Overview of current research areas
Advancing Methods
Accounting for Different Speeds in Comparative Case Studies: Dynamic Synthetic Controls
Synthetic controls are widely used to estimate the causal effect of a treatment. However, they do not account for the different speeds at which units respond to changes. Reactions may be inelastic or "sticky" and thus slower due to varying regulatory, institutional, or political environments. We show that these different reaction speeds can lead to biased estimates of causal effects. We therefore introduce a dynamic synthetic control approach that accommodates varying speeds in time series, resulting in improved synthetic control estimates. We apply our method to re-estimate the effects of terrorism on income \citep{abadie2003economic}, tobacco laws on consumption \citep{abadie2010synthetic}, and German reunification on GDP \citep{abadie2015comparative}. We also assess the method's performance using Monte-Carlo simulations. We find that it reduces errors in the estimates of true treatment effects by up to 70\% compared to traditional synthetic controls, improving our ability to make robust inferences. An open-source \textsf{R} package, \texttt{dsc}, is made available for easy implementation.
Leveraging Temporal Patterns in Forecasting
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. Our methodology is evaluated with three widely used forecasting models - Autoregressive Integrated Moving Average (ARIMA), Random Forest (RF), and Long Short-Term Memory (LSTM) networks. 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 from meteorology to finance and medicine. The results show that the introduction of dynamic covariates significantly improves the prediction accuracy across all three models. Our findings underline the potential of adding recurring motifs to prediction tasks for a variety of algorithms.
Echoes of Conflict: Quantifying Redundancy in Casualty Time Series
The idea that human behavior, societal trends, and global events follow recognisable patterns, leading to similar outcomes, has long intrigued scholars and policymakers, particularly in conflict research. This interest stems from the potential to gain insights into the future by understanding whether wars repeat. If they do, valuable insights can be gained from studying the past. Despite this interest, we know little about the extent of temporal redundancy in conflicts. This article aims to quantify it using time series data on conflict casualties to measure pattern recurrence and determine if these patterns lead to predictable outcomes. By comparing conflict patterns with those in earthquakes, epidemics, and the stock market, we assess how conflict predictability compares to other fields. We find that while conflict data occasionally display repetitive patterns, they are less predictable than patterns in seismology or epidemiology. Determining the redundancy of temporal patterns is crucial for understanding whether conflicts are driven by recurring processes or are more random, thereby offering insights into the predictability and underlying dynamics of conflict events. This study contributes to the discourse on historical event predictability and offers a methodological framework for future research on conflict dynamics, emphasising the complexity and unpredictability of warfare.
Learning Mixtures of Gaussian Processes through Random Projection
We propose an ensemble clustering framework to uncover latent cluster labels in functional data generated from a Gaussian process mixture. Our method exploits the fact that the projection coefficients of the functional data onto any given projection function follow a univariate Gaussian mixture model (GMM). By conducting multiple one-dimensional projections and learning a univariate GMM for each, we create an ensemble of GMMs. Each GMM serves as a base clustering, and applying ensemble clustering yields a consensus clustering. Our approach significantly reduces computational complexity compared to state-of-the-art methods, and we provide theoretical guarantees on the identifiability and learnability of Gaussian process mixtures. Extensive experiments on synthetic and real datasets confirm the superiority of our method over existing techniques.
A Structured Treatment Approach for Time Series Counterfactual Estimation
This paper presents a novel framework for counterfactual estimation in time series data with a focus on continuous treatment scenarios. Counterfactual reasoning is essential for understanding causal relationships and predicting potential outcomes under different hypothetical scenarios. Unlike prior research that predominantly addresses categorical treatment in time series analysis, this work incorporates advancements in time series forecasting models to improve causal inference tasks. The proposed model leverages generalised Robinson decomposition to extract latent features from both treatments and covariates, combined with a cross-fitting approach for robust counterfactual estimation. The effectiveness of the model is demonstrated through experiments on semi-synthetic datasets, comparing it with existing methods, including RNN-based, Transformer-based, and Bayesian models. Applications of the model are explored in two distinct domains: the economic impact of conflicts, using a stochastic differential equation model for GDP, and the effect of greenhouse gas emissions on climate, modeled through an energy balance framework. The results highlight the model's ability to capture complex dependencies and provide accurate counterfactual predictions in time series data. The paper concludes with a discussion of limitations and potential future directions for enhancing counterfactual estimation in dynamic systems.
PanelDiD: A Difference-in-Differences Estimator for Causal Inference with TSCS Data
This paper addresses the challenge of deriving robust causal effects from time-series cross-sectional (TSCS) data. The task is especially complex with multiple treatment status changes, heterogeneous treatment effects, and unobserved time-varying confounders, leading to increased bias and reduced efficiency. Here, we introduce a novel difference-in-differences (DID) estimator to assess the average treatment effect on the treated (ATT), building upon the principles of doubly robust DID estimation. Our approach involves creating matched sets by pairing each treated observation with control observations from different groups that share an identical treatment history. We then employ a combination of propensity score and outcome regression methods, incorporating machine learning algorithms with cross-validation, to calculate both immediate and long-term ATTs. Our simulation and empirical analyses demonstrate the estimator's semi parametric efficiency and resilience to incorrect model specifications. We also introduce an open-source software package for these methods' implementation.
Identifying Temporal Patterns
An Automated Pattern Recognition System for Conflict
This article introduces an automated pattern recognition system for conflict. The monitoring system aims to uncover, cluster, and classify temporal patterns of escalation to improve future forecasts and better understand the causes of escalation toward war. It identifies important temporal patterns in conflict data using novel pattern detection methods and new data. These patterns are used to forecast conflict, with live predictions released in real time. Finally, the discovery of recurring motifs - prototypes - can inform new or existing theoretical frameworks. This article discusses the methodological innovations required to achieve these goals and the path to creating an autonomous conflict monitoring system. It also reports on promising results obtained using these methods, which show that they perform well on true out-of-sample forecasts of the count of the number of fatalities per month from state-based conflict. The monitoring system has important implications for computational diplomacy, as it can alert diplomats of geopolitical risks.
Temporal Patterns in Migration Flows : Evidence from South Sudan
What explains the variation in migration flows over time and space? Existing work has contributed to a rich understanding of the factors that affect why and when people leave. What is less understood are the dynamics of migration flows over time. Existing work typically focuses on static variables at the country-year level, and ignores the temporal dynamics. Are there recurring temporal patterns in migration flows? And can we use these patterns to improve our forecasts of the number of migrants? Here, we introduce new methods to uncover temporal sequences - motifs - in the number of migrants over time, and use these motifs for forecasting. By developing a multivariable shape similarity-based model, we show that temporal patterns do exist. Moreover, using these patterns results in better out-of-sample forecasts than a benchmark of statistical and neural networks models. We apply the new method to the case of South Sudan.
Identifying Motifs and Patterns in War Fatalities: A Matrix Profile Application
Armed conflicts are complex processes characterised by non-linearities, feedback loops, and recurrent patterns of occurrence and diffusion. Recent progress in statistical and computational methods coupled with an increasing availability of highly granular data have increased our ability to forecast armed conflict occurrence. However, conflict dynamics and related variations in conflict related fatalities remain poorly understood. Improving our understanding of the spatiotemporal patterns and recurrent motifs that characterise conflicts is paramount to better understand and forecast conflict dynamics. This paper attempts to fill this gap to identify recurrent motifs and patterns underlying casualties in wars. We apply matrix profiles on monthly data on battle death estimates to uncover, cluster, and classify patterns and motifs in armed conflict. Next, we empirically assess the contribution of the derived motifs in forecasting the future course of casualties, and discuss their implications for theoretical developments. The identification of underlying patterns characterising war fatalities enable us to increase forecasting performance and improve our understanding of conflict processes.
Temporal Patterns in Conflict Prediction: An Improved Shape-based Approach
Existing models for predicting conflict fatalities are risk-averse because they often yield overly cautious forecasts around the mean. Although these approaches tend to be accurate if evaluated by the prediction error, they offer limited insights into temporal variations in conflict-related fatalities, or rather sudden surges and declines in battle deaths. However, accounting for variability in fatalities seems particularly relevant for policymakers. In this article, we introduce a novel risk-taking methodology, the Dynamic Temporal Patterns (DTP) model, to capture variability in fatalities data. The method involves isolating historically analogous sequences of fatalities to create a reference repository, focusing on the most similar historical cases in comparison to the input sequence. Predictions are then generated by analysing the average future outcomes of these reference sequences, enabling the model to make more dynamic and hence riskier forecasts. Our approach maintains high accuracy while significantly enhancing the ability to predict shifts, surges, and declines in conflict fatalities over time, judged by a novel evaluation metric, the difference explained. Empirical tests demonstrate that combining the DTP methodology with existing approaches, in particular the Violence Early-Warning System (ViEWS) ensemble, not only achieves a lower mean squared error, but also better accounts for variability in fatalities data, highlighting the potential of integrating risk-taking and risk-averse models. The DTP model performs particularly well in the context of \textit{high complexity} cases - fatalities time series with multiple fluctuations - while producing overpredictions for \textit{low complexity} observations.
War and Over-optimism: Financial Market Estimates of Conflicts' Economic Impact
Do observers of international relations systematically underestimate the costs of war? Perceptions of the anticipated duration and costs of war matter because they affect the pressure on leaders toward war or peace, and the accuracy of these estimates may determine how audiences reward or punish their leaders. Here, we explore how well the contemporaries of interstate wars have assessed their ultimate costs by studying the evolution of financial asset prices throughout the war. The price of government bonds, in particular, is strongly affected by war through inflation and default risks, and proxies the cost of war as estimated by its contemporaries. As conflicts progress, markets adjust their estimate until the end of the conflict, when bond prices reflect all the information available. Using worldwide weekly government bond yield data from 1816 to 2007, we estimated the market's perceived cost of war both at the onset and at the end of conflict. The discrepancy between the two, measured using an event study with synthetic controls, reveals that costs tend to be correctly estimated. Estimates for autocracies, initiators or strong states exhibit no more optimism than others. Large conflicts exhibit more variance in estimates but no overall bias. This finding has important implications for our understanding of the public's estimates of the cost of war, as well as the ability of audiences to correctly reward the risks undertaken by their leaders.
Triggers of War
The onset of interstate conflict often hinges on seemingly minor events such as the assassination of Franz-Ferdinand in 1914. However, the literature on the causes of interstate war has mostly focused on identifying fertile grounds ("powder kegs"), ignoring these intricacies of history that are typically treated as noise. Yet this approach cannot explain why certain fertile grounds remain peaceful, or why wars start precisely when and where they do. Here, we measure and demonstrate the importance of these micro-level changes with monthly information about short-term events from three different sources: a) one hundred years of newspaper articles; b) two hundred years of government bond yields; and c) fine-grained event-data. We find that these measures of short-term changes significantly improved our ability to explain and predict conflict. In particular, we found that fertile grounds or the occurrence of a trigger are both prone to conflict, but it is their combination that is disproportionately dangerous.
From Protests to Fatalities: The Role of Temporal Sequences in Civil Conflict Transitions
Understanding the temporal dynamics of protests and their potential evolution into civil conflict is critical for both scholars and policymakers. Using a novel method that incorporates time series clustering and machine learning algorithms, this study makes two central arguments: that protests evolve in recurring patterns and second, and that these patterns can help predict the transition to civil conflict. We leverage data from ACLED and UCDP, covering protests and civil conflict events across multiple countries and time periods. Our analysis reveals that protests indeed display recurring patterns that vary both within and across countries. Moreover, incorporating these dynamic sequences into predictive models improves their out-of-sample performance by about 10\%, substantiating our theoretical expectation. The study not only contributes to our understanding of the complex relationship between protests and civil conflict but also offers an innovative methodological framework for analysing sequential data in the social sciences.
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-related 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 analysing preceding dynamics in armed conflict, most importantly battle losses. We argue that preceding dynamics in rebel casualties serve to predict the transition to civilian targeting. Validating our theoretical proposition, we apply dynamic time series techniques, which investigate sequences of events and uncover 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 changes in 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 shifts towards civilian targeting.
Patterns in Protest Cycles: Evidence From Time Series Clustering
Protests are dynamic events that often unfold through consecutive periods of expansion and contraction. These cycles are the observable implications of complex interactions between dissident activity and state repression. This study explores the temporal patterns that emerge from these interactions. We show that protest sequences follow regular motifs that cannot simply be attributed to chance. With data from the Armed Conflict Location & Event Data for India, we extract clusters of similar time series - patterns - using a moving window and Dynamic Time Warping. We then show that incorporating these past patterns as dynamic covariates in our model significantly enhances its predictive accuracy. This confirms our theoretical expectations that protest sequences follow discernible sequences across space and times that are predictable by analysing previous events. Our findings contribute to our understanding of protests by emphasising the importance of sequences---and not merely individual data points---in protest movements. This approach not only enhances our theoretical comprehension of protest dynamics but also has practical implications for anticipating future protest cycles.
Forecasting Conflict
The 2023/24 VIEWS Prediction Challenge: Predicting the Number of Fatalities in Armed Conflict, with Uncertainty (collaboration)
Existing models for predicting conflict fatalities often yield cautious forecasts. Although these approaches tends to be accurate, they offer limited insight into temporal variations in conflict-related fatalities. In this paper, we introduce a novel methodology, which we call "Shape Finder", to capture complex interdependencies in fatalities data. The method involves isolating historically analogous sequences of fatalities to create a reference repository. Predictions are then generated by analysing the average future outcomes of these reference sequences, enabling the model to make more dynamic forecasts. Our approach maintains high accuracy while significantly enhancing the ability to predict shifts, surges, and declines in conflict fatalities over time. Empirical tests demonstrate that combining the shape finder methodology with existing approaches not only achieves a lower mean squared error (MSE), but also better accounts for variability in the time series data.
Agent-based Migration Simulation with Machine Learning Based Conflict Forecasting
The paper presents a novel approach to predicting forced migration by integrating machine learning-based conflict forecasting with agent-based modeling (ABM). The study addresses the challenge of accurately forecasting migration patterns, which are often driven by conflict and violence, using a coupled model that combines a Random Forest classifier for conflict forecasting with the Flee ABM. The Random Forest model is trained on conflict data from the Peace Research Institute Oslo (PRIO) and the Armed Conflict Location and Event Data Project (ACLED), using spatial and temporal inputs to predict conflict events. These conflict forecasts are then used to inform the Flee ABM, which simulates the movement of forcibly displaced people in response to these conflicts. The coupled model is validated using case studies from historical conflicts in Mali, Burundi, South Sudan, and the Central African Republic, demonstrating improved predictive accuracy over traditional methods. The study highlights the potential of this integrated approach to aid policymakers and humanitarian organisations in preparing for and responding to migration crises.
Harnessing AI: How to Develop and Integrate Automated Prediction Systems for Humanitarian Anticipatory Action
This policy paper explores how advanced forecasting systems, particularly those leveraging machine learning, can enhance humanitarian anticipatory action in response to armed conflict and natural hazards. While forecasting technologies have shown great promise in predicting natural disasters, their application to conflict-related humanitarian interventions remains limited. This paper examines the challenges and opportunities associated with integrating these predictive tools into humanitarian decision-making, emphasising the importance of developing common frameworks, ethical guidelines, and best practices for their use. By bringing together insights from both academic research and field practitioners, the paper aims to provide actionable recommendations for improving resource allocation, crisis response, and the overall effectiveness of humanitarian aid. The goal is to offer a practical roadmap for stakeholders to harness forecasting technologies to better anticipate and mitigate the impacts of future crises.