The objective

The focus of this internship project was modelling and analysing large disaggregated panel data sets of supermarket transactions. The overall aim was to develop a holistic model of shopper behaviour, taking into account heterogeneity in the demographic profile, and past responses to pricing, promotion and marketing strategies. Using real-life data, the project sought to validate an agent based model of consumer choice within a particular product category (fruit juice) in an internet supermarket.

The approach

The project took an agent-based consumer choice model developed at Unilever and followed an iterative process of developing a simplified version of the model and implementing it as a computer simulation with the intention of applying validation techniques in order to assist the validation of the full model.

The final simplified model incorporated a weighted consumer choice function of net product price and product characteristics. Product characteristics refer to quantities or qualities of the product, such as brand or pack size, which are mapped to the interval [0,1] based on the volume of sales of categories within each characteristic. Each product was mapped to a point of the unit cube in a three-dimensional characteristics space.

Following the development of the computer simulation model, panel data of the sales of fruit juice in 2007 from an internet shop were used in the initialisation of agents’ characteristic preferences, calibration of parameters and subsequent testing of predictions out of sample. Out of sample tests gave Unilever greater confidence on the ability of the model to predict real life sales.

The advantage of agent-based models lies in their ability to model heterogeneity and emergence within populations. As such, validation needs to occur on multiple levels. Usually this happens at the macro/population and micro/individual levels. In this project, validation at the macro level involved examining overall market shares and at the micro level involved examining consumer choice in terms of product characteristics such as brand, flavour or carton volume, so enabling an examination of individual consumer preferences.

The current validation exercise involved statistical techniques comparing the outputs of the simulation for a three month period to the real data covering the same period. Related ideas on probability evolution are currently being explored in a partial differential equation framework as a means of comparison against the ABM model, with the hope of providing further validation.

related resources:

	Modelling and analysis of supermarket transactions
»	Technical summary

other projects:

[Find other Food and Drink projects]

[Find other Retail Project]