General-to-specific approaches for evaluating multi-step system forecasts

Multi-horizon forecasts from large scale models play important roles in ongoing policy debates across a wide range of disciplines. Although there are many ways to evaluate imperfect forecasts, we show the importance of taking a more general approach and considering the joint distributions over variables and horizons. Once a multi step system context is envisaged, potential singularities need to be addressed, as do tests of significant differences between the forecasts and assessments of the value of improved forecasts. These important issues are the focus of the thesis and are illustrated using applications in the fields of macroeconomics and climate.