The region today faces a set of multi-faceted conflicts that undermine peace, sustainable development, and human rights. Deadly violence presents a major barrier for the Arab States to address deep-rooted structural challenges that prevent sustainable development. Adapting to the impacts of climate change, creating economic opportunities, and addressing social grievances to meet the demands of more accountable, inclusive, and participatory institutions, to name a few, are tremendous collective challenges that Arab societies will be hard-pressed to achieve in the absence of peace and the political resolution of current disputes. Thus, it is critical for the Arab States and their international partners to utilize tools that help in efforts to address the question of how deadly violence and its escalation can be prevented.

The uses of the tool are multiple: it may provide quantitative data to complement the existing qualitative and programmatic assessments that the UN system conducts to anticipate deadly violence, and its forecasts can serve as a benchmark to be usefully contrasted against current policy agendas to assess how societies are faring against the goals outlined by the international community, or national and local governments.

By developing baseline scenarios, the ViEWS-ESCWA model may thus also support a more integrated connection between early warning and policy responses, help implement violence-sensitive policy actions and programming, and contribute to develop contingency planning and risk management in fragile environments.

The ViEWS-ESCWA model generates monthly probabilistic assessments of the risk of deadly violence, based on the original ViEWS system. The model is specialized to predict incidences of deadly violence occurring between two or more armed actors, of which at least one is a government of a state, concerning a contested incompatibility over government and/or territory. This type of violence is categorized as ‘state-based violence’ (sb) by the Uppsala Conflict Data Program (UCDP), whose definitions underlay the ViEWS-ESCWA model.

At the country (national) level, the ViEWS-ESCWA model presents the predicted probability of at least 25 fatalities from state-based violence in each country and month. At the PRIO-GRID level (subnational), the system instead assesses the likelihood of at least one fatality per grid cell and month.

To produce the forecasts for each release, the ViEWS-ESCWA model makes use of historical time-series data on a myriad of violence predictors from 1989 up until and including the most recently available data. Data related to similar themes of violence predictors, such as the history of past violence or indices of good governance, are grouped together into thematic sub-models and fed into a number of so-called random forest algorithms. 

Random forests are machine-learning algorithms that learn from historical observations in order to produce forecasts. When ‘training' the sub-models, the random forest uses some data points to identify combinations of a handful of predictors that are particularly good at predicting deadly violence for another set of data points. It repeats this many times and votes up predictors that consistently produce good predictions. This procedure is especially useful for identifying the most promising predictors from a large number of candidate predictors. 

In a final stage, the outputs from the sub-models are combined into two composite ‘ensemble' models that produce the final forecasts: one ensemble that is trained to generate national level predictions, and one that is trained for the more fine-grained subnational level. In short, they are respectively referred to as the country-month (cm) ensemble, and the PRIO-GRID-month (pgm) ensemble. 

The aforementioned procedure of combining broad collections of sub-models is known as `ensemble modeling' and is one of the main pillars of the ViEWS-ESCWA model. Much like a crowd is wiser than the single individuals composing it, models that collect forecasts from diverse thematic (or so-called constituent) models are known to achieve more accurate predictions. This aspect of the system not only improves how well the prediction system performs, but also helps us understand how individual predictors contribute to the risk assessments described below.

More on the methodology behind the ViEWS-ESCWA forecasts, and a list of the sub-models informing each ensemble, can be found in here. [link to technical paper].

This tool is not the only approach to early warning, nor does it respond to all needs in terms of violence analysis. It complements existing qualitative frameworks and assessment processes that the UN or Member States conduct, but it does not replace them. The predictions are not absolute. Deadly violence is not linear, and its prediction is subject to complex dynamics attached to human and social systems. No model can capture all the dynamism that characterizes violence settings, and predictions are subject to uncertainties that stem from data and modelling choices, as well as from the limitations inherent to a comprehensive setup that cannot possibly capture all context-specific dynamics in detail. Nevertheless, the model can help pinpoint areas that deserve particular attention, contribute to design effective policy intervention, and allocate resources where they are most needed, as well as serve as an entry point for further analysis.

The geographic coverage of this tool is the member states of ESCWA, including Djibouti. The countries considered are: Algeria, Bahrain, Djibouti, Egypt, Iraq, Jordan, Kuwait, Lebanon, Libya, Mauritania, Morocco, Oman, Palestine, Qatar, Saudi Arabia, Somalia, Sudan, Syria, Tunisia, United Arab Emirates, and Yemen.

The forecasts are presented at two levels of analysis – for entire countries (national) and geographic subnational locations. At both levels, forecasts are monthly. The levels are referred to as cm (for country-months) and pgm (for the subnational level, given the PRIO-GRID cell structure).

The forecasts are updated every three months.

The ViEWS-ESCWA forecasts are generated by two different forecasting models: one that is trained to predict deadly violence at the national level, and one that is trained to predict deadly violence at a geographic subnational resolution of approximately 55x55 km. 

As suggested by the term, each of the aforementioned ‘ensembles’ are composed by a number of smaller sub-models that address the forecasting problem from a different angle. More specifically, the sub-models are informed by data on different themes of violence predictors (such as the recent history of conflict, the level of development in a country, or societal vulnerability to climate extremes).

Broadly speaking, the sub-model assessments indicate the relative role that the forecasting system expects each group, or theme, of violence predictors to play in driving future violence in a given country or subnational location. 

For example, if a sub-model on the theme of societal vulnerability to climate extremes generates a high-risk alert for a given country, we can expect the vulnerabilities captured by this model to serve as important drivers of future violence. As such, the forecasts generated by each of the thematic sub-models can be a fruitful resource and complement to the main deadly violence forecasts. They can help identify fragilities and provide informed suggestions for future policy interventions. 

More about the ViEWS-ESCWA methodology, as well as a description of the thematic sub-models that inform each model ensemble, can be found in here. [link to the technical paper].

There are several reasons why the forecasting models trained to generate forecasts at the national level differ from those that are used to produce subnational forecasts. The key reason lies in the different prediction targets. The national level models are trained to predict 25 or more fatalities from violence in a given country and month, whereas the subnational models set out to predict when and where at least one fatality will occur. 

Due to the varying prediction targets, the national level models give particular weight to structural, slow-moving features and patterns that often characterize countries over a longer period of time, such as the stability of political institutions, democracy indices, and socio-economic factors. They also rely heavily on conflict history data that has been aggregated to the country-month (cm) level of analysis in order to allow for broader comparisons. 

The forecasting models trained to generate subnational forecasts instead excel in accentuating effects from local compound risks. This includes – but is not limited to – heightened risks related to local demography, terrain, proximity to natural resources, local precipitation levels, societal vulnerabilities to climate extremes, and conflict history in neighbouring areas. 

The forecasts for each level of analysis do inform each other. The two assessments should nevertheless be seen as separate but related forecasts that are best interpreted in conjunction with each other. 

The sub-national level, outlined by the PRIO-GRID, is divided into quadratic geographic grid structures with cells corresponding to an area of approximately 55x55 kilometers, or 0.5x0.5 degrees. At the PRIO-GRID level (subnational), the ViEWS-ESCWA model assesses the likelihood of at least one fatality per grid cell, per month from deadly violence involving a government of a sate.

The ViEWS-ESCWA model presents forecasts for the predicted probability of deadly state-based (sb) violence, as defined and recorded by the Uppsala Conflict Data Program, UCDP. State-based violence refers to deadly violence over a contested incompatibility concerning government and/or territory that is set between two or more armed actors, of which at least one is a government of a state.

A detailed definition can be found at https://www.pcr.uu.se/research/ucdp/definitions.

ViEWS-ESCWA forecasts are presented at two levels of analysis – for entire countries (national) and for geographic (subnational) locations.

Country-month (cm) level: One of two levels of analysis, in which calendar months serve as the temporal units, and countries – as listed and geographically delimited by the CShapes dataset – comprise the spatial units. Forecasts at this level show the predicted probability of at least 25 fatalities per country and month from state-based violence.

PRIO-GRID-month (pgm) level: One of two levels of analysis, in which calendar months serve as the temporal units, and PRIO-GRID cells comprise the spatial units. The PRIO-GRID is a global quadratic grid structure with cells measuring 0.5 x 0.5 decimal degrees in longitude and latitude, or approximately 50x50 km along the equator (https://grid.prio.org/#). Forecasts at this level show the predicted probability of at least 1 fatality per PRIO-GRID cell per month from state-based deadly violence.

The ViEWS-ESCWA model includes several predictors of deadly violence that are associated with climate change, such as drought. These climatic indicators include information about agricultural production, growing season, and measures of societal vulnerability to climate shocks, which provide interesting entry points to understand how climatic variables might interplay with other known drivers of deadly violence.

For more information about the association between climate change and violence, please refer to the Climate-Conflict risk research paper by ESCWA, which can be found here. (Link to climate-conflict paper)