Using a Neural Network to Estimate Solvent Consumption

摘要

The paper discusses a neural network, using the backpropagation paradigm, that is taught the relationship between employment in the graphic arts industry--(Standard Industrial Classification Code (SIC) 27)--and economic variables and solvent consumption by SIC 27. The project is a proof of concept whose objective is to a relationship using national-level data, which are known, and apply it to estimating solvent consumption on the county level, where data are thus far not available. The network accurately learns a relationship from national data. Although definitive testing is not yet possible due to data limitations, there are indications that the national relationship can be used to estimate county-level solvent consumption. Network inputs are SIC 27 employment, productivity for the current and 1 prior year, and an eight-element 'signature' of quarterly economic changes in output from non-durable industries. One hidden layer of two processing elements connects the 11-element input layer to a 1-element output layer. NeuralWare Professional II Plus Version 4.0 was used as the platform. Training requires 30,000 iterations and results in a Pearson's r value of 0.99. The best result achieved by ordinary least squares regression was 0.93.