In Bayesian Marketing Mix Modeling, the model is given ROI priors — assumptions about the likely range in which the ROI for a marketing activity lies. These priors are expressed as probability distributions, and the choice between informative and non-informative priors is one of the most consequential decisions in how a model is built.
A non-informative ROI prior avoids making strong assumptions about the effectiveness of a marketing activity. Its distribution is wide and flat, telling the model that almost any ROI value is roughly equally plausible before the data is observed. An informative ROI prior, by contrast, encodes outside knowledge about how a marketing activity tends to behave. Its distribution is narrower and centred on a specific range, based on information from sources such as incrementality tests, geo experiments, or shutdown tests.
There are two schools of thought on this question in the MMM community.
The traditional school believes the model should find the true ROI from the timeseries data alone, without being nudged by outside information. The argument is that bringing external data into the prior introduces biases that are hard to control or predict. Practitioners in this camp default to non-informative priors and let the data do the work.
The more recent school of thought, which aligns with how we think at Sellforte, holds that traditional models can be meaningfully improved with informative priors built from randomized control trials, geo lift tests, shutdown tests and other experiments that reveal how a channel actually behaves. This process is called model calibration. Combining the traditional MMM approach with calibration tends to produce results that are both more accurate and more stable over time.
Calibration sits at the centre of how Sellforte builds models. Without informative priors, an MMM can settle on ROI estimates that fit the historical data well but contradict what experiments on the ground have shown. With informative priors derived from credible incrementality evidence, the model is anchored to reality and produces results that hold up across model refits and rolling windows.
Calibration is not easy, though. The value of an informative prior depends entirely on the quality and interpretation of the calibration data behind it, and any bias in that data flows directly into the model. Understanding those biases — when an experiment generalizes, when it does not, and how to translate its result into a prior — is where most of the craft lies. Sellforte has spent significant R&D effort developing this approach so that customers benefit from calibration without having to build the methodology themselves.