Model Center gives you a clear view of the Marketing Mix Models powering attribution and forecasting — so you can monitor accuracy, compare modeled and observed performance, and confirm which configuration is active.
Use the two menus at the top of the page to pick the modeled outcome and the reporting window. Everything — the overview and each card — updates to match.
Start with the summary strip above. It shows how many models and configurations are active, average accuracy, and totals for revenue, spend and ROAS — a quick health read before you dig in.
Each card is one business outcome. It shows the accuracy score, modeled vs observed revenue, and a fit chart — observed as a solid line, modeled dashed. The shaded tail is the forecast/holdout window. Click a card to open its details.
Inside a model, the Health tab plots modeled against observed. The metrics on the right — in amber — test the non-trained period: data the model never saw. The shaded Holdout region on the chart is that test window. Switch to Decomposed to split media from seasonality and trend.
A model can hold several configurations. The Model Selection Criteria compares them across key metrics — model fit, ROAS expectedness, outliers and bounds — and summarizes each as a Preference Score. The best model is a balanced fit, not just the highest single number.
Model accuracy measures how well modeled results align with observed history. It matters — but a high score alone doesn't guarantee every channel result is right. Weigh it with non-trained fit, stability and business context.
The non-trained (holdout) period holds back recent data the model never learned from, then checks its predictions against it. Strong non-trained performance is the best signal a model will forecast well going forward.
Read accuracy and non-trained fit together — never accuracy alone.
Sanity-check the Attribution tab before trusting a config downstream.
Use Decomposed view to separate media from seasonality and trend.
Watch the Holdout window — that's the model's real forecast test.
Confirm the Active configuration before quoting a model's results.
Pick the config that's reliable and realistic — not the highest score.