How to use Marketing Mix Models for forecasting and optimization

So, you’ve built a solid Marketing Mix Model (MMM) to make sense of your past performance, but the real value lies ahead - using it to forecast and optimise. This is where MMMs shine, offering data-backed insights for decision-making that can drive future success. 

Sounds straightforward, right? Well… not so fast. Moving from a backward-looking analysis to a forward-looking application isn’t always as simple as it seems. Let’s explore the challenges and how to overcome them. 

• Forecast or optimization?
• Response curves
• The factor of time
• Aggregating curves  
• The importance of sense-checking your curves 
• What’s changed since your model was estimated? 
• The basic requirements of a robust optimiser 
• The real-time data challenge 
• The future is full-cycle forecasting 

Forecast or optimization? 

The first crucial step is to clarify your objective: what is the business question you are trying to answer? Are you trying to forecast the future or optimise your marketing investments? These are two distinct, and very different, approaches. 

Forecasting involves predicting outcomes by making assumptions about all the factors in your model - pricing, macroeconomic factors, distribution, competitors, and more. It’s complex because getting an accurate forecast means having solid assumptions about these variables. Without that, even the most accurate MMM won’t guarantee success. 

In contrast, optimization usually focuses on refining specific variables, like your media mix. This simplifies the process as it allows you to largely ignore some non-media factors. With fewer variables in play, you can zero in on fine-tuning your investments for maximum efficiency. 

Either way, curves are the heart of your strategy 

Response curves are arguably the most important outputs from your MMM. Curves allow you to simulate various "what-if" scenarios, showing how changes in your inputs (e.g., media spend) affect outcomes (e.g., sales). 

Think of response curves as the blueprint for your optimization. They summarize in one visual how different levels of investment impact your KPIs. But remember – how these curves are calculated in your model might not exactly fit how you want to use them in planning, leading to a potential disconnect. This is where expertise in both modeling and business application comes into play. 

The factor of time 

Time plays a major role in how MMMs should be interpreted for forecasting and optimization. Adstock effects, for example, determine how much advertising impact carries over into future periods. Do you only care about short-term payback, or are you looking at long-term returns? 

If you’re planning with adstocks in mind, you’ll need to consider what happened in previous periods. For instance, sales in Q4 might not just be a result of Q4 media spend but also the cumulative impact of your Q3 efforts. 

Aggregating curves  

The response curves estimated in MMMs can be very granular and detailed but may not align with how decisions are actually made in practice. For annual or quarterly planning, marketers don’t usually need granular weekly insights. They need broader, aggregated insights that show how budgets should be allocated across channels. 

Aggregating these curves - whether across time, campaigns, or media channels - comes with some complexity.  

There are a few different ways you might want to think about aggregating:  

• What you assume about the flighting becomes very important, so what are some sensible approaches? Assuming a flat laydown over time is perhaps the simplest to implement but doesn’t generally reflect how media is planned in practice. A more realistic approach would be to take the flighting for some prior period (eg. last year) and up-and-down-weight the investment levels with the same underlying pattern to simulate the outcome.

• The extra complication is that if the weeks on-air change that can also make a significant difference – spending £100k to upweight existing weeks will have a different impact to adding another week on-air.  

• What should you assume about a) the media mix and b) the flighting?  

• You could base it on an optimized media mix (i.e. the best outcome that could be achieved at each level of spend) or keep the same mix and up-and down-weight each channel in the same proportions. 

The importance of sense-checking your curves 

Sense-checking response curves isn’t just a good idea - it’s crucial for ensuring your Marketing Mix Model delivers reliable, actionable insights. Too often, reviewing the curve shapes can be de-prioritised in the model-building process as modelers focus on fitting the sales data and ensuring the contributions from each driver look reasonable. When that happens, there’s a risk that the resulting curves, while technically sound, can be unintuitive or even misleading. 

One key issue is extrapolation. A response curve might accurately reflect the relationship between spend and results within the range of observed data, but when you start extrapolating beyond that, things can get risky. Pushing spend well beyond the levels historically seen in your dataset could lead to unreliable predictions. Always take a step back and consider the big picture - what happens when you apply these curves beyond the data range? Does the projected impact align with what you know about your market and business? 

Another pitfall is when curve shapes don’t reflect the real-world constraints of media buying. It’s not helpful if your MMM suggests a dramatic increase in spend for a channel where inventory simply isn’t available. For example, a curve might indicate that an extra £500k in Paid Social spend will yield massive returns, but if you can’t secure that ad space due to limited inventory, those insights are irrelevant. 

Without sense-checking, you might end up with a plan that looks great on paper but fails in execution. It’s essential to ensure your MMM isn’t just technically sound but grounded in practical, operational realities. Curve shapes should reflect both the historical data and the limitations of the real-world marketplace, otherwise, you're creating scenarios that can't be actioned. 

In short, sense-checking ensures that the outputs of your MMM are not only plausible but practical.  

What’s changed since your model was estimated? 

Marketing environments are dynamic. Your MMM reflects historical data, but what if things have changed? Did you launch a more effective TV campaign than the one used in your model? You may need to tweak parameters before applying the curves for future planning. 

The basic requirements of a robust optimiser 

An optimiser should do more than just spit out numbers - it needs to be able to handle complexity. A good optimiser should allow for: 

• Handling constraints: Minimum/maximum/threshold budgets by channel, geography, or time period. 

• Scenario planning: Simulating multiple outcomes to see the impact of different spend levels. 

• Time-based optimization: Balancing immediate payback with longer-term gains. 

• Interdependencies: Accounting for the halo effect and cannibalization between channels and products. 

The real-time data challenge 

One of the biggest frustrations with MMMs is how quickly they become outdated. Traditional MMMs might be refreshed once or twice a year, but marketing decisions are made continuously. By the time you get your results, they might no longer be relevant. 

To truly make MMMs actionable, you need a scalable data management system that automates data flows, updates models in real-time, and integrates those results into your planning tools. The future of MMM lies in this ability to reforecast and re-optimise dynamically, as the market evolves. 

The future is full-cycle forecasting 

Imagine a world where you can: 

• Turn next year’s plan into a living forecast. 

• Track progress and immediately see what’s working and what’s not. 

• Re-optimise and re-plan throughout the year, adjusting spend dynamically across brands, channels, and markets - all in near real-time. 

This isn’t a pipe dream - it’s where MMMs are heading. 

Applying MMM output for forecasting and optimization requires more than just a strong model. It demands an understanding of response curves, time effects, and the flexibility to adapt as conditions change. The tools you use need to be nimble, empowering you to make decisions on the spur of the moment - let’s face it, waiting for an annual update is a thing of the past. 

With the right approach, you can turn your MMM from a retrospective tool into a forward-looking, strategic asset that drives both short-term gains and long-term success.