Measuring Demand Interdependencies by Neural Networks

Abstract: Recent research efforts about demand interdependencies (DI) in assortments of retail firms are based on a probabilistic model of purchases of article groups, a multivariate logit model (MLM), and assume symmetric interdependencies. Usuallythe MLM is estimated with asymmetric parameters and symmetry is constructed ex post. We present an artificial neural network (ANN) approach for measuring DI between article groups and show how the restriction of symmetry can be taken into consideration during the ANN estimation process. A likelihood ratio test serves to test the assumption of symmetry. Backpropagation (bp) is an appropriate estimation technique for ANN and ridge regression (rr) is used for the MLM to cope with multicollinearityin the data. We compare the results of ANN and MLM as far as (rr) produced reasonable results. The data base we use consists of 1669 observations of joint purchases of 72 article groups offered in 32 retail outlets.