A neural network approach for marketing strategies research and decision support
Hean Lee Poh · 1991
The primary objective of this thesis is to explore the efficacy of neural network methodology in capturing the relationship between a strategic goal such as market share, and strategic variables such as product quality, and competitive environmental variables such as the number of competitors. PIMS (Profit Impact of Market Strategy) database contains this kind of data. Traditionally, statistical methods such as regression have been used to analyze the underlying relationships using the PIMS data. This alternative method is called the neural network approach, which seeks to mimic brain functions, and has been applied in areas such as pattern recognition where the problems are ill-structured. Rules are implicit in the network. The neural network model adopted in this thesis is a network configured with one output unit (market share), 24 input units, and 10 hidden units between, interconnected in a feed-forward mode with no direct connection between output and input. The learning algorithm used is called backpropagation which seeks to minimize the total sum of squared error between target output and actual output from the network by propagating the error back and thus updating the weights over many passes of the training data sets of inputs and outputs through the network. The 994 cases of data used come from the PIMS database. The network is able to predict market share as well as a linear model using three-stage-least squares (3SLS). It is also able to test hypothesis of marketing strategies. The neural network model also captures the nonlinear relationship between the input and the output variables automatically, without having to specify the nonlinear terms to fit the data, as in the case of regression. In comparison to the linear regression model, the responsiveness of the neural network model to the change in the input differs from the linear model depending on the input variables and their values. The network can be reduced by using 3SLS to identify the most important input variables.