Is there a way to forecast KPI?

[twolf89]

To preface this, i have previously been using PyMC for forecasting a kpi given a future budget.
I have been trying to understand the KPI to tease out what I need to be able to do this since it isn't currently a feature (will it be? Because that seems essential). Looking at some other discussions and online posts, people are trying to optimize future budgets but I haven't come across forecasting KPI. Is there a tip where I should start to be able to do this?

[gousesb]

Hello @twolf89 ,

Thank you for contacting us!

Meridian is a Causal Inference model designed to help you analyze past results, understand key drivers of those results, and assess the efficiency of your budget allocation. It's important to note that Meridian is not a forecasting tool; it doesn't predict future KPI values or provide future budget allocations. The goal of Meridian is to determine the causal impact of treatment variables on the target KPI, not forecasting it into the future.

You may however use the new_data argument in Analyzer class methods which allows you to analyze results predicted by the model for unseen data. For example, the BudgetOptimizer function can determine budget allocations for the what-if scenarios using the new_data argument in the optimizer() function. This may also be used for forecasting if the new_data passed represents the expected future values of treatment variables. We don’t recommend using this method for forecasting, this can only be used for exploratory purposes to gain insights.

If new_data is provided, the optimization is run on the versions of the tensors in new_data and the original versions of all the remaining tensors (i.e., you will only need to provide media data in the new_data and controls, etc. are chosen from the original data). The new_data is expected to have the same time-dimensions as the original input data on which the model is trained. If any of the tensors in new_data is provided with a different number of time periods than in InputData, then all tensors must be provided with the same number of time periods and the time tensor must be provided. We can customize this also by providing media, reach, frequency, media_spend, rf_spend, revenue_per_kpi, and time.

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

Thank you

Google Meridian Support Team

[jmp1989]

@gousesb Let me just ask a followup question to hopefully get some clarity on a hypothetical example. I looked in prior comments on this discussion board and got slightly confused, perhaps with older versions of the tool too.

Background: I built a model on data from 2023-01-01 to 2025-07-01 (2.5 years) to estimate Sales based upon below inputs.

Scenario: I want to know what my sales might be given the media inputs and control variables below.

I'd like to be able to pass in all sorts of combinations of values like below or for 1month or 2mo, etc. I'd need to be able to report to my team "given budget A, sales are 10,000, given budget scenario B, sales are 10,500", etc.

Is this achievable or doable in some way that is not blatantly incorrect?

Date	YouTube Spend	Facebook Spend	TikTok Spend	S&P 500
2026-01-01	$5,000	$3,500	$2,800	6,824.14
2026-01-02	$5,200	$3,800	$3,100	6,842.82
2026-01-03	$4,800	$3,400	$2,700	6,819.38
2026-01-04	$5,500	$4,100	$3,300	6,858.74
2026-01-05	$5,100	$3,600	$2,900	6,845.98

[gousesb]

Hello @jmp1989 ,

Thank you for contacting us!

The Analyzer module contains several methods with a new_data argument, which allows you estimate outcomes under "what-if" scenarios under the posterior distribution of a trained model. In particular, the incremental_outcome and expected_outcome methods may be useful to you.

• The expected_outcome method allows you to pass in new media and controls data. Meridian only estimates time effects within the time window of the data used to train the model. Therefore, the new media and controls data time window must align with the time window used to train the model.
• The incremental_outcome method allows you to pass in new media data. Time effects are do not affect the incremental outcome, so you can pass media data for any past or future time window. Similarly, control variables do not affect the incremental outcome and are not required.

The BudgetOptimizer.optimize method also contains a new_data argument. Similar to the incremental_outcome method, you can pass in hypothetical media data for any past or future time window.

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

Google Meridian Support Team