In our previous blog, we explored some key ideas behind trade promotion optimisation (TPO) and how a vanilla framework is not enough to define the future trajectory of a business in this highly volatile consumer market.

The part 2 of this series, attempts to provide a glimpse of a ‘what-if’ simulation. At a microscopic level, you start to realise that the crux of TPO in B2B is in understanding consumer behaviour across segments and regions, which ultimately define sales patterns for a specific retailer. This consumer market isn’t rule-based and it comes with its own set of constraints and dynamics. An effective TPO strategy is a direct outcome of how symbiotic a manufacturer-retailer relationship is. Apart from modeling consumer behaviour, it is also crucial to map competitor interactions with retailers, which in turn will help you build the final promotional (promo) calendar with necessary action plans to implement retail partner strategies. There is no denying that this entire ecosystem of manufacturer-retailer-consumer dynamics is a complex system.

One of the computational methods to model and understand such a complex system is through agent-based models (ABM). Let us familiarize with 3 important terminologies for easy consumption of ABM.

Agent – Someone acting on your behalf is called an agent. For example, a piece of code, behaving as a trading manager, negotiating a trade promo offer with a store manager becomes an agent. The store manager is also an agent acting on behalf of herself or himself, with similar characteristics as observed within an environment. In addition, we can also model consumers as agents that define purchase patterns. These agents can be seen directly or indirectly interacting with each other at some point in the manufacturer-retailer-consumer journey. In our case, we observe three different kinds of agents – manufacturer, retailer, and consumers. Such a system is known as a multi-agent system.

Environment – The surrounding, where agents live or interact, is called an environment. Consumer marketplace is an example of an environment. This includes consumers’ perception of the value of a commodity (utility function of a product or goods) and their interactions with retailers, while weighing in competitor offerings along with indirect interactions with manufacturers.

Behaviour – An agent perceives certain information from the environment and associates (causal) action(s). This action can either be rule-based, causal, or constraint imposed. For example, there can be situations where a manufacturer agent with an established behaviour of not being able to offer a promo beyond a certain price point, or a retailer agent not willing to buy a promo, if past category movement has been less than a specific threshold percentage. Being able to model these assumptions, rules, and constraints can help us optimize our promotion strategy for better market penetration.

Factoring in real world scenarios

Here are some situations that can benefit from an agent-based simulation for effectively modeling bottom-up and addresses the gap of traditional discrete choice models:

Time sensitivity – Promo negotiations are driven by a sense of urgency. A delayed call or a rework on an existing promo can cause adoption lag for weeks together.

Uncertainty – Consumer marketplace keeps evolving with introduction of competitor SKUs, unprecedented global situations, social sentiments, social influences, and stock prices. These determinants are often overlooked in a traditional approach, and hence, only an agent-based model can take these into consideration ground-up and enable in understanding a dynamic consumer marketplace and can help you design clever promotional strategies.

Adaptivity – As our environment changes, so does our need for a framework that is able to adapt and reflect those changes or be able to simulate situations not observed in the past. Agent-based models can adapt to reactions from the environment and can help in devising future courses of action.

Intelligence – The need for instantaneous responses is key to closing deals that best suit both parties. Agent-based models are designed to respond to such interactions and are capable of explaining emergent behaviours.

What-If simulation framework

We must build a learning tool for understanding the dynamic multi-agent (manufacturer-retailer-consumers) behaviour under various market conditions, particularly one that is evolving constantly and can visualize possible outcomes, by simulating the nature of dynamic interactions.

The quality of simulations depends on the quality of data mapped on underlying distributions of various data points. The following framework allows us to define an agent-based model design and validate its performance before using the final model for simulations. In this design, we conceptualize the model bottom-up, capturing a consumer’s choice of consumption leading to a sale, which translates into weekly sales volume for a retailer. This allows us to leverage granular-level levers for micro-level study of a marketplace. With the design based on consumer purchases – sale by a retailer, we can study the utility value of a commodity to a consumer, the topline KPIs that drive the retailer, and in turn the manufacturer.

Process flow

The data enrichment phase is the starting point and acts as a solid foundation in the ABM architecture. This helps in providing better context to business functions, be it for a golden customer record in marketing analytics for personalized targeting, integration with Nielsen, Information Resources, Inc.(IRI) data for competitive pricing, for price ladder analytics to support a revenue management team, for re-segmentation of vendor-level panel models and inhouse pricing or for choice models to depict pricing functions. Consumer information can be substantially mined via third-party data sources such as Personicx, Experian, bureau data and so on.

Once the data is enriched with context, the next step is to identify and map consumption moments in a sales cycle. Seasonal spikes or holidays can help in designing a promotional strategy, which can not only benefit a manufacturer or a retailer but also the consumer to create positive association with a brand. Leveraging surveys, correspondence analysis, socially foraged data can enable daypart analysis on consumption patterns of consumers. Food and beverage consumption purchase data from vendors like Kantar, used in conjunction with prior mined data, portrays consumer purchasing behaviour and demographics, which further provides insights into brand and category growth opportunities.

After the occasion mapping step, package, pricing, and promo OKRs must be set up. It is often observed that consumer behaviour is governed by her or his ability to shift between choices. Hence, being able to optimize the right volume per pack is crucial to encourage bursts of small-scale sales, especially, for beverages, where we would rather prefer buying multiple small cans than being stuck with gigantic bottles. Targeting specific consumer segments with appropriate promos and pricing are key in understanding volume modeling in the geographic area of interest. Cross-sell and upsell to move certain SKUs to improve category performance and engaging with switcher consumer segments are very sensitive to pricing decisions.

The real world of consumers is complex and filled with constraints. It is important to mimic those uncertainties in your simulation model as well. There could be introduction of new KPIs by manufacturers or a retailer’s shelf space may shrink to accommodate other product categories, or some deal-defining negotiations might get held-up due to change of management. Being able to define business rules, assumptions, constraints are crucial to be able to accurately define simulation environments. They play an important part on what the underlying distributions of data shape each simulation can take, or the probability of a potential purchase at a particular price point.

Simulation training and validation are the penultimate steps to building the What-If simulation engine. Training the model with past data on sales or average transaction volume, comparing with baseline models enable us to see how well the simulations are predicting. In cases where the error is high, we retrain our multi-agent system until we arrive at an acceptable model error. The goal is to reduce the actual vs simulated before leveraging the system for modeling what-if scenarios for confident business decisions.

The above What-if framework, thus, will help you map interactions between autonomous agents and guide you to understand agent behaviour in an environment and predict outcomes successfully.

Be on the look-out for our next and final segment on ABMs. We will share an illustrative ABM network to help you design an agent-based model based on the above framework.