Creating a feedback loop between incrementality tests and Marketing Mix Models with Agent Based Models

‍

The expectations of Marketing Departments regarding MMMs have significantly evolved in recent years. Advertisers now require:

• more granular results, as more data is available, and there is a real need to achieve a sufficient level of precision to enable decision-making,
• more frequent updates, as MMMs should support not only strategic decisions but also tactical ones, particularly regarding digital channels, for which the end of third-party cookies can be compensated by the use of modeling,
• results that are understandable by Marketing teams, who need to multiply, own and trust analysis angles,
• frequent advice on future strategies to implement.

‍

With their by-design granular methodology, Agent Based Models allow for precise modelisation, measure and decision-making. However, granular modeling does not come without effort or relevant data. To provide valuable insights with such approaches, the calibration phase has to be undertaken thoroughly. 

‍

Although classical MMM approaches calibrate models by comparing the model’s prediction of a targeted KPI with true, historical data, ABM calibration compares a very long list of metrics against the model predictions. Certainly, the targeted KPI (for instance the revenue) will be one of these metrics, but ABM allows to go much further. If the advertiser has a loyalty program, we can integrate this information in the agent definition and compare the model’s results on this subpopulation of consumers with the actual revenue generated by this population. Within the same model, we can also easily retrieve results per product category or by customer segments. These are just a few examples of calibration on historical data. 

‍

Once calibrated, the ABM can be used to simulate strategy at a very granular level (geographic, customer segments, catchment areas, …) to help choose the best possible course of action. Since such decision-making is highly strategic, incrementality testing is a natural solution to validate the ABM’s conclusions. Using results from incrementality testing, one can further calibrate the ABM. Indeed, since the ABM is highly granular, it can easily simulate/replicate the performed incrementality tests at the exact same granularity level and thus fine tune its parameters by updating the model parameters to become closer to the reality . In classical approach, the result of the incrementality test can be used to fine tune the model, but to a lesser extent, since the model can not simulate at the same level of granularity. However, since, by design, the ABM solution simulates consumer behavior, any calibration will improve the relevancy at the consumer level and thus the model as a whole.

‍

In conclusion, the ABM is a natural solution in order to put in production a MMM for granular decision making with a retroactive improvement using incrementality testing.