10.1177/0022002720918923

journals.sagepub.com/home/jcr

Conflict forecasters have long argued that good theory is essential for accurate

prediction. Chiba and Gleditsch (2017, 2 95) argue that “the oretical atten tion to

relevant actors” should improve predictive performance and advocate “anchoring

prediction models in theories.” Brandt, Freeman, and Schrodt (2011, 46) aim to

generate “theoretically grounded, policy-relevant forecasts,” and Goldstone et al.

(2010, 194) test a variety of potential correlates of conflict that are “drawn from the

theoretical literature.” Cederman and Weidmann (2017, 476) conclude their

review of the literature by arguing strongly for the necessity of theory for conflict

prediction.

In recent years, however, claims about the importance of theory have yielded two

related tensions in conflict forecasting research. First, some of the most prominent

theories of conflict have been shown to perform remarkably poorly when used to

predict conflict itself. Ward, Greenhill, and Bakke (2010) find that models based on

the Fearon and Laitin (2003) theory of civil war produce surprisingly inaccurate

forecasts; models based on Collier and Hoeffler (2004) do not fare much better. Hill

and Jones (2014) show that existing theories of state repression generally cannot

predict state repression itself. Blair, Blattman, and Hartman (2017) find that widely

studied causes of local violence perform poorly when used for prediction. Some

interpret this as evidence that purported causes of conflict may not be causal at all.

Beck, King, and Zeng (2000, 21) go so far as to argue that causal theories that fail to

forecast are of “dubious validity and marginal value.”

Second, while almost all conflict forecasting models are motivated to some extent

by theoretical intuitions, many are operationalized using variables that do not cor-

respond to those intuitions in a direct way. This has become especially true with the

rise of “big” event data sets like the Computational Event Data System, the Global

Database of Events, Language, and Tone, and, most prominently, the World-wide

Integrated Crisis Early Warning System (ICEW S). ICEWS contain s data on over

300 categories of events between actor dyads, ranging from “deny responsibility”

to “grant diplomatic recognition” to “engage in ethnic cleansing.” This would

seem to offer many opportunities for applying theory to prediction. Yet while

many conflict forecasters have adopted ICEWS, most have done so by collapsing

disparate event types into broad indices that distinguish conflictive events from

cooperative ones and verbal events from material ones (e.g., Beger, Dorff, and

Ward 2016; Chiba and Gleditsch 2017; Montgomery, Hollenbach, and Ward 2012;

Schrodt, Woodward, and Marshall 2011). As informative as these studies have

been, such broad distinctions between conflict and cooperation are too generic to

capture specific theoretical insights.

One way to resolve these tensions might be to combine the best of two distinct

approaches to conflict forecasting. Early conflict forecasting models were inspired

by “structural” theories of civil war that emphasize the role that regime type, per

capita income, natural resource endowments, rough terrain, and other slow-moving

or time-invariant characteristics play in increasing or decreasing the risk of civil war

onset. But with few exceptions, these structural variables change too slowly to

1886 Journal of Conflict Resolution 64(10)