Model Configuration Phase

There are a number of steps Recast performs in order to ensure that your model is performant. Many of the steps are internal, but there are a few steps that will require your input:

  1. Compiling the dataset
  2. KPI Selection
  3. Model Configuration Call: this is a conversation about your business where we’ll set inputs for the model. We’ll discuss how the marketing program is currently run and the team’s expectations around channel-level performance (priors), and the size of organic (or non-marketing-driven results, also called “Intercept” in Recast terminology)
  4. Recast Model Checks
  5. Initial results call: we'll review your initial results in the Recast platform and revise results based on your feedback

🔢 Compiling the dataset

See 📊 Data Guide for more information on compiling your data and sharing it with the Recast team.

KPI Selection

Your KPI is your outcome variable that Recast will model. This could include sales, acquisitions, conversions etc. depending on the nature of your business. The KPI is central to the model.

Select your KPI based on how you are measuring your marketing results. When you spend on your marketing, what metric are you tracking to measure effectiveness? This will be your KPI.

If you have multiple KPIs you want to model, reach out to your CSM or [email protected] and we will talk through options for building multiple models.

Model configuration call

The goal of the model configuration meeting is to infuse the Recast model with your prior knowledge so that the model will return actionable results that align to your general understanding of your business. You will meet with the Recast Customer Data scientists to discuss your business and your current understanding of your marketing performance. This will help us configure your model.

It is important to note that while we ask for your input on priors, our goal is not to highly constrain the model but rather to use your inputs to constrain the parameters to the space of what is truly believable, based on your expertise.

What we generally cover in model configuration:

  • Channel configuration
    • Channel setup
    • Baseline volume
    • Channel-level ROI
    • Channel-level time shift
  • Any outstanding data clarifications that might be required to proceed with model configuration

Control Variables

We typically take a less is more approach with adding control variables to the model, as each variable that is included becomes an additional dependency to be able to make accurate forecasts about the future.
For example, if we were to include App Downloads as a control variable (aka “contextual variable” in Recast terminology) in order to use the Recast forecasting or optimization tools you would need to also be able to predict App Downloads in the future, something marketers do not directly have control over.
Whether or not these all would be valid to include in the model will also depend on the KPIs that we are modeling.
If these are all readily available and simple enough to provide it might be worth sharing them with us, but we would take a parsimonious approach on determining whether or not to add them to the model.

Contextual Variables

Contextual variables in Recast’s model do not have to be expressed exclusively in terms of spend. Recast will ultimately use these variables to better inform our estimates of the Intercept as well as the efficiency of marketing spend.

Channel Breakdown

We need to limit each model to roughly 30 channels or less. How you choose which channels to use will largely be determined by how you think about setting budgets. When you think about channels, it’s important to consider:

  • What you need to be able to explain when making budget decisions
  • Collinearity, interdependence of predictors
  • Granularity of spending

For larger channels, we normally recommend splitting the channels into tactics (e.g. prospecting and retargeting).
For other channels, we consider the following:

  • Will there be enough spend in the smaller components for the model to get a reasonable signal?
  • Will you be able to take action based on the model’s recommendations?
  • Are the channels highly correlated? Unless there is some way of breaking the correlation (like going dark in one of the channels or raising spend in one while lowering the other), the model will not be able to distinguish the effects from two highly correlated channels
  • Are you planning to spend into channels in the future? If not, we always recommend focusing on the channels where you are planning future spend
  • Does that channel have significant spend? Channels that make up less than 2% of the budget or have total lifetime spend of <$100k-200k are good ones to consider grouping

Note: We want to make sure all dollars are accounted for. We don't want to leave any marketing spend out of the model completely, as that will bias the results. Therefore, all spend needs a category in the model.

Priors

With the flexible model specification Recast uses, there are infinite combinations of parameters that will fit the data equally well. We want to use expert information (generally from the marketing or marketing science team at the brand we’re working with) in order to constrain the parameters to the space of what’s truly believable.

We do this through a prior-setting process that we run in conjunction with marketing experts at the brand we are modeling. Priors are your business estimates about channel ROI, time shifts, baseline volume and saturation. The Recast model uses your estimates as a guardrail to make sure that it is optimizing your budget in a practical and actionable way for your business.

Your priors act as a starting point for the model to base its optimization on. They are assumptions the model makes about the dispersion of your data. If your priors are incorrect (for example, if your estimate of your baseline volume is too high), this may lead to less attribution to your channels, a negatively skewed ROI calculation, and a suboptimal spend recommendation. If you think your priors are set incorrectly, contact [email protected].

The goals of prior setting include:

  • Extract true prior beliefs via a set of simple questions non-statisticians are capable of answering
  • Set relatively uninformative priors so that the results are not unnecessarily biased by those beliefs

There are two places where the priors are especially important to collect from marketers:

  • Determining the priors on the impact of the intercept vs. marketing variables
  • Determining priors for the time-shift effects of marketing spend

We ask marketers to give us bounds for these parameters and to identify their most plausible values. The best practice is that the bounds should just be informative enough to exclude values of the parameter that are implausible. We generally avoid using highly informative priors unless there is hard data to support them.

Intercept Prior

When thinking about your intercept (organic) prior, consider:

  • What percent of your business do you believe would remain if all marketing spend were turned off?
  • What percent of the KPI is driven by organic demand?

Example: prior setting for intercept or baseline sales variable

In this example, the midpoint of the distribution will be 25% with plausible values ranging between 10% and 70% of the total amount of the KPI.

In this example, the midpoint of the distribution will be 25% with plausible values ranging between 10% and 70% of the total amount of the KPI.

ROI Priors

To set ROI priors:

  1. We get rough constraints on your ROIs by looking at your historical performance divided by historical spend. This gives us an upper limit on what historical ROI has been. We then take the intercept bounds you provide and “subtract out” the organic component of sales. This gives us a range for the overall marketing ROI, whose width is dependent on how wide the range was for the intercept
  2. Next, we expand this range to incorporate the idea that some channels are better than other channels and some time periods are better than other time periods. As a rule of thumb we assume that some channels on some days will be 3x better than the average, and some channels on some days will be 3x worse than the average. By looking at the range created from simulating a “low organic, high performing channel” and a “high organic, low performing channel” we end up with a broad range, but one that is in the realm of plausibility for a company with their spend and sales volume

While we can tweak these ROI ranges for specific channels based on your knowledge about performance, we usually do not because clients should look to our MMM as an independent, confirmatory voice regarding the performance of their channels. If we change the ROI priors to reflect previous biases, it becomes difficult to interpret whether the final results are driven by the data observed or by previous biases.

During our model configuration conversation re: ROI priors, we’ll ask you “What do you expect the CPA/ROI to be for each channel?”

Note: Recast support provides a breakdown of implied ROI/CPA vs. Intercept (baseline performance). Leave these set to the implied value based on the Baseline Performance, allowing the model to entirely determine channel efficacy

Example: ROI configuration

Time Shift Priors

For each channel that Recast will be modeling, we’ll ask “How long do you believe it takes for the full effect of the spend to be realized?”

We elicit time shift priors by asking you to estimate the range of days it takes for the full effect of spend to be observed.

Typically we expect channels that are higher up the funnel to have more long-term impact and the effects of channels lower down the funnel to be shorter-lived. For instance, we would generally recommend setting the time shift bounds for a linear TV channel to be higher than the bounds for a non-branded search channel.

Example: time shift prior configuration

These numbers set prior bounds on the number of days since day 0 (the day of spend) until 95% of the effect has been exhausted. The bounds on “facebook_retargeting” in the image above are 2 to 10 with a midpoint of 5, which means that 95% of the effect could be exhausted by day 2 or by day 10, or any time in between, and that the most likely length is 5 days.

These numbers set prior bounds on the number of days since day 0 (the day of spend) until 95% of the effect has been exhausted. The bounds on “facebook_retargeting” in the image above are 2 to 10 with a midpoint of 5, which means that 95% of the effect could be exhausted by day 2 or by day 10, or any time in between, and that the most likely length is 5 days.

Tips for setting good time shift priors:

  • Time shift bounds should represent the range of effect lengths that are plausible
  • If a channel is expected to have a longer time shift, don’t set the lower bound to 0. You wouldn’t really believe all the effects could be exhausted on the day of spend, right?
  • Be wary of setting very long time shift priors. It is very hard to distinguish an effect lasting longer than 90 days from background organic demand changes

Channel-level saturation

Channel level saturation is how we determine the scalability of your channels. To identify the saturation, it’s important to consider:

  • Which of your channels do you feel could not be scaled up further from their current spend levels under business as usual conditions? (e.g. search, retargeting)
  • Which of your channels could you increase spend in without expecting a large decrease in efficiency? (e.g. channels with a large addressable audience, low frequency)

Changing your priors post-model configuration

If your business has seen a large change in ROI, timeshifts, baseline performance, or saturation of any given channel, it is essential to update your priors. Contact [email protected] and we will help you configure your model accordingly.