An Empirical Test of New Forecasting Methods Derived from a Theory of Intelligence: The Prediction of Conflict in Latin America

Abstract

The "compromise" method is a new computer-based forecasting tool, available within the conversational CS package on the MIT Multics. Like regression (least squares) or new forms of Box-Jenkins methods, it estimates the parameters of a multivariate dynamic model and may be used for causal analysis or policy impact analysis. Unlike those maximum-likelihood methods, it does not assume that errors are "white noise," random and normal. It follows the newer robust philosophy of trying to minimize estimation errors on the assumption that noise will be inextricably dirty. In the case of "strong" dynamic models¿models which predict that changes in present variable values lead to comparable changes in future variable values it may reduce parameter errors by an order of magnitude. Forecasting errors will also be reduced, although the degree of reduction depends on how much randomness exists in the process. When we used the compromise method according to the new "bias" procedure, in order to reestimate the J-5 model (a nonlinear multiequation model used by the Department of Defense in long-range forecasting), forecasting errors were reduced by between 0 and 45 percent (with a median of about 20 percent) across different variables, as compared with regression. With simultaneous-equation econometric models, it has reduced them by 50 percent. The procedure has been documented for use by nonprogrammers [1]; it incorporates a new quasi-Newtonian method which can handle many parameters.