Improving Forecasts of International Events of Interest

Abstract

The paper compares the forecasting utility of the new GDELT - Global Data on Events, Location and Tone - dataset with some early versions of the ICEWS - Integrated Conflict Early Warning System - data using several alternative methods, including random forests, ADABoost, and Bayesian model averaging. Generally we find that the GDELT data performs as well or better than the data in the original ICEWS - quite possibly due to excessive attention in ICEWS to the eliminate of false positives, Kahneman's "what you see is all there is" pathology - and that these newer methods are quite promising as forecasting methods.