ViEWS2020

Identifying and Weighting the Predictors of Conflict

Building the constituent models

The ViEWS2020 forecasts were produced by advanced models that compiled, analysed, and evaluated historical time-series data from 1989 up to the month prior to each run of the model (each update of the forecasts). The data covered a multitude of variables that decades of peace research had shown to correlate with political violence, or – conversely – with the lack thereof.

Variables that shared a common theme, such as conflict history or different measures of the strength of political institutions, were grouped together into so-called “constituent models”, which were trained and fitted independently. They are described at length in our 2021 Special Data Feature in Journal of Peace Research.

Within the constituent models, each theme of variables was fed into a number of so-called random forest algorithms – machine learning algorithms that learn from historical observations in order to generate forecasts for the future. The algorithms used a subset of the available data to identify predictors that performed particularly well in predicting conflict for a later subset of the same data. It repeated this multiple times, generating a list of the predictors in each theme that perform well over and over again – even taking into account the prevalence of non-linear relationships and interactive effects amongst the pool of predictors. Along with a calibration procedure, the result from this exercise was used to determine the relative weight that was placed on each variable when generating the constituent model forecasts, and (if needed) to weed out variables that have no bearing on the results.

Forecasting Violence With the “Wisdom of the Crowd”

Compiling the model ensembles

Once the thematic constituent models had been trained and fitted, they were combined into broader models known as “ensembles” – a key tenet of all VIEWS models. Much like a crowd is wiser than the single individuals composing it, broader models that make use of forecasts from a number of smaller and specialized models are known to generate more accurate predictions. In addition to the benefits of incorporating multiple themes of conflict predictors and thus becoming more comprehensive forecasting models, ensembles are less sensitive to overfitting and more robust to new data.

The ViEWS2020 forecasts were generated by means of two such model ensembles: one that incorporated forecasts from constituent models trained specifically to predict conflict at the country level, and one that was trained for geographically refined locations spanning approximately 55x55km each (0.5×0.5 degrees). Both ensembles used calendar months as the temporal unit of analysis. They were known as the country-month ( cm) ensemble and the PRIO-GRID-month (pgm) ensembles and each contained a list of constituent models that were interpretable on their own and that had shown to improve the predictive performance of either one of the two ensembles. 16 models met these criteria for the cm ensemble, and 12 for the pgm ensemble. An overview of these is presented in the model section below, described in depth in the ViEWS2020 Special Data Feature in Journal of Peace Research.

Estimating the Model Weights

Similar to the evaluation procedure that the individual conflict predictors were subjected to in order to single out the most important variables in the constituent model forecasts, also the constituent models themselves underwent a weighting procedure upon incorporation into the final ensembles.

Up until February 2020, when the ViEWS2020 model was launched, simple unweighted model averaging emerged as the preferred weighting solution for both levels of analysis, as this method produced similar results to more complex weighting alternatives. This meant that the final ensemble forecasts were estimated as a simple average of the forecasts generated by each of the included constituent models. Following the launch of the ViEWS2 data infrastructure, which provided more data for model weighting, we however shifted to Ensemble Bayesian Model Averaging (EBMA) for the country-month ( cm) level. EBMA allows for inclusion of more models that specialize for subsets of the data, in addition to broader ones, resulting in more accurate forecasts. At the geographic (pgm) level, unweighted model averaging however continued to be used, since the EMBA procedure did not improve the performance of forecasting system enough to justify a change.

The two procedures above are discussed at length in the ViEWS2020 Special Data Feature in Journal of Peace Research. Additional information is also found in Appendix D to that article, available on our publications page. 

More on model estimation (pdf)

Computing the Forecasts

To compute the forecasts, the ViEWS2020 model made use of two strategies: dynamic simulation (ds) and one-step-ahead modeling. The former built on the procedures discussed in Hegre et al. (2013) and Hegre et al. (2016), where it is discussed at length. Both strategies are also discussed in ViEWS’ Special Data Feature in Journal of Peace Research and its appendices.