Client’s Problem: “We operate in a hypercompetitive industry where our products have slim margins, and the differentiation within a market is service-level and not product-level. Our issue is not knowing how competitor behaviors and strategies impact our sales. We struggle with the basics, like pricing, to capture the value proposition of our service over competitors. Even simpler things, like when and how to execute product decisions, are poorly understood outside of the tribal knowledge of our product managers, making long-term strategic decisions extremely difficult. All of this work is also very labor intensive, and the drudgery of it all distracts our talented people from doing other, more valuable work.”

We have a couple of former product managers on our leadership team, a widespread problem for retailers of any scale. Most companies avoid the problem of competitor intelligence altogether by washing it out (sometimes rightly) using one of the standard product management policies, e.g., margin or premium. During scoping, we determined there was sufficient evidence and data to pursue a more sophisticated approach, which was ultimately adopted.

The Client was a large national company, and they did indeed have many competitors. Our first step was to use the Client’s existing industry data to determine if there were patterns in sales, demand, and competitor data. We found signatures of competitor behavior impacting sales that were well-described by game theory. With this discovery, we automated the codification of the rules and actions seen in the games to build a data set for modeling. Using this game characterization data, we created machine learning-based models ensembled to predict when, where, and how a competitor would make moves to capture demand.

Interestingly, these models performed poorly during the experimentation phase. We learned in the experiments that an active spatial component to the competitor behavior was not evident in the original data. Competitors also had complex but quantifiable interactions over time, indicating structured network effects. We modeled both patterns in tandem using a novel approach based on neural networks. We automated the entire modeling backend into an intelligent agent for interaction with our Client’s analysts. The speed and efficiency of this new capability enabled rapid testing of sophisticated strategic scenarios they had never contemplated previously.