Seasonality and/or trend modeling when only using national level data

[AdimDrewnik]

What would be a recommended way to approach that issue? In case of national level data knots are equal to 1 and baseline is a flat line. Are there any recommendations as to increase knots if there is seasonality and/or trend in the revenue? Or is it better to try manually create seasonal dummies and linear trend variable?

[cpulavarthi]

Hello @AdimDrewnik,

Thank you for contacting us!

We recommend starting at the default 1 knot and increasing the number of knots from there for the national level models. You may continue to increase the number of knots until overfitting becomes extreme or media effect estimates become unrealistic. This approach can help you capture the seasonal variations without the requirement to manually create any seasonal dummies or trend variables as Meridian handles seasonality and trend capture through the knots argument alone. Please refer to our documentation on Setting Knots for more details regarding this.

Please note that you may still add dummy variables in Meridian if you wish to capture seasonal peaks that may have been smoothed out using knots alone. You may refer to our documentation on when to use binary dummies instead of knots linked here. Meridian also allows you to specify knot locations while setting knots as demonstrated here.

Feel free to reach out for any further queries.

Thank you

Google Meridian Support Team

[AdimDrewnik]

Thank you.
Was it considered to make number of knots probabilistic with its own prior probability so that MCMC would select its "optimal" value during the sampling?

[cpulavarthi]

Hello @AdimDrewnik,

Meridian currently only supports manual choice of the knots argument and allows manual selection of the placement of knots. We understand that automatically obtaining an optimal value for knots will be a great feature to have but #knots is a discrete variable and can't be handled by MCMC.

Thank you

Google Meridian Support Team