Abstract
The monitoring of conflict risk in the humanitarian sector is largely based on simple historic averages. The overarching goal of this work is to assess the potential for using a more statistically rigorous approach to monitor the risk of political violence and conflict events in practice, and thereby improve our understanding of their temporal and spatial patterns, to inform preventative measures. In particular, a Bayesian, spatiotemporal variant of the Hawkes process is fitted to data gathered by the Armed Conflict Location and Event Data (ACLED) project to obtain sub-national estimates of conflict risk in South Asia over time and space. Our model can effectively estimate the risk level of these events within a statistically sound framework, with a more precise understanding of uncertainty than was previously possible. The model also provides insights into differences in behaviours between countries and conflict types. We also show how our model can be used to monitor short and long term trends, and that it is more stable and robust to outliers compared to current practices that rely on historical averages.
Introduction
Monitoring conflict risk is vital for humanitarian "anticipatory action"—allocating resources before a crisis fully unfolds. However, current industry standards often rely on simple historical moving averages or "black-box" machine learning models. These traditional methods face two major hurdles: sensitivity to outliers and a lack of "self-excitation" modeling. Violence is rarely independent; one riot or battle often triggers another. Standard averages fail to explicitly model this "triggering" effect or how it spreads to neighboring regions. This research aims to provide a statistically rigorous, interpretable alternative that quantifies uncertainty, allowing decision-makers to assess the reliability of risk estimates.
Methodology: Discrete-Time Hawkes Processes
The core of this framework is the Discrete-Time Hawkes Process (DTHP). While standard Hawkes processes operate in continuous time, they are adapted here for the daily/weekly count data typical of conflict databases.
The model decomposes the event rate into two primary components:
- Baseline Rate: Spontaneous events modeled using an Intrinsic Conditional Auto-Regressive (ICAR) approach to account for regional differences like population density.
- Self-Excitation: Events triggered by the past, modeled using Spatiotemporal Kernels. A Geometric distribution handles temporal decay, while a Radial Basis Function (RBF) manages spatial decay.
The model is implemented using Bayesian Inference via Markov Chain Monte Carlo (MCMC) sampling in Stan. This approach produces full probability distributions for risk, incorporating a zero-inflation parameter to handle the frequent absence of events in specific regions and weeks.
Data and Study Area
The study focuses on South Asia, specifically Bangladesh, Sri Lanka, Nepal, and Pakistan, utilizing data from the Armed Conflict Location and Event Data (ACLED) project. The dataset comprises weekly event counts from 2010 to 2014 across six categories: Battles, Explosions/Remote violence, Violence against civilians, Protests, Riots, and Strategic developments.
Key Insights and Results
The Hawkes model reveals distinct behavioral patterns across different conflict types and geographies:
- Conflict "Memory": Conflict in Bangladesh and Pakistan exhibits a "longer memory" (slower temporal decay) compared to Nepal or Sri Lanka, where the influence of past events vanishes more rapidly.
- Spatial Contagion: In Nepal, the influence of an event has a broader geographic reach, whereas in Bangladesh, the triggering effect is more localized.
- Spontaneous vs. Triggered Events: Protests and riots generally have higher baseline rates, suggesting more spontaneous origins. Conversely, battles and explosions show stronger self-exciting characteristics; once initiated, they are highly likely to trigger subsequent events.
- Stability and Predictive Power: The model proved significantly more robust than "naive" rolling averages. In out-of-sample tests (predicting early 2015), the model's 95% uncertainty intervals successfully encapsulated observed event counts. The "97.5% trigger threshold" provided by the model serves as a stable, statistically grounded early warning signal, far less volatile than standard deviation-based approaches.
Conclusion
By accounting for the contagious nature of political violence, the Bayesian spatiotemporal Hawkes process offers a superior framework for conflict monitoring. While limitations such as a one-week systematic delay and the absence of exogenous covariates exist, the model provides a more precise and stable understanding of conflict risk than traditional historical averages. Future work will explore integrating event severity (fatalities) and continuous-space approximations to further refine these predictive capabilities.
Summary updated