Coherent optimal prediction with large nonlinear systems: an example based on a French model

The drawbacks of predictors obtained with the usual deterministic solution methods in nonlinear systems of stochastic equations have been widely investigated in the literature. Most of the proposed therapies are based on some estimation of the conditional mean of the endogenous variables in the forecast period. This however provides a solution to the problem which does not respect the internal coherency of the model, and in particular does not satisfy nonlinear identities. At the same time, for analogy with univariate skewed distributions, the conditional mean may be expected to lie on the wrong side of the deterministic solution, meaning that it moves towards values of the variables which are less likely to occur, rather than towards the most probable values. Estimation of the most likely joint value of all endogenous variables is proposed as an alternative optimal predictor. Experimentation is performed on a large scale macroeconomic model of the French economy, and some considerations are drawn from the results.