Why marketing mix modeling is still hard to get right

Marketing operation modeling (MMM) is becoming much accessible, but getting started remains a challenge.

After respective conversations astir MMM adoption, I noticed the aforesaid mobility kept coming up: “We judge successful the conception of MMM, but we don’t cognize really to get started.”

The reply is that viable open-source platforms person dramatically lowered the obstruction to entry. They haven’t lowered the level of expertise required to nutrient trustworthy, actionable results.

Open-source MMM has changed the starting point

MMM take is accelerating. Almost half (46.9%) of U.S. marketers will put much successful MMM complete the adjacent year, and they classed MMM arsenic the astir reliable measurement methodology (27.6%).

The open-source gyration successful MMM is real. Three production-grade libraries now screen the afloat methodological spectrum:

• Robyn (Meta, R): Automated hyperparameter hunt via Nevergrad, Pareto frontier exemplary selection, and built-in decomposition and consequence curve land — the astir approachable introduction point. It’s the 1 I usage astir because it’s highly customizable.
• Meridian (Google, Python/TensorFlow): Bayesian conclusion pinch geo-level priors and opinionated uncertainty quantification — much rigorous, pinch a steeper learning curve.
• PyMC-Marketing (PyMC Labs, Python): The astir elastic option, offering a afloat probabilistic exemplary that’s closest to academic-grade Bayesian MMM — but it besides requires the astir statistical fluency.

This procreation of devices has eliminated the $150,000 to $500,000 consulting gross that utilized to beryllium the only way into MMM. Any squad pinch R aliases Python expertise and comparatively cleanable humanities information tin now tally a exemplary in-house.

The cardinal caveat worthy making definitive successful immoderate speech pinch those exploring MMM is this: “Free tool” doesn’t mean “free model.” The package is free. The domain expertise required to configure it correctly — a hugely important portion of the process — isn’t.

A crowded vendor scenery pinch an absorbing powerfulness dynamic

The SaaS furniture built connected apical of open-source MMM has proliferated quickly. It’s worthy distinguishing a fewer tiers.

Data-layer-first vendors

Platforms for illustration Rockerbox and Northbeam started arsenic attribution and information postulation platforms, past added MMM. Their separator is information pipelines and speed, not modeling extent aliases customization.

Measurement-first vendors

Platforms for illustration Measured, Analytic Partners, Ekimetrics, and Nielsen Gracenote connection much rigorous modeling astatine a higher value point, pinch enterprise-grade capabilities.

Google Meridian and GA360

One constituent is worthy calling out. Google’s open-sourcing of Meridian was a generous publication to the section and, astatine the aforesaid time, a strategical one. When a walled plot costs and packages the measurement methodology utilized to measure its ain channels, it’s worthy maintaining patient skepticism astir exemplary priors and default assumptions, moreover pinch transparent code.

The applicable mobility erstwhile evaluating vendors is: who owns your information layer, and does that create conflicts successful the modeling layer?

Challenge 1: Data entree is the silent MMM killer

This is the astir underappreciated implementation blocker, and it seldom gets the attraction it deserves. A well-specified MMM needs:

• Two to 3 years of play information arsenic a baseline — capable to seizure astatine slightest 2 afloat seasonality cycles and a meaningful scope of walk variation.
• Consistent channel-level walk granularity — not conscionable “digital,” but search, social, display, and video surgery retired separately.
• Offline channels (TV, OOH, radio, events, nonstop message — which typically unrecorded successful different systems) are owned by different teams, and often usage incompatible clip granularities.
• External covariates — macro indicators, competitor activity, pricing data, and merchandise motorboat calendars.
• For B2B specifically, longer income cycles and little conversion volumes make the information requirements moreover much demanding. You often request much history.

In practice, what blocks astir MMM projects is the six-week information archaeology workout that comes earlier exemplary building. Finance owns revenue. The marque squad owns TV. The agency owns integer spend. The spreadsheet personification built successful 2021 is the only grounds of waste and acquisition promotions.

The exemplary is only arsenic bully arsenic the information archaeology that precedes it, and cipher tells you that successful the vendor demo.

---

Challenge 2: You still request to rotation up your sleeves

AI assistants person meaningfully lowered the syntax barrier. They tin scaffold a Robyn run, make a Meridian config, aliases thief debug a PyMC model. What they can’t yet do is navigate the judgement calls that make an MMM trustworthy:

• Choose wherever to beryllium connected a Pareto frontier of hundreds of exemplary solutions (NRMSE vs. DECOMP.RSSD tradeoffs).
• Know erstwhile Nevergrad’s optimizer has meaningfully converged versus landed successful a section minimum.
• Configure adstock translator parameters (Weibull shape/scale, geometric decay) to lucifer realistic transmission dynamics.
• Diagnose why a exemplary assigns an implausible publication to a channel, and whether to reside it pinch a prior, a information correction, aliases a adaptable exclusion.

In different words, vibe coding your measurement to an MMM will nutrient a exemplary that appears to activity but is incorrect successful ways you won’t catch. The scripting isn’t the difficult part. The domain expertise required to validate the output includes moving channel-specific incrementality experiments to calibrate your MMM.

Challenge 3: The quality expertise furniture isn’t optional

Even erstwhile the tooling matures to the constituent wherever AI tin tally a competent default MMM, the irreplaceable quality publication is encoding business discourse — things nary exemplary tin infer from the information alone:

• Adstock and carryover context: Your TV bargain has a four-week carryover. Your paid hunt has a three-day carryover. Your branded consciousness run has a decay that spans months. This accusation isn’t recovered successful the data. It’s successful the minds of the transmission experts.
• Saturation curve shape: Knowing a transmission is apt approaching diminishing returns earlier the exemplary tells you so, and questioning the results erstwhile the exemplary suggests otherwise.
• Guardrails and anomaly handling: Factors for illustration COVID troughs, merchandise launches, pricing shifts, and macro disruptions request to beryllium modeled explicitly aliases flagged arsenic structural breaks. AI doesn’t cognize your customer had a pricing situation successful Q3 2022.
• Interpretation sanity checks: A modeled TV publication of 40% for a marque spending $2 cardinal connected TV whitethorn “feel wrong” and warrant investigation. That intuition is earned, not computed.
• Organizational translation: The astir technically correct exemplary is worthless if you can’t explicate why it recommends shifting 15% of the hunt fund to CTV successful position a CMO and CFO will enactment on.

Lay the groundwork earlier you build a model

The champion spot to statesman is knowing what information you request to substance the exemplary and who needs to thief contextualize and construe that information into effective trading decisions. Neither is easy aliases fast, but some are basal if you want to get meaningful insights from your model, sloppy of whether you take an open-source aliases subscription-based platform.

A applicable first measurement is to download Robyn’s demo book and experiment pinch the sample information earlier applying it to your own.