Marketing Mix Modeling Insights

The Future of App Measurement: How MMM Drives Growth and Optimizes Budgets  As privacy regulations tighten, traditional methods of marketing measurement are becoming less effective, creating a need for more robust solutions. **Marketing Mix Modeling (MMM)** offers a comprehensive approach to understanding how marketing spend drives business outcomes, providing the insights necessary for growth and budget optimization. The Shift in Marketing Measurement Recent privacy regulations, such as Apple’s App Tracking Transparency (ATT) and GDPR, have limited data collection and hindered traditional attribution models. These models, which rely on user-level tracking, are no longer sufficient in today’s privacy-conscious landscape. As a result, businesses need to explore new methods for measuring marketing effectiveness, and **MMM** is quickly emerging as the solution. Marketing Mix Modeling is a statistical method that aggregates data from various sources, such as sales data, marketing activities, public competitor information, and macroeconomic factors like seasonality. By using MMM, app businesses can gain a comprehensive view of how their marketing spend and activities drive sales and optimize budgets for sustainable growth. Why MMM Is Essential for App Businesses MMM provides three key advantages for app businesses: 1. Sustainable, Privacy-Centric Measurement Unlike traditional attribution models that rely on individual-level data, MMM uses aggregated data like advertising impressions, spend, and industry trends. This approach is immune to privacy regulations, ensuring that businesses have reliable and future-proof data for measuring marketing performance. As privacy concerns grow, MMM offers a sustainable method for tracking and optimizing marketing efforts. 2. A Holistic View for Informed Decisions MMM evaluates all factors that influence revenue, including media spend, product updates, seasonality, and even competitor activity. This comprehensive approach provides a clear understanding of what drives user acquisition and engagement, enabling businesses to make well-informed, data-driven decisions about where to focus their marketing efforts. 3. Data-Driven Budget Optimization MMM doesn’t just provide performance data; it also offers recommendations for optimizing marketing budgets. By analyzing the effectiveness of various marketing activities, MMM helps businesses allocate resources more efficiently to maximize ROI. This enables companies to reduce wasted spend and ensure that every marketing dollar is contributing to growth. Real-World Examples: How MMM Drives Growth Aloha Factory, a Korean gaming developer, used MMM to adapt its marketing strategy to privacy changes and optimize its growth. By analyzing media impressions and contextual factors like weather and COVID-19, the company found that 37% of its app installs came from Google App campaigns—much higher than the 13% predicted by last-touch attribution. With this insight.

We just merged AI skills for Bayesian MMM into the PyMC-Marketing repo 🎉 ! If you've tried asking an LLM for help with media mix modeling, the advice is pretty surface-level. It doesn't know that your priors should reflect spend shares, or that adstock carryover changes how you compute ROAS, or when time-varying coefficients actually matter versus when they just overfit. So we wrote skills that bake in that kind of domain knowledge. They plug into Cursor, Claude, and similar tools, and walk you through everything from setting up the model to running budget optimization. They ship with the library, so they won't drift out of date. The thing we kept running into: people have the data and can install the package, but the hard part is knowing which modeling choices matter. Priors, saturation curves, and lift-test calibration are where MMMs succeed or quietly fail. The skills try to make those decisions less opaque. I am curious what others find trickiest. Try them out and give us feedback! https://lnkd.in/d6EMiiz5 #PyMC #MediaMixModeling #BayesianStatistics #MarketingAnalytics #OpenSource

MMM is having a moment…and a misunderstanding. More brands are adopting marketing mix modeling, which is great to see. But there’s a growing problem in how teams are using it. MMM tells you what’s incremental. It doesn’t tell you what’s fixable. I’ve seen too many brands look at their model, see a channel underperform, and immediately write it off. “TikTok isn’t incremental.” “CTV doesn’t work for us.” “Meta’s not driving lift.” So they cut spend. Move on. And then run out of channels that “work.” That’s not what MMM is for. The question isn’t “Does this channel work?” It’s “Why isn’t it working for us?” Maybe the creative’s wrong for the format. Maybe the audience targeting is off. Maybe your measurement window is too short to see the impact. MMM should spark curiosity, not closure. Before cutting a channel, ask:  • Has this channel worked for others in my category?  • Is my customer active there?  • Have I truly tested the right variables? Because if you keep taking MMM results as gospel, you’ll optimize yourself into a corner and run out of ways to grow.

MMM vs MMM. 𝗠𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝗠𝗶𝘅 𝗠𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴 𝘃𝘀 𝗠𝗲𝗱𝗶𝗮 𝗠𝗶𝘅 𝗠𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴. Same thing? Not really. Not the same in terms of data inputs, model structures or what you are trying to learn. 𝗠𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝗠𝗶𝘅 𝗠𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴 typically looks at 𝗮𝗹𝗹 𝗺𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝗹𝗲𝘃𝗲𝗿𝘀 to understand what is driving incremental growth across the business.  The data involved is very specific to each business, and covers areas such as pricing, promotions, distribution, seasonality, economic performance, competitors and external factors (Eg weather). Paid marketing is a key input but the models aim to identify the impact of paid marketing with respect to the impacts of these other factors. 𝗠𝗲𝗱𝗶𝗮 𝗠𝗶𝘅 𝗠𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴 only looks 𝗽𝗮𝗶𝗱 𝗺𝗲𝗱𝗶𝗮 𝗰𝗵𝗮𝗻𝗻𝗲𝗹𝘀. All those other factors are not actively considered. This limits the data collection scope, but speeds up data availability and allows the models to focus on most recent Paid Marketing performance. So if I only care about paid marketing optimisation, 𝗱𝗼𝗲𝘀 𝘁𝗵𝗲 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗺𝗮𝘁𝘁𝗲𝗿? In short, yes.  Because without including those other factors you don't actually know whether an increase in business performance was caused by the 10% increase in paid marketing spend, or the 15% increase in market demand. In a lot of cases the directional insight "where is the best place to spend the next $" will be aligned between both Marketing Mix and Media Mix models.  But in several cases it won't be:  • Media Mix Models 𝗼𝘃𝗲𝗿 𝘃𝗮𝗹𝘂𝗲 𝗽𝗮𝗶𝗱 𝗰𝗵𝗮𝗻𝗻𝗲𝗹𝘀 which increase spend at high category demand, and during discount periods (eg Search, Shopping, Affiliates)  • Media Mix Models will 𝘂𝗻𝗱𝗲𝗿 𝘃𝗮𝗹𝘂𝗲 𝗽𝗮𝗶𝗱 𝗰𝗵𝗮𝗻𝗻𝗲𝗹𝘀 which spend in buildup to sales events or product launches  • Media Mix Models will struggle to isolate the 𝗶𝗺𝗽𝗮𝗰𝘁 𝗼𝗳 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗲𝗳𝗳𝗲𝗰𝘁𝘀. If you suddenly get featured in high reach press or creator content, then Marketing Mix Models will identify this while Media Mix Models will not. The marketing industry is terrible with acronyms, it really doesn't help when two measurement methods share the same one. So next time you hear MMM, make sure you check with flavour.

The Econometric Illusion: When Marketing Models Mislead In today's data-driven marketing world, econometrics promises to unveil hidden connections between media spend and sales. But as we delve into regression analyses and p-values, a troubling question emerges: Are we building strategies on statistical quicksand? The shift from gut feelings to data-driven decisions in media planning has been revolutionary. Marketers now confidently declare that a 10% increase in TV spend will yield a 2.3% uplift in sales. These precise figures, presented as scientific fact, can make even seasoned professionals feel like soothsayers with supercomputers. But many of these pronouncements rest on shaky statistical foundations. At the core is the misuse of statistical significance, particularly p-values. A p-value below 0.05 often transforms tenuous correlations into "significant findings" driving million-pound decisions. This fixation has led to a replication crisis in marketing research, mirroring issues in psychology and medicine. Another pitfall is mistaking correlation for causation. When models show a strong relationship between social media engagement and sales, it's tempting to assume direct causation. But marketing is a web of interconnected factors, and apparent causal relationships may be spurious correlations. Even with sound methods, data quality remains a challenge. Capturing clean, comprehensive data on consumer behaviour is Herculean. External factors like economic downturns can derail sophisticated models. The final hurdle is interpreting and implementing insights. Overconfidence in model outputs can lead to rigid strategies ignoring inherent uncertainty. Focus on marginal gains can distract from transformative initiatives. Consider these cautionary tales: - An FMCG company increased TV ad spend based on models showing strong sales correlation. Failing to account for diminishing returns, they wasted millions on ineffective advertising. - A bank shifted budget to digital channels, missing the crucial role of traditional media in building trust. This led to a decline in valuable long-term customer relationships. - A supermarket optimized promotions using models that didn't account for cannibalization, eroding overall category profitability. Should we abandon econometrics? No. We need a more nuanced approach: - Embrace uncertainty. No model captures full market complexity. - Foster statistical literacy among marketers. - Prioritise data quality and comprehensive measurement. - Reaffirm the value of human judgment alongside data. The future of marketing econometrics lies in understanding the interplay between data, analysis, and strategic thinking. By combining rigorous analysis with strategic insight and creativity, we can develop strategies both data-informed and attuned to human complexity.

Here's the uncomfortable truth: You can build thousands of different MMMs where all of them fit your data perfectly. But they'll give you 1,000 different answers about what works and what to do next. The more complex a model is, the easier it is to perfectly fit a given data set. And MMMs are extremely complex, so it’s pretty trivial to generate a model that fits the data well, and there will be a huge number (approximately infinite) of candidate models that fit the data equally well. For this reason, measures of in-sample model fit like R-squared or MAPE simply aren’t useful for determining if a media mix model is “good” or not. You need external validation – proof that following the model actually improves outcomes – before you start following its forward-looking recommendations. Here are the options: → Out-of-sample forecasting: hold out outcome data from future weeks and check if the model can accurately predict what happens to the business. → Backtesting decisions: simulate what would’ve happened if you’d followed past recommendations. → Live intervention tests: change your budget based on model guidance and see if the model remains stable and accurate. Sure, MMMs train on historical data. But the entire reason we build these models is to decide what to do next. If you're not using your MMM to make better decisions tomorrow, you risk engaging in expensive measurement theater. In order to use your MMM, you have to know if it can actually make good recommendations about what to do in the future. To do that, you need to consistently verify that the forecasts continue to be correct. By actually testing real predictions that can be measured in the real world, you have a falsifiable hypothesis that can be proved or disproved, and that’s what really builds trust.

Your MMM might be lying to you. A common issue in marketing mix modeling (MMM) is determining what assumptions and configurations you're baking into the model. • Inputs and outputs • Granularity (daily vs weekly) • Adstock and carryover • Dummy metrics and confounds And most germane to this post: the shape of the time-based saturation curve. Many MMM providers and tools bake this assumption in as a default setting, and can give you WILDLY different results based on if you go in and manipulate it or not. For example: • Day 1: peak impact, 10% of all impact • Day 2: slightly off peak, 9% of all impact • Day 3: reduced again, 8% of impact And so on... You get a curve that looks like a ramp, with the peak on the left (starting at day 1), and a steep drop off over time. If you've run any series of incrementality experiments (such as a matched market test), you know that very often channels show NO impact the first day(s) or even weeks, and then all of a sudden the incremental impact starts to show up. (Side note: across our thousands of tests, we see an average MINIMUM of 22 days to see statistically significant incrementality). This means that if you run an MMM including a channel like CTV, which is very likely to have limited lift in the first days or weeks of a campaign, you'll significantly UNDERESTIMATE the impact and incrementality of that tactic. Compare that to something like a Weibull curve, which has this more bellcurve or hill curve shape, and you get a significantly different incremental return on that ad spend. In this illustration below, you can see the iROAS is 3.1x on a Weibull curve, and only 1x on a standard monotonic curve. That makes a channel go from terrible to amazing, and can have dramatic consequences on the business. Why does this happen? • Impact isn't the same every day • Impact takes time to "build up" in an audience • Impact continues to survive after the peak • Impact being measured opposite to how it functions gives you a weaker formula, weaker relationship, and overall lower result How to fix: • Ensure that you understand the assumptions in your MMM • Where possible, deploy more "realistic" curves by tactic/funnel • Potentially hack or offset your data if your tool doesn't do this • Validate and calibrate your models with incrementality experiments The number 1 mistake I see brands making when going into MMM and incrementality land is underselling their top of funnel channels due to bad experiments, bad stats, and a fundamental misunderstanding of how long it takes for marketing to "work." #mmm #incrementality

If your MMM vendor says they can isolate the lift from individual creatives, don’t be afraid to press on this. Ask them: - Are you using actual modeled lift estimates for each creative, or just allocating channel-level lift proportionally by spend? - What assumptions are you making to estimate creative-level impact, and can you show me how sensitive your results are to those assumptions? - Can you show a prior-predictive or posterior-predictive check that validates creative-level signal detection? Media mix models are top-down models. They operate on aggregate outcomes (revenue, conversions, etc.) and aggregated inputs like spend or impressions. For the model to detect lift, the creative has to move the needle more than the system’s natural noise. And that bar is higher than most people think. In most cases, a new creative variant (a new background color, a new headline, or a different CTA) doesn’t drive a big enough delta in sales to stand out above baseline noise. So when a model spits out “Creative A had 3.4x ROI and Creative B had 1.9x,” you should ask: did those creatives actually drive statistically distinct outcomes? Or did the model just estimate that they did? Sometimes vendors fudge it: dividing lift by spend, or filling in gaps with untested priors. That can be fine as long as they’re honest about what’s being modeled and what’s being assumed. But don’t confuse assumptions for evidence. And don’t let a modeler (Recast or anyone else!) tell you they’re seeing something the data can’t actually support.

I simulated data and fit a Bayesian hierarchical MMM to test how multicollinearity impacts model accuracy. Without highly informed priors, adding a second correlated media channel distorted total ROAS estimates and obscured the true performance gap between channels. Marketing Mix Models (MMMs) are powerful, but they can struggle when media channels are highly correlated—leading to misleading ROAS estimates and suboptimal budget allocation. 🎥 Watch here: https://lnkd.in/eMurXGcb In this video, I break it down: ✅ Simulating MMM data with known parameters ✅ Seeing how correlated media spend distorts estimates ✅ Implications for media budget decisions In upcoming videos, I'll test more informed priors and explore Google Meridian as a potential solution. Follow and engage to stay ahead of the latest insights on MMMs, causal inference, and data science. 🙏🚀 #datascience #datascientist #marketinganalytics #causalinference #datasciencetutorial #marketingmixmodel #mmm

Friday morning client call. Their head of growth asked why I wanted to cut their "best performing" ASC campaigns. The Facebook dashboard showed ASC driving 70% better efficiency than manual campaigns. Every metric screamed "winner." But here's what happened when we ran marketing mix modeling on their data. Those same ASC campaigns? Actually less efficient at driving incremental sales. The manual campaigns were quietly outperforming by 20-30% in true business impact. The problem isn't ASC itself. It's running multiple ASC campaigns across different ad accounts for competing products. Each campaign claims credit for the same purchase. Platform attribution becomes a hall of mirrors. Health brands face this constantly. Someone searches for "gut health" but buys a "weight management" product. Six different ASC campaigns take credit. Dashboard metrics look fantastic. Real incrementality gets cannibalized. Most media buyers optimize for what platforms show them. Smart ones optimize for what actually moves business metrics. The difference? We've seen 40% gaps between attributed efficiency and incremental efficiency. That's not measurement error. That's real money flowing to the wrong campaigns. Before your next budget allocation meeting, ask this: → Are you scaling campaigns that look good or campaigns that actually work? Platform dashboards tell you what happened. MMM tells you what you caused. Have you noticed similar gaps between your attribution and incrementality data?