Agent-based models of market dynamics and consumer behaviour

Workshop Report

The first 'ConMod' workshop was successfully organised by the Centre for Research in Social Simulation at the University of Surrey on the 17th and 18th of January 2006. It attracted 46 participants from world wide destinations, including the US, South Korea and New Zealand. During the workshop 15 papers and 2 posters were presented, covering a broad range of empirical and theoretical issues around the theme of the agent based modelling of consumer markets. The workshop was sponsored by Unilever Corporate Research and organized by Nigel Gilbert, University of Surrey, UK and Iqbal Adjali, Unilever Corporate Research.

Program committee:

Guillaume Deffaunt (Cemagref, France)
Wander Jager (Groningen, the Netherlands)
Nigel Gilbert (Surrey)
Iqbal Adjali (Unilever)

Day 1

The four P's in social simulation
Wander Jager opened the workshop by presenting an account of how simulation could be used to explore the four ps of marketing: product, price, placement and promotion. He used this framework to show how simulation could benefit the understanding of marketing practices, presented a set of formalisations of each p and discussed some of the issues involved in this process.
Co-opetition for the diffusion of resistant innovation: A case study in the global wine industry using an agent based model
The second paper looked at the problem of resistant innovation adoption, focusing on the case study of screw top caps for wine bottles. Rosanna Garcia and Paul Rummels empirical research identified a difference in adoption between the US and Australian consumers, which they concluded was due to the differences in mutual marketing between co-operating wineries. They presented a Netlogo model of production, marketing and consumption, which they are using to explore how different co-opetition strategies affect uptake of resistant innovations.
Development and validation of a MAS simulation of consumers, retailers and manufacturers
David Midgley recounted his experience developing a model of consumer, retailer and manufacturer interactions based on theory and personal experience. He found that by trying to include too much in the model he had made it too complex, which led to problems of verifying complex code and validating complex models against available data. He concluded that this approach was too difficult and advocated starting simpler, for example producing models which aimed to capture 80% of reality.
Using empirical data to build an agent-bases model of innovation diffusion
Nina Schwarzs project examined the diffusion of innovation by modelling the adoption of water saving devices in the home, in Germany. This work combined empirical research, qualitative interviews and demographic data, with theory, Maslows pyramids of needs and Bourdieus concept of lifestyles. These are being formalised into a simulation.
Promotion strategies for the takeoff of diffusion
Deire and Jager are using simulation as an artificial laboratory to explore marketing strategies in a hypothetical market. The model uses social influence to represent two markets: fashionable with high social influence and utility with no social influence. They compared different marketing strategies , targeting individuals or cliques, and found that a mixed strategy was most effective. They also used this model to explore the outcomes of different timings of post product launch marketing campaigns.
Consumer heterogeneity evolving from social dynamics
Alex Frenzel used Social Comparison Theory to argue that consumers use comparisons with others to evaluate possible behaviours in uncertain conditions. From this he suggested that consumers will copy their social group in some consumption decisions, hence creating groups organised by product type. He used Latent Class Analysis to model and identify consumer groups from sales data for shoes in Germany from 1980 to 1991.
Modelling market competition
To conclude the first day, Ian Wilkinson presented a personal history of his involvement in the marketing and simulation fields. He reviewed historical approaches including Chaos, Complexity and the N-K models of Stuart Kauffman and argued that market systems need to be studied as Complex Adaptive Systems, the approach developed by the Santa Fe Institute.

Day 2

Propagation effects of filtering incongruent information
Day two started with Guillaume Deffuant presenting a simple model, based on abstract theory, which looks at the propagation effects of filtering incongruent information. The model assumes individuals filter information that is incongruent to their global perspective and are reluctant to talk about any such information. Analysis of this model found surprising results, where the rational filtering of information led to non-rational final attitudes, for example an individual with a positive attitude about an object, can end with a negative one, even though the object is globally neutral. This paper also promoted the simulation methodology of developing simple abstract models from theory, with complexity then added once the simple model has been developed and understood.
Social interaction and low involvement products
In the second paper of the day Johanna Kuenzel presented her empirical research which aimed to assess the level of social influence, if any, on purchase decisions for low involvement products, in this case everyday food products. She presented the initial results from a survey which found that not all food products are perceived as low involvement and that consumers perceive significantly different involvement between different food types. Preliminary analysis suggests there is some social interaction effects on some of these low involvement products.
The impact of quality uncertainty without asymmetric information
This paper, by Segismundo Izquierdo, presented a model of a market with asymmetric information about quality. That is, where the seller is better informed about the quality of a product than the buyer and uses this to sell lower quality goods at the same price as higher quality ones, for example the used car market. In the model consumers decision to buy is based on their perspective of the quality and value of goods in the market which is developed through individual experience and the experience of their peers. The model found that when quality is assessed only on consumer experience the market collapsed, but that social networks helped maintain consumer confidence.
Agent-based simulation of consumer behaviour in grocery shopping on a regional level
Tilman Schenk presented a model of supermarket choice which used a GIS interface to visualise the included spatial dimensions. The model was built from a variety of micro and macro level empirical data. It models agents spending across grocery stores based on their geographical and socio-demographic settings. The model is calibrated against revenue data and validated by a goodness of fit analysis against actual shop revenues.
Agent based simulation of fashion shopping in the functional region of Regensburg
Daniel Schrodels paper followed a similar approach, but looked at the revenue of fashion shops. The model was built from empirical data, macro level data on socio-demographics and micro-level data from interviews. The model was calibrated and validated against revenue data and found that price, offers, display, floor space and distance were the most important variables in explaining shop choice.
Simulating dynamically changing consumer preferences
This paper, by Andras Vag, presented a framework for combining multi-agent simulations with the data collecting technique, Conjoint Analysis, to model the dynamics of consumer preferences. He suggested that Conjoint survey data can be used to set up initial consumer preferences, which can then be updated over simulated time to model dynamic preferences. He proposed that the simulation could incorporate social aspects, product experience and marketing variables.
An agent based simulation of consumer purchase decision making
This first part of Tao and David Zhangs paper developed a dynamic model of consumer behaviour derived from the static model of Engel et al. In the model agents have a set of preferences decreed by their socio-demographic details, which are updated by social interaction and brand management changes. The second section aimed to use the model to demonstrate and explain the empirically observed phenomenon of the decoy effect.
Socio-dynamic discrete choice on networks in space
Elenna Dugundji and Laszlo Gulyas reported on their model of transport decisions for commuters in Amsterdam. They develop a Discrete Choice analysis by including social influence effects modelled with a network. The network is constructed from macro level demographic data, where contacts are assumed to exist based on social and spatial variables. Each agents choice is then additionally influenced by the choices of its neighbours, resulting in various stable proportions of transport choice.

Themes

The workshop covered many important themes pertinent to the application of agent based modelling to the study of markets and consumer behaviour. The main themes identified are the different methodological approaches to developing an agent based model, which can be broadly categorised into theory, empirical and mixed.

Empirical Based: Within the empirical based models there was a wide variety of data types utilised in many imaginative ways. There was the direct linking of Conjoint data into agents preferences, converting socio-demographic data into group- and individual-level variables, using spatial data and GIS maps to visualise the model and analysing quantitative interview data to gain insight into actor behaviour, which can be transferred to agents. Other important work focused on analysing the real world for potential application within agent based models, which reminded us of the importance of validating presuppositions, for example that there are indeed social influences even on milk buying.
Theoretical Based: The theoretical models presented a way to formalise theory and explore their subsequent complexity. Some speakers used the models as artificial laboratories, where different markets could be modelled, different strategies tested and different phenomena analysed. This formalisation process is an interesting one where we need to be careful: many papers seemed to assume they have achieved the only one to one mapping of the world, whilst the formalisation process is inherently interpretative and so a model needs to be seen as one of a range of possible formalisations. Cross model validation can then compare differences, if any, between formalisations.
Complex v. Simple: When building the model the issue of simple or complex models came up repeatedly. The consensus seemed to be in favour of simple models where additional complexity can be added after analysis and understanding. However this issue is less straightforward with models with empirical targets, as these do not have the option of starting simple, because they will then not capture the real world properties that are aimed at. The solution seems to be to developi mechanisms and rules within a simple theoretical model which can then be implemented in an empirical one. Other suggestions were to start aiming for 80% accuracy, although that seems quite high and which 80% do you include?
Validation and verification: Validation and verification issues were a small part of proceedings, mainly because most models were prototypes and so had not reached that stage. These remain important unresolved issues within the field in terms of how a non-linear stochastic model can be validated. However it was good to see that the development of most models were firmly linked to data and theory, which helps strengthen the argument that they can be applied to the target system.

Next steps

At the end of the workshop, there was a great deal of enthusiasm for continuing the dialogue that had been established between the participants, many of whom had not previously been in contact. Plans were made to:

publish selected papers from the workshop in the Journal of Business Research. The Programme Committee would act as guest editors for a special issue, which is likely to be published in Spring 2007.
set up a wiki page to exchange contact details and research interests.
plan another, follow-up workshop in a year's time.

Alan Roach and Nigel Gilbert
26 January 2006, revised 5 February 2006