Meridian MMM model refresh (to generate outputs) without model re-fit

[asos-jenniferlim]

Hi

We have build a national MMM model for the first time on 2021-2024 data and we would like to have refresh the model with 2021-2025 H2 data.

We are confused with what's the best approach for this.

• Refit the model with 2021-2025 data so generate a new MMM model (without changing previous ROI priors), or

• Pass the new dataset 2021-2025 to the existing fitted model (on 2021-2024 data) to generate summary outputs / response curves?

For the second option, I tried using class analyzer.DataTensor to generate new_data variable to be inputted into analyser.summary_metrics with the latest 52 weeks (list(new_meridian_input_data.time[-52:].data)) as selected_times and got this error ValueError: selected_times must match the time dimension names from meridian.InputData. And if I do not specify using a specific time-period, there's another issue on ValueError: new_controls.shape must match controls.shape.. There's more data in new_data which is why the TensorShape is different. Do you have any working examples that I can use?

[gousesb]

Hi @asos-jenniferlim ,

Thank you for contacting us!

In short, the best thing to do is re-run the model on the full dataset including the new data. You can still use the posterior distribution from an older model to inform the priors. However, consider that setting priors that match an old MMM's results effectively counts the old data twice. So, you might want to relax the prior a bit (e.g., increase the standard deviation) so that you aren't over-weighting the older data too much.

With that said, you may also model the new data disjointly and use the priors from the previous model. The decision to discard old data when appending new data is a bias-variance trade-off. A longer time window reduces variance because you have more data, but it can increase bias if media effectiveness and strategies have changed drastically over time.

You may check our documentation on Model Refresh for detailed considerations regarding the best approach for your use case.

The new_data argument is used for analyzer class methods to incorporate new data while generating outputs like expected or incremental outcomes or even summary metrics, not for model refreshes. The dimensions of these new tensors must correspond to those of the original tensors in input_data. When new_data is utilized, relevant metrics are calculated using the tensor values from new_data and the original values of all other tensors. This is the reason for the dimension mismatch errors encountered during the creation of the new_data object as the time dimensions don’t match for the data you are attempting to add.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team