NRC: Uncertainty of Forecasting Fatalities (UFFAC)
The Uncertainty of Forecasting Fatalities in Armed Conflict (UFFAC) project
About
Taking uncertainty into account when forecasting future conflict is of paramount importance to facilitate use of data-driven predictions for anticipatory action and crisis response. Uncertainty of Forecasting Fatalities in Armed Conflict (UFFAC) is the first comprehensive research project of its kind that addresses this need.
Building onto and expanding the Violence & Impacts Early Warning System (VIEWS), the UFFAC project sets out to develop prediction models that forecast the number of fatalities in armed conflict while carefully exploring and assessing the multiple sources of uncertainty thereof with each monthly release of new VIEWS data. In addition to a thorough review of the extent to which prediction uncertainty can be fully assessed, the project will also explore how best to evaluate the quality of the forecasting models, taking both forecasts and their associated uncertainty into account.
Over the course of the UFFAC project, we will expand the VIEWS model from predicting only state-based conflict to also predicting direct deaths from one-sided violence and non-state conflict, per the UCDP definitions thereof.
Objectives
WP1: Uncertainty
To address statistical uncertainty, UFFAC will made systematic use of bootstrapping approaches, explore missing data and measurement uncertainty issues, and also explore whether uncertainty itself is a useful early warning signal.
WP2-3: Model development
UFFAC will review and specify a number of algorithms and combination of predictors that together form a set of specialised forecasting models. It will then explore techniques to combine the insights from these constituent models into an aggregated ‘model ensemble’ (a set of models), while processing their forecasts so that the end result comes as close to the distribution of the outcome variable as possible.
WP4: Evaluation
UFFAC will develop a protocol for evaluating whether the forecasting models are optimal, discussing the criteria for specifying what is optimal as well as the technical solutions. The entire model system will be validated by a gradual integration with the VIEWS early-warning system.
Funders
UFFAC is funded by the Norwegian Research Council (NRC).
