Abstract: Recurring temporal patterns emerge naturally from underlying processes and interactions in a variety of disciplines, ranging from epidemiology and ecology to social sciences and physics. These patterns and motifs hold considerable promise for enhancing the precision of time series forecasting. This study introduces a method that identifies these repeating patterns and incorporates them as dynamic covariates in traditional time series forecasting models. Using subsequence time series clustering, we obtain cluster solutions for various window length and k combinations, and only retain the ones with the lowest standard deviation at t+1, indicative of the level of signal entailed in the clusters. Our methodology is evaluated with three widely used forecasting models---Autoregressive Integrated Moving Average (ARIMA), Random Forest (RF), and Long Short-Term Memory (LSTM). Each model is implemented in its standard form and subsequently augmented with dynamic covariates. The empirical evaluation draws on a data set comprising 1,000 time series spanning a broad array of domains. The results show that the introduction of dynamic covariates significantly improves the prediction accuracy. Through combining static and dynamic variants in a compound model, our algorithm effectively filters out conditions under which time series clusters pick up meaningful signal and discards noise. Including dynamic covariates is particularly salient for time series with disrupted seasonality. Our findings underline the potential of adding recurring motifs to prediction tasks for a variety of algorithms and problems.

Abstract: Protests often follow cycles of escalation and decline, influenced in part by the interaction between protest tactics and government responses. These dynamics can produce protest waves that vary in intensity and timing. Understanding these patterns is critical to anticipating shifts in political instability, assessing state responses, and analyzing protesters’ strategic behavior over time. Using protest data for India (2016–2023) at the local level, we apply Dynamic Time Warping and k-Means clustering to identify recurring patterns in protest time series. We find that protests evolve in sequences characterized by alternating phases of increasing and decreasing intensity. Incorporating these patterns greatly improves protest prediction accuracy. Protest patterns enhance our ability to distinguish between spontaneous, short-lived mobilization and sustained protest waves.

Abstract: Why do some protests escalate into deadly civil conflict while others do not? Protest can turn violent when participants see non-violent action as ineffective, or when rebel groups perceive protests as a strategic opportunity. Yet, armed violence is rarely spontaneous; escalation typically follows periods of calculation and adaptation. This suggests that recent protest activity (e.g., at t-1) is insufficient to explain transitions to civil conflict. Instead, it is the evolving interaction between protesters and the state that gives rise to “dangerous” protest patterns. We identify three such patterns that consistently precede high fatalities in civil conflict: a gradual escalation of protest intensity, a sharp decline following peak activity, and a U-shaped trajectory. Using country-month data on protest and battle events, we identify these patterns through statistical inference and predictive validation. Our results show that protest-to-conflict transitions depend not just on protest volume, but on protest sequences.

Abstract: Civil wars are highly complex phenomena, often described as puzzling, random, or stochastic. Rationalist explanations contend that civil war is caused by uncertainty, or rather, the unique aspects of each case. If civil war is caused by uncertainty, onsets become inherently unexplainable. Rather than seeking to establish sufficient conditions, it might be useful to think about civil war onsets in terms of stereotypical pathways. This paper uncovers stereotypical pathways to civil war by leveraging machine learning tools, which are advantageous when studying intricate phenomena. Every onset is assigned to one pathway, while allowing for the possibility that each case presents a unique manifestation of the underlying logic. The analysis reveals eight distinct civil war pathways: Economic crisis, opportunity, grievances, weak state capacity, oil curse, bad neighborhood, climate, and large population. The derived clusters illustrate how the combination of grievances and opportunity produces civil war onsets. While common pathways to civil war exist, every onset has a unique character, limiting the extent to which causal mechanisms apply across cases.

Abstract: Existing research on the causes of civil war focuses on correlations between structural background factors and armed conflict. However, the great risks and logistical challenges of armed violence make a direct link between structural factors and civil war seem implausible. Civil conflict is a dynamic phenomenon and commonly emerges from preceding and less violent collective action, such as protests, riots, terrorism, and low-intensity armed conflict. Instead of singling out civil war, it appears more reasonable to consider different types of collective action together. This paper uncovers similarities and differences in the causes of collective action and finds that structural risk factors perform equally when predicting low-intensity collective action in comparison to civil war. Similar theoretical arguments apply across outcomes, most evidently for population size and the cyclical nature of political contention. Simultaneously, the analysis highlights the unique character of less violent forms of collective action, i.e., protest and riots, for example, regarding the importance of mobile phones. While structural risk factors might prompt civil war by transitioning through protest and/or low-intensity armed violence, collective action is a non-sufficient condition for civil war, which applies particularly to protest.

Abstract: Anticipating conflict onsets—the emergence of fatalities following periods of peace—remains a core challenge in conflict prediction. Structural indicators evolve too slowly to offer timely warnings, while more reactive signals (e.g., news) capture the early stages of conflict rather than true precursors. This paper introduces a novel conflict prediction model—the Onset finder—which leverages preceding temporal patterns in collective action, particularly protests, riots, remote violence, violence against civilians, and non-state conflict, to forecast conflict onsets. Interactions between non-state groups and the government produce patterns in low-intensity collective action (i.e., protests and riots) as non-state actors attempt to pressure the state directly. Alternatively, rebel mobilization may give rise to temporal patterns in low-intensity violence, as these actors extract resources from civilians or rival groups. Based on these theoretical considerations, the Onset finder matches patterns in collective action with historical analogs and uses the aggregated future fatality sequences of the closest matches as a prediction. The Onset finder demonstrates good performance on onset cases if compared to existing conflict prediction models and several technical benchmarks.

Abstract: Why do rebel groups target civilians in some cases, and not in others? Existing research suggests that civilian targeting might constitute a “cheap” alternative to conventional tactics if the group is otherwise close to defeat, as measured by the number of battle deaths. Here, we argue that the number of deaths is not sufficient to understand the rebel groups’ strategic decisions. Instead, the decision to shift towards unconventional tactics originates from the groups’ evaluation of the likely path the armed conflict might take in the future, based on analyzing preceding dynamics in armed conflict, most importantly battle losses. We argue that preceding dynamics in rebel casualties serve to predict civilian targeting. To validate our theoretical proposition, we apply dynamic time series techniques to uncover relevant sequences of events and patterns in time series data. Time series of rebel deaths are clustered using k-Means and the Euclidean distance of time warped time sequences as distance metric. The derived clusters are included as additional covariates in a model, predicting civilian deaths based on the specific past n observations of rebel casualties. We show that accounting for temporal patterns in rebel deaths adds important information when predicting civilian targeting.

Abstract: Research on violent and nonviolent conflict is becoming increasingly disaggregated. While past literature observes contentious processes at the level of countries and campaigns, nascent literature seeks to generate new insights and improve inference by zooming in on the level of individual events. Therefore, violent event datasets have come under increasing scrutiny in recent years, with methodological research critically comparing their scope, validity, reliability, and transparency. For nonviolent event datasets, however, we still lack a systematic understanding of how they compare, and of how sensitive research findings are to their choice of data. How do protest event datasets differ in their conceptualization and operationalization? We are the first to offer a conceptual and empirical appraisal of all major data projects that record information on protest events, systematically comparing definitions, measurement, and coverage. By outlining the strengths and unique features of different data projects, our study provides an important resource for future work on contentious politics, and helps researchers and stakeholders make informed decisions on their choice of data for scholarship and practice.