How to solve large recursive-dynamic Computable General Equilibrium models

We discuss approaches to solve and speed up baseline generation with recursive-dynamic global Computable General Equilibrium (CGE), using a set-up extending the GTAP standard model (Corong et al. 2017) in GAMS (van der Mensbrugghe 2018) with modules such as GTAP-AEZ or carbon taxation with endogenous process emission factors. G-RDEM (Britz and Roson 2019) depicts structural change dynamics based on a MAIDADS demand system, sector specific productivity changes, endogenous saving rates, debt accumulation from foreign savings and cost and expenditure shares depending on per capita income. The underlying data base splits the GTAP-Power Data Base Version 11 to 103 products and 143 sectors based on FAOSTAT data to provide detail on the agri-food system including irrigated and rainfed crops. Besides real GDP, crop land expansion, crop yields and calorie intakes are exogenous and swapped with productivity and preference shifters. Baselines are constructed until 2050 under projections of GDP and demographics for the Socio-Economic Pathways (SSPs) with matching scenario assumptions, testing (1) using last year’s solution as the starting point, (2) using last year’s solution plus update factors derived from GDP, population and past endowment changes, (3) adding previous year’s changes to last year’s results, with a recursive definition of start values of derived variables, (4) add previous year’s changes relative to last and current GDP change to last year’s solution, again with the recursive definition. All approaches are additionally tested with pre-solves of single country models derived from the global one. Moreover, we discuss further options to speed up solution and decrease memory needs more generally in GAMS based equilibrium models. The thirteen data bases used for the tests jointly cover Europe and North Africa at country level, and where available, at sub-national administrative regions, complemented by a smaller test set with nine global regions. Using updates factors based on projected changes in GDP, capital and labor stocks and past other endowment and shifter changes performs best, an option easily added to code copying last results to the current period.