How does the Recast Model work?

What methodology does Recast use?

Recast is a fully Bayesian statistical model powered by a Hamiltonian Monte Carlo sampling algorithm implemented in Stan.

What level of detail is shared?

Recast contains a high level of information detail about the model specifications, raw data such as ROIs/CPAs with confidence intervals, saturation curves, modeled time shift, and the proportion of realized impact vs. expected impact. You can expect this data the weeks following your model build

Is there a minimum conversion threshold needed for a model readout?

There needs to be "signal" in the model, so >100 conversions per day will probably be okay, though it's tough to give a definitive answer before seeing the data.

What qualifies a variable as significant, and how often are existing variables and new variables re-evaluated?

"Significance" is not a concept in Recast; it's a fully Bayesian model. There's no need to remove insignificant variables. All variables related to marketing activity are included, and there's no re-evaluation of variables.

How are non-media variables determined, and is there a standard set of variables considered for all clients?

Recast's approach differs, focusing on media variables. Non-media variables (e.g., earned media, promotional events) are exceptions.

Do your models typically account for customer type? Are you able to support additional custom client segmentations?

Customer type splits can be modeled in the dependent variable. Different approaches, such as modeling revenue by customer type or using a hierarchical approach, are possible.

How does your model balance short and long term ROI? How is branding media considered?

Brand spend impact is estimated for new customer conversions within a reasonable timeframe (e.g., one quarter). Brand effects not showing up in short-term purchases are recommended to be tracked separately.

Does your model take a specific post-impression or consideration window into account? If so, how is this determined?

The model incorporates sales cycle, by-channel attribution windows, and time shifts based on client input. Bayesian priors are used to estimate different time shifts for each marketing channel.

How does your model incorporate lift/incrementality test results tested in the modeling window?

Recast produces ROI and marginal-ROI estimates for every channel, allowing validation against outside incrementality tests. Results of external tests can be incorporated to anchor the model. Talk to our support team to add lift tests to your models.

Why does Recast use Bayesian instead of Machine Learning/Random Forest?

Random Forest is optimized for prediction, leading to misinterpretation when correlated channels are equally effective. The Bayesian approach provides confidence bounds for every parameter.

How does Recast measure and account for saturation curve variation over time?

The model allows for different saturation curves over time through a saturation multiplier, considering long-term trends and spike effects localized around holidays.