This paper considers the problem of cross-sectional aggregation when the un
derlying micro behavioural relations are characterized by general non-linea
r specifications. It focuses on forecasting the aggregates, and shows how a
n optimal aggregate model can be derived by minimizing the mean squared pre
diction errors conditional on the aggregate information. The paper also der
ives model selection criteria for distinguishing between aggregate and disa
ggregate models when the primary object of the analysis is forecasting the
aggregates, and establishes the consistency of the model selection criteria
in large samples. In the case of standard non-linear micro relations with
additive errors it also provides suitable small sample corrections. For mor
e general non-linear specifications we consider bootstrap techniques to cor
rect for small sample bias of the proposed model selection criteria. Some o
f the ideas in the paper are illustrated using log-linear micro relations,
often employed in applied research. The paper also contains an empirical ap
plication where log-linear production functions are estimated for the UK ec
onomy disaggregated by eight industrial sectors and at the aggregate level
over the period 1954-1995, (C) 2000 Elsevier Science S.A. All rights reserv
ed.