Forecasting FAQ

Frequently Asked Questions

Do the models consider impression share? 

Yes, the models themselves will be different depending on impression share; if you're already at 100%, expect far flatter slopes. 

Google simulations are sensitive to impression share, and Marin models have dedicated models for high/medium/low/unknown impression share brackets.

Other constraints such as budget cap, imp share cap, and folder traffic status are ignored such that we forecast the unconstrained metrics (constraints are handled further downstream)

The only constraints we honor in forecasting are:

  • Campaign Status (we do not forecast future metrics if the campaign is PAUSED or DELETED)
  • Excluded Dates (we do not consider excluded or rolling excluded dates when building the time-series forecasts, but still output the reported metrics)

Do the models use historic changes in target to improve accuracy?

Yes. We expect more accurate simulations when a campaign has have multiple recent targets

What is the accuracy of the models?

Marin doesn't report accuracy statistics for the models (neither do Publishers). There are many independent variables that make it challenging to define accuracy.

Typically, we can expect 89% of campaigns to be within 15% of the forecast

Do you share data across customers?

No directly. We take great care to protect customers' data.

Simulations are campaign-specific. Where possible, we will use publisher sims, then customer-specific models, but we do have fallback models that aggregate publisher cost data across Marin customers, but we never share revenue or conversion data.

Are there cases where a model is unavailable?

In general, no. There should always be a model available for a given campaign regardless of publisher, provided there is recent performance (last 7 days spend and clicks)

We still have a model even for zero conversion campaigns, however, we might not be able to determine the effective current target (e.g. if ManualCPC and not on traffic, or if on smart Bidding with no target), and in which case if we cannot fallback to recent performance as a proxy target, we'll not recommend a new target (allocation will only be updating budget). 

We also never kill off a campaign directly - the minimum we can go in a single day is -50% of current target, or damping (if applied), whichever is the smallest. We would only change to -50% if it made sense to do that relative to other campaigns and the strategy target

Can I access campaign-level simulations and forecasts within Marin?

Yes, and we are extending capabilities to provide better support in the following months via:

  • Allocation Dashboard
  • Custom report (ask your Marin rep CE)
  • Aggregates to folder-level are available in chart form in the UI (Optimization > Forecasting view)

​​Is There A Way To Export Forecast Data? 

Yes, via Custom User Reports

Are Forecasting Available For Awareness Targeting Strategies? 

Yes - forecasting is available for all campaigns regardless of Strategy goal, Publisher, campaign type, biddable type, and other settings.  

The main requirement is that there is recent data available (non-zero clicks last 7 days)

How are the models used in bidding and optimization?

Campaign forecasts are used in Strategies with dynamic allocation goal (check out the dedicated article on Forecasting)

Do we honor excluded dates on the data used for Dynamic Allocation? 

Yes: we do not consider dates flagged by excluded or rolling excluded dates when building the time-series forecasts

Excluded dates are typically used to flag problematic/partial/inaccurate revenue data, or promotional periods (where revenue data is not representative of general performance, critical for criterion bid calculations within Marin Bidding

Typically, spend data can be considered accurate, so excluded dates will be ignored when forecasting spend and reallocating to Spend Targets (Max. Convs. to Spend Target, Max. Rev. to Spend Target, Pace to Spend Target)

For CPA & ROAS targets, data from excluded dates will be ignored (Jira needed)


Written by Joe Southin

Last published at: March 15th, 2024