Predictions are good and predicting specific conflicts using analysis of particular motives and conditions empirically succeeds

Our argument is that simply identifying inappropriate methods as the key source of the problem in forecasting conflict may give us the wrong diagnosis and lead us down less productive avenues. A more fundamental problem is models that provide a poor basis for forecasting by disregarding the motives for conflict to arise, or by only considering motives in a relatively superficial manner. Models that have been proposed for research on the democratic peace, notably the work of Russett & Oneal (2001), are primarily intended to examine whether certain characteristics of liberal institutions, such as democracy and trade, make conflict on average less likely relative to baseline risks of conflict. Although these approaches may be appropriate for testing the original propositions of interest, they essentially ignore the contentious issues that might cause states to resort to violence and instead treat these contentious issues as exogenous features, typically hidden inside a so-called ‘black box’ of the baseline risk of conflict. Our own initial foray into out-of-sample prediction for a state-level model indicates that spatial information about other conflict events can help to improve forecasts (see Ward & Gleditsch, 2002). Although this allows predictions of conflict to be conditional on other observed events rather than treating each conflict as an independent observation, the approach still ignores the issues over which such conflicts may have arisen initially.