Balancing International Financial Data: Fun And Games With Gams And Cross-Entropy

In balancing international financial data (Lane and Milesi-Ferretti, 2006), contrary to most matrix adjustment problems, the only constraints are that, for any given category of securities, the sum of assets over countries must equal the sum of liabilities. Obviously, these constraints determine a solution only up to a factor of proportionality. This paper examines various ways of dealing with this problem, based on the minimum information-gain principle (minimum cross-entropy), as generalized to accommodate negative entries in the data (Junius and Oosterhaven, 2003; Lemelin, 2010). It is shown that : (i) the errors-in-variables approach adapted from Robinson et al. (2001) seems inappropriate in the present case; (ii) the use of a weighted loss function to take into account information that might be in Cumulative errors and omissions does not make much difference; (iii) the naive approach yields results that are very close to the basic model. In addition, it was observed that the choice of initial values is critical for the MINOS 5.5 GAMS solver, and, more importantly, that the default tolerance levels were too high for the problem at hand.