Marketing

A Set of Applicable Neural Network Models

By Assistant Professor Pasi H. Halmari

Department of Marketing and Corporate Geography

Swedish School of Economics and Business Administration

Helsinki, Finland

PROJECT SUMMARY

Neurocomputing has rapidly gained in importance in a multitude of different application areas. Neurocomputing does not require algorithm or rule development, a feature that often significantly reduces the quantity of software that must be developed. As a consequence, the last decade has seen an uncharacteristically rapid growth of applications in virtually all disciplines, including sensor processing, pattern recognition, data analysis, and control. Since we focus on model building and applications in business and in the social sciences, pertinent applications of neural network modeling include investment and portfolio management, commodity and stock trading, forecasting, logistics management, image-based recognition of hand-written and fax messages, adaptive control problems, mailing list development for specific markets, seat allocation on airlines, loan application evaluation, quality control, and facility.

We will attempt to build/expand three model families with a clear anchor in business and in the social sciences. All three models are expected to serve several purposes ranging from the direct applicability of each individual model to their diverse theoretical contribution. Multilayer perceptron (MLP) algorithm is used in all three models to train the neural networks models.

Model Cluster 1 is an expansion of a existing (prototype) demand forecasting model. Here, we started by building an aggregate model of the demand for cars in Finland. The model is based largely on macro-economic data. We trained the model with four years of data: background variables and outcome information. Our initial testing is very encouraging. We intend to expand into alternative model superstructures, into forecasting other forms of demand, and into the largely uncharted territory of regional demand. The models concerned with regional demand, interestingly, may become an alternative and/or a complement to the theoretically richer diffusion models outline here as Model 3.

Model Cluster 2 is based on data from a set of Canadian shopping malls. Despite the predominance of the mall in North American commerce and the mall's ever greater appeal in other parts of the world as well, very little research work has focused on this important economic/social phenomenon. Ultimately, we are interested in understanding this economic microcosm as a full dynamic system. The mall phenomenon to a large extent is built on a set of classic economic and market terms: agglomeration economies, economies of scale, product mix, etc. Yet, attempts to optimize these systems --even on the most basic and uni-dimensional level of profit maximization-- tend to be based on qualitative models and judgments. The mall is a microcosm where everything seems to have a potential impact on everything else. We intend to build two initial submodels based on our shopping mall data.

Model Cluster 3 again combines potential direct applicability with a strong theory- and method-building component. We intend to build a set of neural network models of a few key diffusion processes. It is well known that market penetration is a slow and sometimes unpredictable process.

We know that the spread of innovations frequently start in a small number of central places, sometimes even in just one. From these centra, the product then works its way into an expanding hinterland. We know quite a bit about product lifecycles in general --mostly expressed on a temporal axis only-- but can today only speculate about the characteristics of product/idea lifecycles in a time-space continuum. At the same time as the product/idea enters an acceleration phase at the center, it may just start to penetrate outlying markets. Likewise, long after the market at the center is saturated, the outlying markets may still be in a growth phase. We intend to build separate models of the diffusion of a small number of traceable products. We are particularly interested in the diffusion of products like the mobile phone. First, the societal and economic impact of the innovation is vast, second the nature of the product makes it more traceable than other products. We are interested in training neural networks to predict the acceptance rates of the innovations through time and space. Whence a small set of models covering different products has been accumulated, initial hypotheses may be possible concerning more comprehensive characteristics of innovation diffusion.

To date, the theoretical advances on diffusion theory within geography, marketing and other social sciences, originate in Hägerstrand's (1953, 1968) classic studies. Our models will break with the traditional Monte Carlo based diffusion modeling framework. In doing this, we hope to build models that are far less mechanistic and more responsive to the specific environment being modeled.

Publications and presentations:

Halmari, P.H., Ivars, C. and Rosenbröijer, C-J (1992), Using Neural Network to Forecast Automobile Sales in Vogt, W.G and Mickle, M.H. the Twenty-Third Annual Pittsburgh Conference on Modelling and Simulation, University of Pittsburgh, USA

'The 4th International Conference on Recent Advances in retailing and Services Science', June 30 - July 3, 1997, Scottsdale, Arizona, USA. A paper 'Optimizing Tenant Mix and Tenant Location in a Shopping Mall - Applying Geographical Information Systems and Neural Networks' submitted for presentation.

The 17th Annual Applied Geography Conference', October 12 - 14, 1994, Kent, OH, USA. 'Application of Neural Networks in Shopping Center Tenant Mix Optimization'

'Location, Location, Location, - Advanced Retail Location Techniques Seminar' May 31, 1994, Toronto Canada. 'Using Neural Networks in Decision Support Systems'

'The Twenty-Third Annual Pittsburgh Conference on Modeling and Simulation 'April 30 -May 01, 1992, Pittsburgh, PA, USA. Halmari, P., Ivars, C. and Rosenbröijer C.-J.: 'Training Neural Network to Forecast Automobile Sales'.

'Second European Conference on Geographical Information Systems', April 02 - 05, 1991, Brussels, Belgium. Gustav Lundberg, G. and Halmari, P.,: 'Computer Aided Logistics Management: Demand Forecasting with Neural Nets'.

'Annual Meeting of Association of American Geographers' April 13 - 17, 1991, Miami, FL, USA. Halmari, P. and Lundberg, G.: 'Bridging Inter- and Intra-Corporate Information with Neural Nets'.

Doctoral thesis expected in fall 1997 'Tenant Location in a Shopping Mall - Applying Geographical Information Systems and Neural Networks'

Download this Application Summary in Word97 format.
If you don't have Word97....Download the Word Viewer.