Forecasting fatalities from state based conflicts using Markov models

Abstract:

In this contribution to the VIEWS 2023 prediction challenge, we propose using a set of different Markov-type latent state models to make prediction of fatalities from state-based conflicts on the country-month level. Partly building on the Markov modeling strategy from the VIEWS 2020 prediction contest, we propose three types of Markov-style models. First, we use an observed Markov model (OMM) which utilizes domain knowledge about conflict states to define observed states through which countries can move over time. The OMM is flexible as it does not require any parametric assumptions, and can be viewed as a set of classification and regression problems. Second, we propose a hidden (pseudo-) Markov model (HPMM) which utilizes unknowable, latent or hidden, states which the countries can move through over time. The HPMM is not strictly Markovian as we relax the assumption that the transition matrices are conditional on discrete states and instead model transitions conditional on weighted states from the posterior state probabilities. Finally, we propose aa Gaussian process continuous Markov model (GPCMM) which utilizes a continuous observed Markov ‘state’ through which countries move over time.

Authors:

David Randahl and Johan Vegelius <br/>