Report

Teaser, summary, work performed and final results

Periodic Reporting for period 2 - ViEWS (The Violence Early-Warning System: Building a Scientific Foundation for Conflict Forecasting)

Teaser

The aim of the project is to carry out the basic research to build ViEWS -- an ambitious political violence early-warning system. A useful system must have uniform, global coverage and frequent updates to avoid blind spots and specifically provide alerts that are as location...

Summary

The aim of the project is to carry out the basic research to build ViEWS -- an ambitious political violence early-warning system. A useful system must have uniform, global coverage and frequent updates to avoid blind spots and specifically provide alerts that are as location- and actor-specific as possible. It must be transparent and publicly available to anyone in the sense that all risk assessments may be traced back to a fully specified argument and publicly available data and tools with which assessments were created. It must build on the most advanced theoretical and empirical research within the field, and be based on systematic evaluation of predictive performance on data not used to train the system. It must cover a range of related but distinct forms of political violence, most importantly armed conflicts between organized state armies or non-state armed groups, violence against civilians perpetrated by governments and armed groups, and the forced displacement of civilians that often follows in conflicts\' wake. Warnings must be issued with respect to the geographical location of the violence as well as the actors involved.

Succeeding in this objective is important for society. Large-scale political violence kills and maims thousands of people every month across the globe. For every person killed, hundreds are forced to relocate within countries and across borders. Armed conflicts have disastrous economic consequences, undermine political systems and public services, prevent developing countries from escaping poverty, and hinder international actors in providing humanitarian assistance. The challenges of preventing, mitigating, and adapting to large-scale political violence are particularly daunting when it escalates in locations and at times where it is not expected. Policy-makers and first responders would benefit greatly from a system that systematically monitors all locations at risk of conflict and assesses the risks of conflict escalation. If fully successful, ViEWS will help crisis responders to prepare for and prevent conflict-induced humanitarian disasters. Such a system will be useful for domestic actors and international NGOs, ensure maximum transparency and credibility regarding decisions made on the basis of specific warnings. In addition, it provides the scientific benefits of better understanding the causes, connections and consequences of conflict.


When planning the project, we have broken it into five sub-objectives.

- Develop the data-collection and programming routines required for a pilot ViEWS system

- Formulate theoretically informed models of violent processes and integrate in an ensemble forecast

- Solve methodological challenges to incorporate spatial and temporal dynamics across multiple levels of analysis to integrate, weight, and improve forecasts

- Highlight the theoretical implications of the integrated and evaluated modeling approach

- Compare ViEWS with other forecasts in the field

Several of these sub-objectives imply conducting basic research to develop methods for training of predictive models and for evaluating their performance, and the adaptation of theoretically founded empirical models from the extant relevant literature. Is it even possible to construct a high-quality political violence early-warning system based on openly accessible data? How good are such systems?

The first sub-objective, however, entails setting in production a live forecasting pilot that provides monthly forecasts for all of Africa. Currently, the pilot produces forecasts at the country level -- the probability of a conflict event in any African country -- and the geographic level -- the probability of a conflict event in any of approximately 10,000 grid cells with a size of 0.5x0.5 degrees (about 55x55km), as defined by the PRIO-GRID project. The live pilot has been running since June 2018 and we are currently finalizing the first major revision of the system, based on the basic research carr

Work performed

\"In the following, and in all other text elements in the report, we refer to items in the project publication list as \'PP#\' where # is the item number. PP10, for instance, is the 2019 article in Journal of Peace Research that presents an overview of the forecasting pilot. We refer to items in the list of dissemination and outputs as \'DO#\'.

The most visible output from the project is the initiation of the live forecasting pilot, accessible at the project website. As elaborated on in the section on Project achievements, this has entailed constructing a large and complex database and an associated data ingestion system, developing forecasting models that maximizes predictive performance while retaining interpretability, implementing simulation routines for prediction models where useful solutions are not available elsewhere, and construct a pipeline that processes all steps required to run monthly updates to the forecasting pilot.

From June 2018, ViEWS has had a complete first version of the pilot running, with updated forecasts for the coming 36 months published every month. The figure \'ViEWS flowchart\' summarizes the entire process underlying it. The routines ensure that the forecasts have excellent consistency and that operations are reliable and secure. The procedures handle a large amount of data and complex models, so considerable efforts have been invested in maximizing efficiency. Efficiency and automatization is critical for operating a live pilot with regular updates. The investments put down to ensure this, however, also opens up new possibilities for exploring new models and methods on top of the excellent infrastructure the team has built.

The current pilot produces forecasts for all the three forms of political violence coded by the Uppsala Conflict Data Programme - state-based conflict between a government army and the army of another government or an organized, armed non-state actor; one-sided violence in which government or non-state armed groups kill unarmed civilians; or non-state conflict where non-state armed groups fight each other. The first step in the procedure is to monitor past conflicts. In the figure \'History of non-state conflict, June 2019\', locations with recent non-state conflict events are shown with red color, whereas locations that have been calm over the past few years have purple color.

ViEWS provides forecasts for such conflict at the country level as well as at a 30x30 arch minute grid cell. Figure \'Forecasts for August 2019, state-based conflict, country level\' shows country-level forecasts from the run we published in early August 2019. Figure \'Forecasts for August 2019, one-sided violence, geographical level\' shows the grid-cell forecasts from the same run, but for one-sided violence. Figure \'Change in predicted probabilities, state-based conflict, geographical level\' shows how our forecasts for state-based conflict have changed from what we published one month earlier.

The research has also entailed developing an evaluation system. This system must be tailored to the particular prediction problem ViEWS is tasked to meet. We have chosen (and justified) a scheme for splitting the data into partitions for training, calibration, and testing.

The ViEWS forecasting models are put together in a set of model ensembles, one for each level of analysis and each form of political violence. The individual models constituting the ensembles have been constructed partly to maximize interpretability, partly predictive performance.

In the current pilot, all the input data are openly available quantitative data. The final sub-objective above is to compare ViEWS forecasts with other types and sources of forecasts. One such type of forecasts are assessments by scholars with deep knowledge of individual countries and regions in Africa. In order to compare ViEWS fairly with expert assessments, we need forecasts from experts that are cast in the same metric as ViEWS. To obtain these, we have develo\"

Final results

The ViEWS pilot is the most ambitious live, open-source conflict early-warning system in existence. Other open-source projects (e.g. Coupcast and the US holocaust memorial museum\'s Early warning project) are begin updated regularly and of excellent quality, but only publish forecasts at the country level and for a more restricted set of outcomes. ViEWS is unique in its multi-level, multi-outcome forecasting ambition. This ambition level has merits on its own, but a main rationale for ViEWS is that pursuing multiple related forecasting problems simultaneously allows the solutions to inform each other. ViEWS has already demonstrated (in PP10) that this strategy bears fruits, but we also believe that there is considerable untapped potential to be explored in the next iterations of the forecasting system.

Other forecasting efforts have been developed, but they are not publicly available and lack the transparency and replicability of ViEWS and the projects mentioned above. We believe that transparency is the main virtue of an academic project like this. ViEWS is unlikely to be able to compete with large countries\' intelligence agencies\' systems, but public availability allows the results and the forecasts to enter the public debate about the serious social problems that individual armed conflicts are for those affected.

Methodologically, ViEWS is using several models that have only recently been introduced to social sciences from the fields of machine learning.

The evaluation procedures used to evaluate the current ViEWS pilot build on established practices, but the adaptation to the particular problem at hand contains novel elements. In the next iteration of the system, we will introduce a more sophisticated data partitioning setup, where we time-shift data systematically and individually for each forecasting horizon (i.e., having separate time shifts for 1-month-ahead and 12-month-ahead forecasts). This allows ViEWS to match all possible combinations of forecasts and outcomes inside a test window. This, in turn, yields improvements both in the certainty of our evaluation procedures and the performance of the forecasts themselves.

ViEWS has made good progress in formulating a new evaluation metric that allows us to credit forecasts that miss by only a small amount (in time or in space) more than forecasts that miss by large distances. Such metrics are currently not in use in peace and conflict research.

Another activity that goes well beyond the state of the art is the ViEWS expert survey which we are gradually building up. We have recruited 35 experts on individual countries in Africa. We ask them two sets of questions. In the first set, we ask the experts to identify the most important conflict issues and to assess the positions of the main stakeholders within the country with respect to these issues. Over repeated surveys, we will trace changes in these positions as well as new demands and threats of the use of violence that stakeholders make to attempt to influence the outcomes. From this analysis, we will derive \'implicit predictions\' of violence using game-theoretic models. In the other set of questions we ask the experts how likely they believe it is that the political violence situation will change based on their knowledge of the country.

Website & more info

More info: https://views.pcr.uu.se.