How is the intercept handled in Robyn, and how can I improve model stability?

[shantisb]

I have a few questions around how the intercept is assigned in Robyn and how to improve model stability and decomposition consistency over time.

I'm noticing that one of the main challenges I face with Robyn (and MMM in general) is that small changes—such as adding or removing variables like seasonality, trend, or competitor sales, or even simply adding more recent data—can cause the model to become unstable. In some cases, this leads to non-convergence, or previously converging models producing very different decompositions.

Additionally, I’ve observed that in several runs, a large share of the revenue is attributed to the intercept, even when all relevant media and control variables are included. This makes it difficult to draw actionable insights from the media contributions.

Could you please clarify:

1. How is the intercept assigned in Robyn, particularly under ridge regularization?

2. What best practices can I follow to improve model stability across updates (e.g. tuning suggestions, validation)?

3. Are there specific parameters or penalty settings I can tweak (e.g. lambda, adstock, Hill function settings) to reduce how much variance is attributed to the intercept and make media contributions more meaningful?

Thank you in advance—really appreciate any advice on how to get more stable and interpretable outputs from Robyn!

[AdimDrewnik]

One problem is that trend component also has intercept included, so when both intercept and trend are in the model then they both compete for the same attribution which is one source of instability. Also seasonality is taking both positive and negative values which causes seasonality to get zero attribution most of the times in the waterfall graph, which is counterintuitive to many users. And there are many more similar traps here and there in Robyn.