Back to search

FRIPRO-Fri prosjektstøtte

The Uncertainty of Forecasting Fatalities in Armed Conflict (UFFAC)

Alternative title: Usikkerhet forbundet med å forutsi antall drepte i væpnet konflikt (UFFAC)

Awarded: NOK 12.0 mill.

The UFFAC project studies the methodologies of producing forecasts for the number of people killed in organised political violence, with a comprehensive evaluation of the uncertainty such forecasts must have. The project systematically reviews various sources of uncertainty, such as that stemming from selecting what is the best prediction model, or whether processes and patterns that we have observed in the past will be repeated in the future. Another fundamental uncertainty relates to evaluation -- the ambiguities of ranking or weighting models according to multiple, reasonable but perhaps mutually exclusive criteria. We make use of 'bootstrapping', a technique where we compare how different parts of the available data yield different results. To better understand the relative advantages and disadvantages of various modeling strategies, we explore a large variety of predictors and statistical/machine-learning algorithms. The predictions we develop are in the form of probability distributions over all possible outcomes, such as the probability of observing zero deaths in a country in a month, that of observing one, two, three, etc. We explore a range of evaluation metrics designed to assess the performance of prediction models that generate probability distributions, comparing predictions from models learned from one part of the dataset when they are used on a different part of the dataset. UFFAC contributes to the ongoing VIEWS project (https://viewsforecasting.org) which provides monthly updated forecasts of the number of fatalities in organized political violence and is developing similar forecasts for some of the main humanitarian impacts of such violence.

The UFFAC project will study the methodologies of producing forecasts for the number of people killed in organised political violence, with a comprehensive evaluation of the uncertainty such forecasts must have. The project will systematically review various sources of uncertainty, such as that stemming from selecting what is the best prediction model, or whether processes and patterns that we have observed in the past will be repeated in the future. Another fundamental uncertainty relates to evaluation -- the ambiguities of ranking or weighting models according to multiple, reasonable but perhaps mutually exclusive criteria. We will make use of `bootstrapping', a technique where we compare how different parts of the available data yield different results. To better understand the relative advantages and disadvantages of various modeling strategies, we will explore a large variety of predictors and statistical/machine-learning algorithms. The predictions we develop will be in the form of probability distributions over all possible outcomes, such as the probability of observing zero deaths in a country in a month, that of observing one, two, three, etc. We will explore a range of evaluation metrics designed to assess the performance of prediction models that generate probability distributions, comparing predictions from models learned from one part of the dataset when they are used on a different part of the dataset. UFFAC will contribute to the ongoing VIEWS project (https://viewsforecasting.org) which provides monthly updated forecasts of the number of fatalities in organized political violence and is developing similar forecasts for some of the main humanitarian impacts of such violence.

Funding scheme:

FRIPRO-Fri prosjektstøtte

Funding Sources