A comparison of the forecasting performance of a constructed monetary index with component data using neural networks

Accuracy in the measurement of money is important for economists. The conventional method is to simply sum the various constituent liquid liabilities of banks. This method of arriving at broad money aggregates is seriously flawed and an important alternative is the Divisia weighted index, in which the components are weighted on the basis of the monetary services provided by each. However while there is much evidence in favour of the Divisia index, governments largely continue to work with the 'simple sum' index. This study uses neural networks to demonstrate the superiority, for inflation forecasting purposes, of working directly with the component assets as a first step towards optimising U.K. index construction. The possibility is demonstrated of constructing a new, weighted index based on weights empirically derived from neural network models. An experimental index is shown to have a lower forecasting error than the Divisia index published by the Bank of England.