.jpg)
Several existing forecasting systems attempt to predict various forms of political violence, like conflict fatalities, in the near future. Researchers and end-users of such forecasts prefer ‘risky’, over-predicting forecasts rather than careful, under-predicting ones. Within this community many believe that selecting forecasts using the mean-squared error on the log-scale (MSLE) rather than on the original scale (MSE) will lead to the selection of a riskier forecast. In this talk, we study this ‘conventional wisdom’. We discuss different interpretations and investigate their validity, while placing particular emphasis on data distributions which are typical in the political violence context – with many zeroes, and a long tail.
This talk is based on joint work with Gudmund Hermansen, Håvard Hegre and Michael Colaresi.