How to predict the kpi for the future ?

[Aminshnn]

I know that MMM models are not for prediction but they should be used for causal inference. However, I want to set targets for the future. So, given a new dataset, I can get the incremental contribution of paid media variables from the model easily. However, my question is how I can get that for non-media treatment variables and the baseline. Here is how I calculate the kpi obtained from non-media treatment variables. If those values are stored in raw_input, I use the following code:

``
transformer = mmm.non_media_transformer

normalized_data = transformer.forward(raw_input, apply_population_scaling=True)

gamma_gn_samples = mmm.inference_data.posterior['gamma_gn'].values

kpi_scaled = np.einsum('gtn,cdgn->cdgt', normalized_Data, gamma_gn_samples)

kpi_original = mmm.kpi_transformer.inverse(kpi_scaled)
``

I am running the model at national level, so, there is no geo level data.

To check my procedure, I tried to retrieve the baseline value from the trained model (not for the new data). To construct the baseline, I use the following code:
``mu+t = mmm.inference_Data.posterior['mu_t'].values

mu_transformer = mmm.model_xontext.kpi_transformer

kpi_original = mu_transformer.inverse(mu_t)
``
The issue is that when I use this code, the results are different from what I get from

``
media_sumamry =visualizer.MediaSummar(mmm)

media_summary.contribution_metrics(aggregate_times=False, included_non_paid=True)
``

Both codes shows the same patterns but values are different.

I am wondering how I can predict the baseline for the future periods (like the next 6 weeks)? Can I do that based on the trained model? If it is not possible, should I train an ML/time series model to predict the baseline for the future?

Thanks

[cpulavarthi]

Hello @Aminshnn,

Thank you for contacting us!

In a national model with no trend (knots = 1), you can broadcast the static intercept, combined with forward effects of other variables (paid media, organic media, non-media, controls), to obtain forecasted KPI (you might have to do your own custom einsums to combine the future variable values with estimated posterior coefficients). This isn't possible if the model has a trend. In this case would need to build out a separate trend forecasting model (Meridian does not support this use-case).

Feel free to reach out if you have any further questions or suggestions for us.

Google Meridian Support Team