PROJECT TITLE: IMIS - Intelligent Marketing Integrated System
PROGRAMME NAME: ISIP
PROJECT NUMBER: IED/4/1/8022
Participants: Project Manager:
Michael Campbell
Henley Centre for Forecasting Ltd
2-4 Tudor Street
London EC4Y DAA

Background and Objectives

Conventional marketing modelling techniques derive from economics and econometrics and assume that consumers behave in line with the characteristics of economic consumer choice theory. In many markets for fast-moving consumer goods, these characteristics are not descriptive and, in particular, do not help firms to identify which alternative brands are competing with one another and what are the important aspects of the competition. The prime objective of the project was to integrate linear and non-linear statistical techniques with more qualitative rulebased approaches in order to identify competitive sets in markets where there are many products and brands and where reliable data is limited.

Progress

The MMU team have built several demonstrator models which incorporate linear statistical methods provided by Henley Centre. These models have been applied to data from supermarkets provided by Distillers. The statistical analyses were supplemented by new, non-economics approaches to consumer choice analysis as well as qualitative judgements about the nature of different demands and such product characteristics as `specialness' and `uniqueness'. The models were implemented in SDML — a strictly declarative, agent-based modelling language which runs on top of Smalltalk-80.

The Henley Centre has developed the techniques for non-linear analysis based on local regression. One of our priorities is now to integrate the local regression into the MMU software in order to identify important non-linearities in the data and to exploit the non-linearities in the determination of the competitive sets of individual brands or financial products.

Interesting results and ideas

We have found that our models yield more accurate simulations of market shares than do conventional statistical models estimated by Ordinary Least Squares and Maximum Likelihood methods. In generating these simulations, our models are much more parsimonious in their data requirements but computationally much more expensive.

Our results entail some methodological advances for qualitative modelling in that we have had to devise procedures for estimating qualitatively valued parameters using mainly genetic algorithms on numerical mappings from the qualitative values. "Qualitative modelling" in economics and management sciences refers only to numerically valued partial orderings and consists of relative sizes (up, down, larger and smaller).

Perhaps the most important aspect of our results from the point of view of the marketing professionals is the facility to develop scenarios based on qualitative changes in the markets starting from a base of qualitative judgements already shown to be consistent with reliably accurate numerical data. These scenarios can entail intelligent learning and response strategies by simulated agents.

See the papers: "A Formal Preference-State Model with Qualitative Market Judgements" and "Modelling Learning as Modelling" for some more detail.


Monitoring officer: Michael Falla

Back to Centre for Policy Modelling