We are in the concluding part of our trade promotion optimization (TPO) series. In our last two blogs, we have explored the industry mature discrete choice model and the concept of building a simulation framework for agent-based modeling. In this blog, we will deep-dive into some design principles of a multi-agent system and how we can use the framework to evaluate and respond to changes in a consumer marketplace.

Simulating a multi-agent system 

There are a finite set of interactions in a TPO world, such as buying and selling between consumers and retailers, negotiating and fixing promotion (promo) calendars between manufacturers and retailers. Each agent in a simulation environment can be described by a set of characteristics that represents how an agent is modeled. There are fixed attributes that remain unchanged during a simulation exercise.  Packsize and the range of margin percentage to be maintained for the duration of the promo are some of the examples of fixed attributes. Variable attributes on the other hand may vary during the process of a simulation exercise, like the duration of a specific promo and the promo offered to a retailer. To add to the set of characteristics, there are behavioural attributes as well that must be considered. These represent actions emerging out of various interactions that these agents have with each other such as acceptance of a specific promo plan.

The figure above illustrates the attributes of a retailer and a manufacturer in a multi-agent system. Prices, promos, and duration can be varied based on business KPIs of a manufacturer. During negotiations with a retailer, margin% and household penetration (HH) act as constraint-based variables that need to be maintained by a manufacturer. An acceptance of a specific promo calendar is considered as a behavioral trait governing a manufacturer and a retailer relationship.

Utility of a product

Before we take a closer look at the agent-based model (ABM) system design, it is crucial to understand the underlying concept of utility. In modeling a consumer choice behaviour, we may follow the classical approach of defining a utility of a product to a consumer as the behavioural propensity of a consumer to choose a particular product of interest that would eventually lead to a sale in a marketplace. For a product that is defined by a bundle of attributes such as brand, packsize, price, day of week, health index, and display feature at a store, we can build a utility function using the above-mentioned attributes along with a preference quotient associated with each of them to be mapped against a region or a segment during a time frame. Following this, we can then derive a probabilistic choice for a specific product, based on the calculated utility for a specific region or at a segment level.

ABM system design

Let us consider a typical simulation setting, where for a region, we take a sample set of mathematically derived agents. We simulate a marketplace with 50,000 consumers agents, 2500-3000 retailer agents, and 200-300 manufacturer agents. Retailers and manufacturers represent the B2B dynamics and the consumer agents represent the dynamics of product movement and purchase behavior in the marketplace of a particular geographic region under study. This allows us to tie demographics to consumption, which in turn enables us to model our choice model and adjust the KPIs for both types of agents. Sales are simulated, based on how the consumer marketplace reacts to offered promo packs across time. For all other information on consumer data, we use past syndicated data to build probabilistic relationships between various psychographic and demographic variables leading to a purchase decision.

With recent advances in technology and computational power, we are now capable of building network-oriented causal structures to determine consumer choices and identify purchase behaviours. Multiple state-of-the-art techniques such as Bayesian networks, Markov decision processes, posterior probability or conditional probability tables can be leveraged in replicating the consumer choice that contributes towards retail sales. This purchase decision model is derived from various socio-demographic features that define the consumer and identify personality traits triggered by consumer agents interacting with the marketplace and retailer agents, and indirectly with the manufacturer, that ultimately leads to a purchase decision or sales.

Given this construct, business leaders are keen to study the emergent behaviour among retailers, who show propensity towards specific promo calendars. Retailer inclination is often influenced by the margin earned and sales of a product. In our sample case study of 50k consumer agents, we must define an average basket or transaction value, household size, income – to be derived through prior probability distributions, gender, ethnicity, segment, and age to build a model that is in proportion with geographic distribution. After validating the framework on how accurately we can simulate this environment and measure actual vs simulated sales per retailer, we can then use this model for studying the market dynamics.

The way we are evolving the validation framework

Learning from enterprise-level production work, Enquero (A Genpact Company), has been designing and incorporating a new stage in the model validation framework – guardrails. We have often observed that agents in a simulated environment pick up irrational behaviours such as transmitting sudden spikes in consumption moments in offices during weekends even though offices are closed during that time. Such erratic behaviours shouldn’t ideally be simulated. Enabling a detailed behavioural constraint framework to feed in agent restrictions (in the above example of excluding office consumption during weekends) assures the simulated behavior to be as close to actuals as possible.

In conclusion

With the improvements in technologies, there are a varied array of tools that we can leverage to build ABM-oriented products in an enterprise. From traditional tools like NetLogo to enabling complex system analysis with Mesa, building advanced simulation techniques have become very user friendly in the last few years. Apart from building a business case around optimizing topline KPIs on margin, revenue and penetration that is apt from a manufacturer-retailer perspective, the simulation framework also enables in understanding changing market behaviours and evaluating impacts of new product launches.

Designing systems ground-up with evolving dynamics in mind, continues to leave businesses with the following key questions and it is crucial that we strive to find those answers today:

1. What if there is a rise in a specific demographic of students by 2% on the beverage category movement?
2. What if there are whitespaces in the healthy snack category price ladder? Will it be possible to control x% of market shares if a coconut flavoured snack is introduced?
3. What if there is I run a BOGO offer for the summer? Will this improve the household penetration by x% within the occasional customer segments?
4. What if we introduce an offer of wider assortment of canned fish cat food in a 12 pack bundle? Does this improve the profit margin by 2% in the wet food category?

The what-ifs are endless… but with a nimble agent-based model, enterprises are in a better position to make clever decisions in the B2B marketplace with learnings from consumer dynamics.