Delayed or Weibull Adstock

[rajeshpaleti999]

In several marketing contexts, the peak effect of a channel is not immediate but is delayed.

The attached paper from Google Research also alludes to this important point and discusses delayed adstock as an option to capture this effect.
46001.pdf

Is addition of this delayed or Weibull adstock in the roadmap of the Meridian team? Thanks!

[Jon77Ruler]

https://github.com/google/meridian/blob/main/CHANGELOG.md

Think this has just been released? Might be what you're looking for.

- Add binomial decay option to Adstock.

[rajeshpaleti999]

The binomial adstock allows for slower decay but the maximum impact is still at lag 0 (i.e., immediately after adspend). Weibull and delayed adstock, on the other hand, allow maximum impact at a different lag.

[dattatreyam23]

Hi @rajeshpaleti999,

Thank you for your feedback! We appreciate it and shall take it into consideration for our upcoming releases.

In the meantime you may want to experiment with the available customization options in the ModelSpec. In particular, the adstock_decay_spec can be set to either geometric or binomial decay. Binomial has slower initial decay, which makes it more similar to the delayed peak function. You can also try setting the alpha_m prior or the max_lag to further customize the adstock decay.

Our team is constantly working on improving Meridian's functionality and you may check the unreleased features on the CHANGELOG page in GitHub.

Feel free to reach out if you have any questions or suggestions regarding Meridian.

Thank you,

Google Meridian Support Team

[isaacmalka]

hey, i have a somewhat similar question and was wondering if there's any update on the plans here. say we are trying to model upper funnel channels and understand their contribution to the kpi and estimate their lag which could sometimes be even a few months - is it possible to use the current framework to estimate the above?

[dattatreyam23]

Hi @isaacmalka ,

Thank you for contacting us!

Yes, it is possible to use the current Meridian framework to model upper-funnel channels and estimate lags of a few months.

To recap my previous response, you may want to experiment with the available customization options in your ModelSpec:

• adstock_decay_spec: This can be set to either geometric or binomial decay. The Binomial adstock decay function often results in a larger proportion of the estimated media effect being "long term". [See here]

• You can try setting the alpha_m prior to further customize how the adstock behaves. Specifically, you could experiment with setting a Beta prior that places more probability mass on values closer to one.

• The default max_lag value is 8, but you can try increasing it to 12 weeks or more.

Feel free to reach out if you have any questions or suggestions regarding the same.

Thanks,
Google Meridian Support Team