Abstract
An agent-based behavioral model incorporating utility-based rational choice enhanced with psychological drivers is presented to study a typical consumer market. The psychological drivers incorporate purchase strategies of loyalty and change-of-pace, using agent-specific memory of past purchases. Attribute-specific preferences and prices drive the utility-based choice function. Transactions data is used to calibrate and test the model. Results indicate that prediction accuracy at both macro and micro levels can be significantly improved with the incorporation of purchase strategies. Moreover, increased agent memory does not improve predictions in the model beyond a threshold, indicating that consumer memory of past shopping instances is finite and recent purchase history is more relevant to current decision making than the distant past. The chapter illustrates the use of agent-based simulations to model changes or interventions in the market, such as new product introductions, for which no history exists.
TopIntroduction
Consumer behavior as a field of study is highly interdisciplinary in its approach, and that is evident in the amount of literature on this topic in multiple fields of study – whether in economics, psychology, sociology, computer science or even applied mathematics. Correspondingly, the traditional methods of analysis used by researchers in this field are numerous, and they range from quantitative (statistical and regression based) to qualitative (surveys, interviews, ethnographic studies etc.). However, over the last few years, new studies are being increasingly seen in the literature, which use modern computational techniques based on computer simulations, data mining, big data analysis etc., which mirror the changes and technological progress in societies and markets the world over. This chapter introduces one such method – agent based simulations, and links real world empirical data with models of behavior from multiple disciplines. Additionally, the chapter also provides an example of how such models can be used to “explore the future”, with their ability to incorporate “what-if” scenario building techniques. This chapter is based on Sengupta and Glavin (2013) and Sengupta and Glavin (2010), and introduces the models and methods used in both, and extends them by illustrating how radical changes in the market (such as new product introductions) can be modeled robustly using computational methods.
Markets often exhibit noisy dynamics in the form of volatile movements in market shares (Jager, 2007). Frequent competitive interventions by manufacturers, such as introduction of new products, aggressive marketing policies such as multiple pricing and promotion strategies – is definitely one reason behind this widespread phenomenon (Ailawadi et. al., 2001; Blattenberg & Wisniewski, 1989). However, the presence of a wide variation in tastes and preferences amongst a reasonably large and demographically varied consumer population is also a key factor leading to the noisy character (Allenby & Rossi, 1998; Sengupta & Glavin, 2010; Sengupta & Glavin, 2012). Such markets do not lend themselves easily to traditional statistical and econometric analysis. Nor do markets where major interventions or events have occurred in the immediate past, which have moved these markets “out of equilibrium” (Reid & Brentani, 2004; Mathews, 2006), such as new product introductions, innovations etc. Additionally, the presence of potential non-linear interactions such as social networks, word-of-mouth influences etc. means that they may also exhibit a “complex” character – hence making traditional techniques further redundant. Not surprisingly, markets in general and consumer packaged goods (CPG) markets in particular, are increasingly being brought under the purview of “complex systems” analysis – whereby more modern “bottom up” methodologies such as agent based modelling are being used for analysis, inference and predictions (Gilbert et. al., 2007).
Systems which exhibit “emergent behavior” of some kind cannot be fully examined and analyzed by traditional “top-down” methodologies. Simulation based techniques – relying on agent based constructs – where constituents of the system (in this case, shoppers, firms etc.) are treated as individual modelling units (or agents) with the ability to follow independent rules of behavior and engagement have become increasingly popular and are widely advocated (Epstein & Axtell, 1996; Gilbert & Troitzsch, 2005; Tesfatsion, 2006). CPG markets have been extensively studied in the mainstream literature, but in spite of exhibiting many characteristics of a complex system, have only recently been brought within the purview of complex systems analysis (North et. al., 2009; Sengupta & Glavin, 2010; Rand & Rust, 2011). This paper builds a behavioral model of consumer choice, which is then incorporated within a multi-agent simulation framework to illustrate the accuracy and usefulness of such an approach in predicting market phenomena. It builds on earlier simpler models by the same authors, by incorporating crucial psychological factors into the choice model, modifying and extending the validation methodology and finally showing that both market and individual level predictions can be significantly improved by using enhanced choice models.
Key Terms in this Chapter
Agent Based Modelling: Refers to a computer simulation methodology which is commonly used for modeling complex social systems. The system is constructed bottom up , that is, individual constituent units of the system are programmed as independent autonomous units (agents), following simple to complicated behavioral rules, and with the ability to interact with their environment and other agents in the system. The models are seeded with an initial set of values for its parameters, and the simulation is set to run for a given length of time. Behavior of individual agents as well as the system as a whole is recorded and analyzed once the simulation has run its course.
Testing: Involves using the calibrated simulation to make predictions on a particular time period (or cross section) from which the data has not been used for calibrations. These predictions are then compared with the corresponding data to ensure the goodness of fit of simulated data with real data. Note that the data being used here is not the same as the one used for calibration. The testing exercise provides further validity of the models and ensures a degree of robustness for practical use.
Calibration: The process of exploring the right range of values for parameters in the model, which ensures realistic and accurate behavior at the micro and macro levels in any computer simulation. Involves re-running simulations with varying parameter values to see which provide the best match to real data.
Initialization: The values of parameters that any simulation experiment starts with. These parameters could be agent specific or global.
Model Validation: Validation in the context of agent based modeling refers to the process of ensuring the behavior exhibited by a simulated system matches the real target system which it is trying to mimic. This involves a systematic exploration and calibration of model parameters, ensuring robustness of the model, eliminating logical errors in the model code and finally, ensuring that the model replicates the target system with a high degree of accuracy both at the macro level and at the micro level.
Complex System: A system which is usually composed of large number of possibly heterogeneous interacting agents, which are seen to exhibit emergent behavior. Emergence implies that system level behavior (macro level) cannot be inferred from observation of individual level behavior of its constituents (micro level). This absence of explicit links between the micro and macro levels makes complex systems especially difficult to analyze using traditional statistical and analytical techniques to study the dynamics of behavior. One typically requires the use of bottom up simulation based methods to study such systems. Complex systems are ubiquitous – markets, societies, social networks, the Internet, weather, ecosystems, are just a few examples.
Agent Memory: In the context of agent based models, this refers to the far back any individual agent in the model is able to recall past actions, behavior, parameter values etc., which may be relevant for its current state. The memory could refer to its own variables or variables of other agents or global and environmental variables, and this depends on the context and nature of the model being considered.