Lesson (un)replicated: Predicting levels of political violence in Afghan administrative units per month using ARFIMA and ICEWS data
The aim of the present paper is to evaluate the use of the Autoregressive Fractionally Integrated Moving Average (ARFIMA) model in predicting spatially and temporally localised political violent events using the Integrated Crisis Early Warning System (ICEWS). The performance of the ARFIMA model is compared to that of a naïve model in reference to two common relevant hypotheses: the ARFIMA model would outperform a naïve model and the rate of outperformance would deteriorate the higher the level of spatial aggregation. This analytical strategy is used to predict political violent events in Afghanistan. The analysis consists of three parts. The first is a replication of Yonamine's (2013) study for the period beginning in April 2010 and ending in March 2012. The second part compares the results to those of Yonamine (2013). The comparison was used to assess the validity of the conclusions drawn in the original study, which was based on the Global Database of Events, Language, and Tone (GDELT), for the implementation of this approach to ICEWS data. Building on the conclusions of this comparison, the third part uses Yonamine's (2013) approach to predict violent events in Afghanistan over a significantly longer period of time (January 1995 to August 2021). The conclusions provide an assessment of the utility of short-term localised forecasting.
Keywords：forecasting, time-series analysis, georeferencing, political violence, Afghanistan
This repository consists of the main code for replication of 'Lesson (un)Replicated' paper. It has four components. 'original_timeframe' holds the code for the timeframe equivalent to Yonamine's (2013) paper. '25_vs_75_percent', '50_vs_50_percent' and 'third_vs_two_thirds' hold the code to replicate the different cutting points for the full duration of the ICEWS dataset used for the analysis - January 1995 through August 2021.