Is it possible to force the base to be positive?

[uomodellamansarda]

I have an overattribution of media impact that is pushing down the base to negative values.
I truly don't believe on that, but instead of setting priors to the media activity I would like to have the base always positive.
It seems similar but not the same.

Additionally I think having the knowt free to float can lead to strong overfitting situation. What is your experience with that?

Is something possible in Meridian?

[cpulavarthi]

Hello @uomodellamansarda,

Thank you for contacting us!

You are right that a negative baseline does not align with business expectations. Baseline values represent the expected KPI in the absence of media execution. Essentially, it is the portion of KPI attributed to media subtracted from the total expected KPI. Therefore, a negative baseline implies the brand would have negative sales without media, which is implausible.

Regarding Overfitting and Knots, based on the screenshot you shared, overfitting doesn’t seem to be a concern. The model’s expected outcome actually misses the sharp peaks and troughs of the observed data, suggesting it is capturing a mild trend rather than "chasing the noise" (overfitting).

Negative Baseline:

• The negative baseline issue usually occurs when all of the important confounding variables aren’t accounted for and this leads to over-attributing the KPI changes to media execution - which means that the model isn’t able to accurately capture the causal impact of media execution. You may want to carefully evaluate the causal graph of your business use case and identify any confounders missing in your current model.
• Choosing the correct prior type and customizing them appropriately to reflect your expectations can also help mitigate this issue. You may choose to use total media contribution priors. This allows you to set a "cap" (e.g., "Media cannot be more than X% of sales") without setting strict priors on individual channels.
• Setting knots can capture potential seasonality effect in the data to reduce bias in baseline time-effect estimate. Please incorporate business context to understand if the regular peaks and troughs in the "actual" data are indeed seasonality effects that the model has yet to capture instead of just noise. You may refer to this documentation page for more details on how to set knots and consideration about bias-variance tradeoff.

Detailed guidance regarding mitigating negative or low baseline is available in our documentation which can help you better understand the causes and potential fixes to this issue.

Let us know if you are still facing this issue despite following the guidance available in our documentation, with details regarding the same, so that we can attempt further debugging.

Google Meridian Support Team