What Is This?
This tool combines two powerful concepts: Media Mix Modeling (MMM) and Predictive Forecasting.
What is Media Mix Modeling (MMM)? An MMM is a statistical technique used to understand how different "mix" variables (like ad spend, pricing, seasonality, promotions) contribute to a business outcome, such as revenue. Instead of just looking at the last click, an MMM looks at all marketing activities over a long period (2+ years) to determine the true, incremental impact of each one. Our tool uses Google's Meridian platform, the same one our MarSci team uses, to do this.
What is Forecasting? A forecast is a prediction of future performance. Traditional marketing forecasts are often inaccurate because they assume the future will be just like the past (e.g., "if we spend 10% more, we'll get 10% more revenue"). This fails because it ignores the law of diminishing returns.
Why do they work so well together? Our tool doesn't just forecast. It uses the results of the MMM to power the forecast.
It works in a two-step process:
Step 1: The MMM (Meridian): The tool first runs a complex Media Mix Model on the client's historical data (2+ years required). This model understands the true incremental impact of every channel and generates iROAS (Incremental ROAS) values and response curves.
Step 2: The Forecast (nova): You then use our new UI to input a budget (either manually or using an optimized suggestion). Our tool applies the iROAS values from Meridian to that budget to generate a 12-month predictive revenue forecast.
This method inherently accounts for diminishing returns, giving you a realistic view of what to expect as you scale spend.
How to Use It: A Step-by-Step Guide
Follow this process to generate your first forecast and compare scenarios.
Step 1: Run the MMM (Meridian)
1. Navigate to the nova Forecast tool for your client.
2. Review Priors (Important!): The model is pre-loaded with "priors" (starting assumptions) from our MarSci team. These are often based on client-specific MMT tests or deep industry data.
Warning: Do not edit these priors without first consulting the MarSci team. Changing them without validation will make the forecast inaccurate.
3. Upload Non-Media Data (Optional): The model is good, but it's better when it knows about offline activity. You can optionally upload a .csv file with data on promotions, holidays, or other non-media events that affect sales.
4. Run Model: Click the "Run Meridian Model" button. This may take a few minutes as it processes 2+ years of data.
5. Check Model Quality: Once the model runs, first look at the "Model fit metrics". You'll see metrics like R-squared and MAPE (more on those in the next section). The tool will also warn you if the model "didn't converge" (couldn't find a stable answer) or if there isn't enough data.
6. Review iROAS Curves: Go to the "Response curves by marketing channel" chart. This shows you the diminishing returns for each channel.
Step 2: Create Your Forecast Scenario
1. Configure Your Budget: Go to the "Budget Configuration" table. You have two options select the time period to forecast:
Default: Use historical budget (optimized). This is the default setting. The tool takes the client's historical spend and re-allocates it for maximum ROI based on the new iROAS curves.
Manual: Use manual budget entry. Select this to enter your own spend figures by channel and month, creating a custom plan.
2. View Forecast: As you adjust the budget, the "12-Month Revenue Forecast" chart will update, showing you the projected revenue and spend.
Step 3: Compare Scenarios
1. Save Your Scenario: Give your forecast a name (e.g., "demoCS," "Aggressive Q4 Budget") and save it.
2. Create Another Scenario: Repeat Step 2 with a different budget configuration (e.g., one with more spend on Google, one with less on Facebook) and save it with a new name (e.g., "demoCS2").
3. Compare: Go to the "Scenario Comparison" tab. Here, you can see all your saved scenarios side-by-side, allowing you to easily compare the projected revenue and spend for each plan. This is the perfect view for strategic client conversations.
How to Interpret It: Key Concepts
This tool gives you new, powerful metrics. Here’s what they mean.
What is Meridian & Bayesian MMM?
Meridian is Google's open-source MMM solution. It uses a Bayesian statistical model.
In simple terms: most models try to find one "right" answer. A Bayesian model instead figures out the most likely answer and also tells you how confident it is. This is perfect for "noisy" marketing data. It also allows us to feed it "priors" to make it smarter.
What are Priors?
A prior is an expert assumption we give the model before it runs.
Without Priors: The model starts blind and has to guess.
With Priors: We can tell the model, "Hey, our MarSci team ran an MMT test and we're 90% sure the real iROAS for Meta is between 1.5 and 2.0."
This guides the model, making its output vastly more accurate and aligned with our internal MarSci methodology.
What is iROAS vs. Platform ROAS?
Platform ROAS (What Meta/Google reports): This is a total average. It takes credit for users who would have converted anyway (e.g., brand search, repeat customers).
iROAS (Incremental ROAS): This is what Meridian calculates. It measures the true, incremental revenue generated only from ad spend. It answers: "How much extra revenue did I get for the next dollar I spent?"
What are Response Curves?
This chart shows the law of diminishing returns. As you spend more on a channel, it gets less efficient (iROAS goes down) because you're reaching less-qualified audiences.
This curve is the most important tool for budget planning. It shows you the "sweet spot" of spending before efficiency falls off a cliff.
What is Baseline vs. Media Contribution?
The "Contribution" chart splits all your revenue into two buckets:
Baseline: Revenue you would get even with $0 ad spend. This comes from brand equity, direct traffic, organic search, and repeat customers.
Media Contribution: The incremental revenue driven by each of your paid channels.
How to Read the Model Quality Report
You will get a "Model fit metrics" report. Here’s what the key metrics mean:
R-squared: "How well does the model's prediction match the past?" It's a percentage. A value of 0.48 means the model explains 48% of the changes in historical revenue. For noisy marketing data, anything above 0.4 is decent, and 0.7+ is excellent.
MAPE (Mean Absolute Percentage Error): "On average, how far off was the model's historical prediction?" A MAPE of 35% means the forecast was typically off by +/- 35% from the actual revenue. Lower is always better.
wMAPE (Weighted MAPE): A "smarter" version of MAPE. It gives more weight to the errors from high-revenue periods (like holidays), making it a more relevant business metric. A wMAPE of 26% is a good result.
Benefits: Why Use This?
Builds Trust: Finally, a forecast tool that isn't a black box. It's transparent, explainable, and backed by the same methodology our MarSci team uses.
Provides a Holistic View: It moves you away from last-click attribution and shows how all channels (paid, organic, and baseline) work together to drive the total business.
Enables Strategic Budgeting: Stop asking "What was my ROAS?" and start asking "What is the most efficient way to spend my next dollar?" The response curves and scenario planning features are built to answer exactly that, helping you optimize client budgets.
Improves Client Conversations: You can now confidently walk into a QBR with multiple, data-backed budget scenarios to support your strategic recommendations.
Frequently Asked Questions (FAQ)
1. What data is required to use this tool?
To run, the client must have 2+ years of historical data and active integrations for their e-commerce platform (Shopify or GA4) and their primary ad channels (Google Ads, Meta Ads, TikTok Ads, Bing Ads, Pinterest Ads). The model also uses Google Trends data automatically.
2. Do I have to upload non-media data?
No, it's optional. If a client runs major promotions and you don't tell the model, it will incorrectly think your paid ads were responsible for the sales spike, which will inflate your iROAS and make the forecast inaccurate.
3. Can I edit the priors?
The tool allows some priors (like those based on broad industry data) to be edited. You should not edit priors based on client-specific MMT tests. Always consult the MarSci team before changing any prior.
4. What is "model convergence" and what is "R-hat" (R̂)?
Convergence: Since a Bayesian model runs many simulations to find the "most likely" answer, "convergence" means all the simulations agreed on an answer. If you get a warning that the model did not converge, it means the data was too messy or insufficient, and the results are not stable or trustworthy.
R-hat: This is the metric that measures convergence. You don't need to know the math, just this rule: R-hat should be 1.05 or less. If R-hat is high, the model did not converge.
5. What do I do if my model doesn't converge or my quality metrics (R-Squared / MAPE) are bad?
This is almost always a data problem.
First, ensure you have at least 2 years of clean data.
Second, upload non-media data. Unexplained spikes from promos are the #1 cause of bad model fit.
If both of those are done, contact the MarSci team. Do not use the forecast; the model needs to be investigated by an expert.
6. Can I add more channels to my forecast?
No, the channels we will have enabled are the channels that we have the data ingested through nova cloud and have been reviewed to comply with our data quality standards and requirements.
7. Can I forecast a period other than 12 months?
Yes you can forecast from 1 month to 12 months, the shorter the period is forecasted is more accurate.