• Our guiding principles in building Recast — what we call an “Incrementality System” — and how MMMs fit into this system.
• Our philosophy of causality, which is that causal models are best validated by the quality of their predictions under intervention, and not just observation.
• The capabilities of the Recast model and platform in terms of its cadence — how quickly you’ll be able to take action based on new data, composition — what types of structural parameters (like ROI, time shift, etc.) are estimated, and action — what types of interactive decision-making tools we provide.

When we started Recast, we thought existing MMMs were slow, expensive, and built on shaky (at best) statistical foundations. We set out to build an MMM using state-of-the-art statistical technology that is purpose-built to help teams take action with their MMM to improve marketing performance and ultimately drive business performance. 

We believe the true value of an MMM is not in reporting past performance, but in powering the key planning workflows of marketing teams: forecasting, scenario planning, optimizing marketing budgets, and experimenting. 

We call this combination of model, platform, and mindset an “Incrementality System”.

Supporting the system and the underlying workflows requires a comprehensive, causal understanding of the relationship between marketing activity and business performance. Without that, you may run a forecast, but it will be inaccurate; you may optimize a budget, but it will reduce profit instead of increasing it.

This makes a causal MMM capable of making accurate predictions absolutely necessary, but insufficient — the planning tools, and the organizational mindset that embraces making smart bets, are the other half of a system that produces more efficient, more profitable, more scalable marketing programs.

More materials – How to Build an “Incrementality System”

Planning

Plan an initial marketing budget using a combination of planning tools, intuition and business knowledge, and insights and recommendations from your measurement providers.

Experimentation

In tandem, design an experimentation roadmap to gain more insight into your channel-level performance and to provide more signal to your measurement tools. You should prioritize channels whose uncertainty drives the majority of your forecast’s uncertainty, channels where you are hesitating to invest because you haven’t yet gotten a good read, and channels for which there are conflicting recommendations from your measurement systems.

Validation

Execute the experiments and use their results to validate the output of the MMM and to calibrate it to improve the accuracy of subsequent runs; this process of error correction continuously improves the underlying statistical model.

Optimization

With a refreshed model that incorporates new data and new experimental results, your marketing team can improve on existing plans using constrained optimizations for paid media channels that target business objectives (such as profitability, efficient growth, etc.).

Then, you repeat this process. The optimization becomes the plan, and the cycle repeats. 

Our philosophy of causal inference — robust predictions

An accurate causal model should be able to predict what is going to happen when you take action. Correlational models will mostly fail at this task.

In the context of marketing measurement, one can think of it this way: an MMM that simply captures correlations may appear to forecast well when the modeler is backtesting their sample data, but will make bad predictions when the marketers actually use the MMM’s results to make changes to their marketing mix. (Whether or not an MMM can fit in-sample data is completely irrelevant.) Only a “causal” MMM will make good out-of-sample forecasts when it is used to make decisions that drive exogenous change to the marketing budget.

There are infinitely many plausible models that can fit the data well, and many models that mistake correlation for causation that can predict what will happen as long as the marketing budget doesn’t change. For an MMM to create value, the model needs to identify actual causal relationships that can make accurate predictions once a company makes exogenous changes to their marketing budgets.

While we make extensive use of the structural causal model approach when designing the Recast model, we use the robust prediction framework to evaluate the degree to which the Recast model is capturing the true causal relationships between marketing activity and business performance. This emphasis on forecasting performance aligns with our philosophy that an MMM is only as good as the financial impact of the decisions made using it, and that the core workflows that an MMM should power are forward-facing — forecasting, optimizing, and experimenting — vs. simply reporting past performance (although Recast does that, too).

More materials – Invariance, Causality, and Robustness

Our model validation procedure

We often like to talk about MMM as a “small data, big inference” problem. A typical Recast MMM has roughly a thousand observations (a few years of daily data), a few hundred control inputs, and tens of thousands of estimated parameters (although, via the use of priors in the Bayesian model, the number of effective parameters is much smaller). This makes the process of building an MMM a delicate one, and we follow the “Bayesian Workflow” to arrive at a model that is robust, stable, and accurate. The key model building and validation steps are outlined below.

More materials – Bayesian Workflow, The Recast Process, How we do Holdout Testing at Recast

Deep dive business review

Recast’s modeling process begins with a set of conversations and exploratory data analyses to identify a business’s main drivers and patterns:

• Is it heavily driven by promotions or outside economic factors?
• Is it a household name driven by brand affinity, or a new brand driven by marketing?
• Have they run lift tests in the past, and what have they shown?

Every model is configured to each unique business, allowing us to (theoretically) estimate causal effects.

Enabling causal MMM through model validation checks

There are a number of internal- and external- validity checks that we run as part of the model validation process. These checks are specifically designed to help build confidence that the model has estimated true causal relationships and can be used for budget optimization.

Parameter recovery check

The true value of the parameters that Recast (or any MMM) estimates are unknown and unknowable in principle. Whereas the model can make predictions about sales that can be validated, the true ROI of Facebook can never be directly observed. This doesn’t make it any less important, and these ROI estimates drive most of the value derived from an MMM. There is only one possible way to actually know the unknown values that generate data, and that is to simulate it. That’s exactly what the parameter recovery check enables us to do.

It works as follows: we simulate a set of observations using known parameters, then run that simulated dataset back through the model, and compare the model’s estimates with the ground truth. This gives us confidence that when we unleash the model on the real data, the values it recovers will be close to the (unobserved) true values.

Stability loop check

We want to know that the parameter estimates from the model are robust to minor perturbations in the underlying data. If including small amounts of additional data causes the parameter estimates to swing wildly, that indicates that there are fundamental problems in the model and the results cannot be trusted.

At the same time, we do want the model to be learning from new data, and yes, even revising past estimates of channel efficiency as new data comes in, if the new data warrants it. 

In order to check the robustness of the parameter estimates, we re-estimate the model several times over rolling horizons, and compare the results from offset models to ensure that they are consistent.

Predictive accuracy check (backtesting)

Finally, we run the model with several holdout periods (120 days, 90, 60, and 30) to see how well the model forecasts held-out data. We take care to ensure that only data that could have been known in advance is included in the forecasts (so, excluding information like branded search spend and website visits) to prevent information leakage, and we weight our forecasts by their “difficulty”, which measures how different the marketing spend is from what would have been expected based on past behavior. These backtests don’t stop once we deploy the model — they happen every time we get new data and refresh the model. See how it looks in the platform here.

And because insights alone do not actually generate profit, we have built out the platform to support all of the steps in the incrementality system including planning, experimentation, validation, and optimization.

The key features of the Recast platform (documentation here) are:

• Validated, near real-time results
  • Automatic weekly (or on-demand) model refreshes based on daily data to meet the cadence of marketing teams
  • Automatic calibration with time-bounded incrementality experiments 
  • Always on, transparent backtesting
• State of the art model structure
  • Time-varying covariates to estimate how channel ROI evolves over time
  • Hierarchical model to estimate relationship between demand-generation channels and demand-capture channels (like branded search)
  • Sophisticated “spike” specification to account for pull-forward and push-backward effects of promotions
  • Contextual variable specification that accounts for interaction effects between context variables and marketing effectiveness
  • Shift curves (adstocking) and saturation (diminishing marginal returns) by channel
• Robust planning and optimization
  • Scenario Planning: Simulate “what-if” scenarios to forecast how changes in spend by channel will impact key business outcomes.
  • Budget Optimization: Optimize allocations across channels to maximize incremental impact, accounting for diminishing returns and time lags.
  • Planning & Goal Tracking: Set your plan and monitor your spend pacing and marketing performance each week, with alerts when performance deviates from forecasted targets.
  • Always-On Forecasting: Continuously updated forecasts ensure that planning decisions reflect the latest data and market conditions — including in-flight adjustment recommendations.

The holistic platform is designed to support an incrementality system of continuous planning, experimentation, validating, and optimizing.

For companies operating marketing at scale and with speed this approach isn’t optional — it’s foundational. Recast’s Incrementality System is designed to meet that challenge, helping teams confidently make smarter investments that improve performance, increase efficiency, and drive measurable business results.

Schedule a Recast demo

Learn how Recast can help you eliminate wasted marketing spend, optimize your channel mix and plan for the future with reliable forecasting.

Not ready for a demo? You can always reach us at: info@getrecast.com

Recast is SOC 2 compliant and a badged Meta Measurement Partner.