There is a myth in MMM circles that Multiplicative MMM is better. Somehow people confuse the mathematical sophistication for relevant application.

Actually two fundamental questions alone create this fork:

Do marketing / media variables contribute an absolute number of sales (or any KPI) -Go Additive Modeling
Do marketing / media variables proportionally scale sales (or any KPI)? - Go Multiplicative Modeling

The Additive MMM

Yt​ =β0​ + β1​TV + β2​Meta ​+ β3​Search + β4​Promo +β5​Seasonality ​+ε

Here each channel contributes a fixed incremental amount to sales.

Each channel literally pushes sales upward.

Suppose

Base = 40
TV = 20
Meta = 20
Search = 10
Promotion = 10

Total =100

Everything is additive.

The Multiplicative MMM

t ​= α × TV^β1​​ × Meta^β2 ​​× Search^β3 ​​× Promo^β4..

In this example, instead of channels adding sales, each channel scales sales proportionally.

TV might increase sales by 10%.

Search might increase them by another 5%.

The effects compound. This naturally produces elasticities.

Why Multiplicative models are applied?

I would say predominantly due to statistical misconceptions and also not thinking clearly about how the marketing reality.

One of the statistical reasons was to capture the adstock effect.

Modern additive MMMs don't model raw media spends. They model adstocked spends. The final model remains additive and yet it already captures Non linear carryover effects and diminishing effects.

The Elasticity Argument

Multiplicative also gained prominence because it readily gives you the elasticity once you take the log. But the same can be done via additive models too.

The Synergy Modeling Argument

Many people think Multiplicative models model the synergies better. They Don't.

Suppose TV amplifies Search.

An additive model simply becomes TV × Search, without assuming that every marketing variable somehow interacts with every other variable.

A multiplicative MMM assumes every variable acts as a scaling factor on the entire system, rather than contributing independently !!

It doesn't distinguish between relationships that are marketing reality wise plausible and those that are not.

For MMM, the usual goals are:

- Estimating incremental contribution
- Understanding carryover/saturation
- Capturing a handful of genuine channel interactions,
- Optimizing budgets.

Modern additive MMMs with adstock transforms and explicit interaction terms address all of these directly, while producing a decomposition that marketers can interpret and act on.

I would say 90-95% of commercial MMM engagements do not gain enough from a multiplicative model.
It is often introduced because it sounds more sophisticated than because the underlying business problem genuinely requires it !!

Sometimes the best model isn't the one with the fanciest equation.
It's the one that answers the business question with the fewer assumptions.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more - https://www.aryma.ai/

One challenge in Marketing Mix Modeling (MMM) that rarely gets discussed is temporal disaggregation.

So what is Temporal Disaggregation?

Temporal disaggregation is the process of converting a low frequency time series (quarterly / half yearly or annual) into a higher frequency one (monthly or weekly), while ensuring the higher frequency numbers still add up exactly to the known totals.

Why this is needed?

In MMM, business data, brand equity data and some macroeconomic data are rarely collected at the same frequency.

▪️Macroeconomic Variables

Many macroeconomic indicators such as GDP, unemployment rate and inflation are often published quarterly or monthly.

But an MMM may operate at a monthly or even weekly granularity.

Dropping these variables altogether means losing potentially important explanatory signals.

▪️Brand Equity Tracking

Measures such as TOMA, SPONT, Brand Consideration are (still) collected through surveys every quarter.

However, the MMM may require monthly observations.

▪️Campaign Spend Reconciliation

Sometimes the marketing team can only provide quarterly campaign spends for variables like OOH.

But the monthly allocation has been lost, aggregated, or inconsistently reported across systems.

This is surprisingly common in long-running marketing programs.

Hence again the need for temporal disaggregation.

The Denton - Cholette Method

At Aryma Labs, we use Denton - Cholette method along with extra logic added for temporal disaggregation.

Rather than splitting the quarterly value equally across months, Denton Cholette uses an indicator series to distribute the values intelligently while satisfying two important conditions:

• The monthly values add up exactly to the known quarterly / half yearly total.

• The monthly movement follows a reasonable underlying indicator (for example search interest, impressions, media activity, GDP movement) without introducing unrealistic jumps.

🛑 Caveats:

The Denton - Cholette method (DCM) does not recover the TRUE historical values.

Instead, it produces the most statistically plausible high-frequency series, given the available information and constraints.

If the original monthly observations exist, those should always be used.

DCM is not perfect but these methods quietly improve data quality before a single MMM is ever fit.

Good MMM starts long before model building- it starts with respecting data constraints.

Aryma Intellisheet comes equipped with DCM method. Check out the full demo here.

Check out our website here - https://www.aryma.ai/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/

Most discussions around AI Agents focus on what the agent can do. While strictly speaking, the AI Agents don't have agency yet. They get the agency or pretend to have agency by virtue of what it is being prompted to do or not do.

Which brings me to the topic of How we at Aryma Labs train our MMM AI Agents.

We follow a philosophy called 'Via Negativa'. It is basically improvement through subtraction rather than addition or just 'What not to do'.

For many complex domains, it is easier to identify what is wrong than to precisely define what is right.

And MMM is a perfect example :)

I will cite few examples from our prompt templates to illustrate the principle of 'Via Negativa'.

📌Example 1

An MMM model shows:

TV Contribution = 35%

Paid Search Contribution = 20%

An poorly trained AI agent may immediately conclude:

"TV is the most effective channel. Increase TV budgets."

A good MMM practitioner knows that this conclusion lacks nuance.

The agent should first ask:

• What are the current spend and saturation levels?

• What are the marginal ROAS ?

• Is TV already operating near diminishing returns?

Hence Via Negativa prompt is:

"Do not recommend budget increases solely because a channel has high contribution".

or

"Do not recommend reallocations without checking saturation, spend constraints and implementation feasibility."

📌Example 2

AI Agent : OOH only contributed only 1%, let's cut it.

The agent should know:

- OOH often acts as an amplifier and could manifest its effect through other channels.

- OOH may create long-term brand equity.

Via Negativa Prompt: "Do not recommend cutting channels before examining interaction effects and halo effects."

📌Example 3

AI Agent: Base is the top driver of your sales or seasonality is the most efficient channel

The agent should know that:

Base is not technically an active driver of your sales. It is the organic sales or brand equity. The sales that you get even when you have zero marketing or media spends.

Base hence can't be treated like a media channel. It is basically a manifestation of your previous marketing efforts (see related post in comment).

Similarly an agent should not construe seasonality as an active driver of sales. It is just an enabler or catalyst at best.

Via Negativa Prompt: Don't treat Base and seasonality like other marketing channels.

📌 The future of MMM AI Agents

The most knowledgeable agent will not necessarily be the one that has read the most MMM content.

It will be one that has accumulated the largest collection of 'Not to Do':

• Things not to conclude / infer / recommend

• Things not to optimize

In reality, this is how experienced MMM consultants also operate.

They develop a mental catalogue of mistakes they no longer make. We certainly have such a codified catalogue.

The next frontier of MMM AI is not teaching agents more MMM. It is teaching them more MMM restraint.

Check out our website here - https://www.aryma.ai/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/

I have often noticed that Attribution and contribution are often used interchangeably in marketing measurement, but they are fundamentally different concepts.

Here is my POV:

Attribution is about who get's the credit

Attribution attempts to assign credit for a conversion or sale to specific marketing touchpoints.

For example:

let's take a customer journey:

Google ad -> FB ad -> ChatGPT research (yup this is happening nowadays) -> Purchase

A last click attribution model would say:

ChatGPT = 100% credit

A linear attribution model might say:

Google = 33.3%

Facebook = 33.3%

ChatGPT = 33.3%

Attribution hence is largely a 'finger pointing' exercise. But it doesn't answer two major things - Incrementality and Causality.

This bring me to Contribution.

Contribution

Contribution tries to estimate the incremental impact of a marketing activity on the outcome. And when done via MMM, it also helps answer the causality question to a certain extent.

Related post link in comments.

For example:

Suppose a brand sells 100,000 units.

MMM might estimate:

Base sales = 40,000 units

TV contribution = 20,000 units

FB Ads = 10,000 units

Google Ads = 10,000 units

ChatGPT = 10,000

Promotion = 5,000

Seasonality = 5,000 units

Contribution hence answers two questions:

"Who contributed to sales" and also "How much did each of the factors contribute"

That's why many MMM practitioners would argue that MMM is fundamentally a contribution model, while MTA is fundamentally an attribution model.

To summarize:

Attribution = allocation of credit.

Contribution = allocation of credit + estimation of impact.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/

AI adoption in Marketing Mix Modeling (MMM) is accelerating rapidly.

But much of the discussion remains superficial and lacks technical depth.

As a former NLP engineer and as a company that began applying AI to MMM as early as 2024 through MMMGPT, I thought I will unpack the different AI architectures and explain why we believe RAG is the most suitable approach for MMM (at least for now).

Frontier Models

A frontier model is a large foundational model trained from scratch on enormous amounts of data and compute.

Examples - GPT, Claude, Gemini etc.

Training one requires:

• Massive datasets

• Huge GPU clusters

• Hundreds of millions (sometimes billions) of dollars

For a niche domain like MMM, building a frontier model makes little sense.

MMM knowledge is relatively small compared to the internet. You do not need a trillion-parameter model to understand adstock, saturation, contribution analysis and budget optimization.

One needs depth, not scale.

Fine-Tuning

Fine-tuning takes an existing foundation model and further trains it on domain specific data.

The flow is like:

General LLM -> MMM Training Data -> MMM Fine Tuned Model

This changes the model's weights and can improve domain understanding.

However, MMM is not merely a collection of concepts and examples.

Much of the value resides in:

• Client decks

• Consulting notes

• Historical projects

• Business context

• Organization specific methodologies

Fine-tuning struggles when knowledge evolves. Every new project, spend pattern or methodology may require retraining, making it difficult to maintain.

Distillation

Distillation trains a smaller model (student) to imitate a larger expert model (teacher).

The Flow is :

Expert Model -> Generates Knowledge -> Student Model Learns

This helps reduce cost and latency.

However, a distilled model only knows what existed at training time. It cannot automatically access your latest MMM projects, new learnings or evolving methodologies.

The MMM Reality: Why Aryma Labs believes in RAG architecture

At Aryma Labs, our architecture is:

LLM -> RAG Layer -> Proprietary MMM Repository -> MMM Applications / Products

- The LLM provides language and reasoning.

- The RAG layer provides retrieval.

- The repository provides domain expertise.

Our repository contains:

• MMM consulting notes

• Historical MMM projects

• Spend and effect-share patterns

• Optimization studies

• Validation frameworks

• Experimentation and causality learnings

The model does not need to memorize this knowledge. It simply retrieves the right information at the right time.

The future moat in MMM AI will not be who trained the biggest model.

It will be who built the richest and trustworthy domain memory.

When people ask with surprise what model powers our products like Aryma Deck, Singularity Nebula etc.

Our answer is:

It is not the model. It is the data and domain knowledge 😎

Check out our website here - https://www.aryma.ai/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more - https://www.aryma.ai/

I am a former NLP Engineer. Yes the very word NLP now sounds pretty 'Dinosauric'. I can't believe within a decade we have moved from one hot encoding as embedding model to transformers-on-Steroids model.

At Aryma Labs, apart from building MMM and Marketing Experimentations, we build cutting edge Gen AI products focused on Marketing Measurement. Almost all our team uses Claude or Cursor to enhance their code.

Claude Code is genuinely impressive.

We refactored legacy modules in days instead of weeks. It suggested optimizations and patterns we hadn’t considered, perhaps ideas borrowed from other domains and surfaced instantly.

Which raises a deeper question:

📌 Has Claude Code Learnt Everything?

For those of you who are programmers, you might have have experienced that Claude almost has all the 'Design patterns' of most programming languages.

It may even have been trained on the public repos of so called '10x - 100x Engineers'.

These 10x Engineers (I have been lucky to work alongside few in the past), are the Da Vinci or Mozart of programming. I have seen them do functional programming while thinking how to efficienize memory utilization.

This breed of programmers are shrinking and I wouldn't be surprised if we have <10,000 of them now worldwide.

📌 10x - 100x Engineers using Claude Code

The issue isn’t average programmers improving - that’s good.

The issue is :

What happens when 10x engineers start defaulting to Claude instead of their own creative friction? Won't we eventually 'regress to the mean'?

If models are trained on yesterday’s brilliance, and tomorrow’s brilliance is guided by those models, where does true novelty come from?

📌 The Need for Galapagos Island

The Galapagos Islands fascinated biologists because species evolved in isolation. No cross-pollination. Pure adaptive pressure.

In causal terms a clean 'control'.

📌 But why do we need Galapagos Island for Gen AI?

I think it is important that the 10x-100x Engineers don't lose their edge to 'regression to the mean'.

As strange as it may sound, I believe Gen AI progress and perhaps the path to AGI even lies in such Galapagos Islands.

I Imagine every big company hiring these 10x engineers and kind of putting them inside a Galapagos Islands (figuratively of course) where they are left to their own devices and solve some of the cutting edge problem through their own innovative thinking rather than using 'Claude Code' or 'Cursor'.

From these “islands,” innovation flows outward.

Models can learn from the output but the source remains untouched.

📌 10x Engineers will be hired more

The advent of Claude code and cursor doesn't mean the end of 10x Engineers. It just means even average programmer can intermittently become a 10x-100x programmer.

The 10x Engineers will continue to be in great demand in future. The edge and novelty of the world models will depend on the quality of inhabitants in this so called 'Galapagos Islands'.

Check out our website here - https://www.aryma.ai/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more - https://www.aryma.ai/

This led us to develop Aryma Deck.

What is Aryma Deck?

Aryma Deck is an AI powered PPT add-in that deciphers, understands any MMM slides from any domain and generates expert level insights all within the current PPT you are operating in.

The Aryma Intelligence Layer

The Aryma Intelligence Layer is trained on 1000's of consulting notes, Media spend and effect pattern and most importantly how our co-founders Ridhima Kumar and Venkat Raman connect the dots between various MMM outputs, business context to generate high quality insights.

Basically we have taught the AI to think and generate insights like we do.

AI will Immortalize Experts

Every organization will eventually build AI systems that carry the DNA of its best employees.

AI trained not only on what experts know. But how they connect the dots to generate insights.

Key Feature of Aryma Deck

Aryma Deck comes equipped with 7 tools:

/analyse - analyses the current slide and generates deep expert level insights

/analyse all - analyses the entire deck and generates insights for all of them (Press this button, go grab a cup of coffee and the insights will be ready in 10 mins on all slides)

/master insights - this tool generates the way forward recommendations. We saw some Move 37 level insights already !!

/ask ppt - Query and scrutinize the insights generated

/Refine -this tools helps you summarize the insights

/ generic - ask anything about MMM

/ history - Browse every saved analysis

How to use Aryma Deck

We can help you set up Aryma Deck add-in leveraging our existing trusted and reliable intelligence layer in just 5 easy steps.

or

You can bring in your own intelligence layer (we can custom train the models on your MMM knowledge). Our AI engineers can help you design it from scratch.

Enterprise Grade Security

We know organizational knowledge is one of the most valuable assets a company possesses.

That is why Aryma Deck has been built with encryption principles in mind. Every variable, company name is triple layer encrypted and LLM doesn't see any of the real variable names.

The real benefit

Imagine a new employee joining your organization.

Instead of spending five years learning from experience, they can immediately access AI systems trained on the best thinking your company has accumulated over decades.

The gap between novice and expert begins to shrink.

Not because expertise disappears. But because expertise becomes scalable.

That is the future we are building toward.

Welcome to Aryma Deck - Deep MMM Insights, Now Democratized

Check out Aryma Deck for more details and feel free to schedule a demo: https://www.aryma.ai/products/aryma-deck

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/

Real AI Solution Providers Speak in Specifics

Recently I came across a post from a MMM vendor benchmarking their “AI capabilities” against a few other vendors.

Most of the listed capabilities were extremely surface level:

- “Conversational AI”

- “AI insights”

- “AI reporting”

- “Natural language recommendations”

At some point it felt like the same capability was being stretched into 2-3 different bullet points with different wordings 🙂.

Also each capability was rated differently (one was on a scale of 4, 5, some even on scale of 8. Obviously it has no statistical or mathematical meaning except to pad up the final score which was not even weighted 🤡!

MMM vendor making such elementary statistical mistake is shocking.

And then came the marketing theater.

One vendor who was benchmarked, pretending to be “angry” .

All roads conveniently leading the audience back to the same blog post where these vendors were benchmarked and led to an impression that only these vendors are the 'best' AI tools for Marketing Measurement 🙂.

📌 The Kayfabe

Kayfabe works in WWE (may be not the recent version but the one I grew up with 90's - 2010 did).

But Kayfabe in marketing is one of the most deceptive practice because it creates an illusion of competition while quietly steering the audience toward a pre scripted narrative.

I will write more about this in future.

Unlike the above vendor, I am not saying Aryma Labs is the only Marketing Measurement vendor with advanced AI capabilities. There could be others too.

But there is one simple telltale sign that helps distinguish real AI solution providers from AI wrapper companies:

📌 Real AI solution providers speak very specifically about:

• What exact problem they are solving

• Why the problem exists

• Why current workflows fail

• How their system changes outcomes

• What operational bottleneck gets removed

Not just:

“XYZ with conversational AI.”

At Aryma Labs, our AI systems were built around very explicit MMM problems.

▪️MMM Synapse: Solves organizational memory decay in MMM.

▪️MMM Singularity: Connects fragmented MMM outputs into one unified intelligence layer.

▪️Aryma Nebula: Solves budget optimization grounded in business reality, not just mathematical elegance.

▪️MMMGPT: Solves knowledge fragmentation across the marketing measurement ecosystem.

▪️Aryma Intellisheet: Solves repetitive EDA pain directly inside Excel where most MMM workflows still begin.

▪️Aryma Deck: Helps analysts create executive-grade MMM insights and narratives like seasoned strategy consultants.

As you can see each product has:

• Clearly defined problem

• Clear operational bottleneck solved

• Clear business outcome

This depth usually comes only when builders are deeply hands on with the domain.

If you are curious to see the depth of these AI systems in action, feel free to talk with us - https://www.aryma.ai/

You will see the depth at which we have thought through the problems of MMM.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai/

From Excel Add-ins to Gen AI: Aryma Labs Is Building Both Ends of the Future

Innovation does not always mean throwing away the past.

Sometimes the real innovation is modernizing what the industry already loves using.

Anybody who has operated in the MMM space will know that MMM is a small data problem (relatively speaking).

Despite all the hype around modern data stacks and dashboards, a huge amount of MMM data still arrives in Excel sheets.

Most MMM datasets are monthly, weekly or daily aggregates. They are manageable in size, business facing and most importantly highly collaborative.

Excel remains incredibly efficient for such workflows.

Most Clients and vendors still live inside Excel 🙂.

Excel is still the operating system of the MMM world (At least the EDA phase is).

📌 Don't go against the flow

From our experience of delivering nearly 500 MMM models across 5 continents, we realized that Excel is still a popular and useful tool.

So instead of forcing users into disconnected workflows, we asked ourselves: Why not bring sophisticated MMM EDA directly to where the data already resides?

📌 Aryma Labs' Custom MMM EDA Excel Add-in

The above led us to build an Excel Add-in specifically for MMM Exploratory Data Analysis.

With this add-in, our analysts can do the following:

- Automatically aggregate the data across any variable and by any format

- Perform data variation / spend diversity analysis

- Perform missingness diagnostics

- Outlier surfacing

- Correlation exploration

- And Finally call the logic applied in our MMM Readiness Checker tool to check the entire dataset readiness for MMM (If you haven't checked out MMM Readiness - link is in comments.)

All the above happens directly inside Excel itself !! Without maintaining fragile scripts. Without opening 15 notebooks.

And importantly...Without generating those visually generic Plotly charts.

📌 Simplicity over Complexity

A good EDA system should help the analyst think better, not merely produce more tables and charts.

This is where we believe the future becomes interesting.

At Aryma Labs, we are not choosing between: "Legacy tools" or "Gen AI". We are combining both.

We have built MMMGPT, MMM Synapse, MMM Singularity, Aryma Nebula and now soon to be launched Aryma Arena.

However, we also deeply respect why Excel survived for decades.

Tools survive when they reduce friction. Not when they merely look futuristic.

📌 The Add-in Impact

EDA processes are complex. What used to take our data scientist few hours, now takes them few minutes. The time spent on thinking 'Have I aggregated the data accurately', 'Have I checked for anomalies, duplicates in the data' can now be spent on thinking about business questions more deeply right from the EDA phase.

The future of MMM may not be replacing Excel. It may be making Excel intelligent.

Innovating the past and the future simultaneously is the philosophy we are building at Aryma Labs 😎.

If you are interested in subscribing to Aryma Intellisheet or creating your own AI powered Excel Add-In to solve the EDA problem on MMM, give us a call - https://arymalabs.com/book-a-demo/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Should we forget the older data in Marketing Mix Modeling (MMM)?

One problem in MMM that very few discuss deeply is:

How much should the past influence the present?

The industry thumb rule is to use at least 2 years of monthly data or 1 year of weekly data, mainly to capture seasonality effects reliably.

But people also recommend "More the better" for data.

But what if the oldest year behaved very differently from the current?

In year 1, lets say the brand experimented on a digital channel for the first time and experience good ROAS, low saturation effects and very high contribution. But in later years, the effect wanned a lot.

The old behaviour can quietly influence the following even in the latest year.

• Adstock decay

• Contribution shares

• Channel efficiency

📌 The Statistical Take

Statistically, in OLS regression, an old data point has no inherent decay in influence based on its temporality. All data points are weighted equally. However, where the data point starts having influence on downstream data points is based on two things:

1) Leverage

2) Residual size

Leverage:

Imagine a lever for convenience. Few years back I also wrote a post on Moments in Statistics.

If you a have lever, you would notice that it is far easier to push down or pull up the lever from the edges, assuming the fulcrum is right in the middle.

Now imagine a linear regression. For simplicity, a regression straight line (though in reality it can easily be any squiggly shape, the linearity in regression is wrt to only parameters).

Now if you have data points that are outliers in the very beginning, it can have a huge impact on the slope of the regression line. (see interactive demo).

Residual Size:

This is basically how far the regression prediction is from ground truth, in MMM we often notice that the first data point is never accurately predicted. This can lead to inflated MAPE for the whole time period.

Now that we have understood Leverage and Residual size, we get to the useful metric - Influence. Influence is defined as Leverage * Residual size.

📌 The Solutions

Rolling window

Drop old data points, keep only recent n observations. Now this could mean if you built a model using 3 years of data (2023, 24, 25, 26), one can drop the year 2023.

Cook's distance

We at Aryma Labs compute something called the cook's distance. It basically tells you how much the model changes when omitting a data point. So you get to know the influence of that data point.

Recursive least squares

This is a bit advanced algorithm. This algorithm basically allows you to selectively 'discount' the influence of old data points.

Exponential smoothing

This Weights recent observations more heavily with exponentially decreasing weights for older data.

But this may not be good application for MMM as it impacts the adstock effect.

Overall, I just wanted to spark a discussion on a technical MMM topic that rarely gets discussed. Your thoughts are welcome.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Marketing Measurement Has Its Own Schrödinger’s Cat Problem

According to Quantum Mechanics:

Before measurement, a particle exists in multiple possible states (a superposition).

But the moment you observe it, reality “collapses” into one observed outcome.

Schrödinger illustrated this paradox with his famous cat experiment.

A cat is placed inside a sealed box with a radioactive trigger mechanism.

Until the box is opened and observed, quantum mechanics implies the cat exists in a superposition: Both alive and dead.

Only observation resolves the uncertainty into one realized state.

To explain this, Many Worlds Theory (MWT) proposes:

Every possible outcome actually happens, but in separate branching universes !!

So in one branch, the cat lives. In another, the cat dies.

Now whether one agrees with MWT, there is an interesting parallel to Marketing Mix Modeling (MMM).

📌 Bayesian MMM=Marketing's MWT.

In MMM, we observe realized sales. We observe realized media spends.

The campaign has already happened.

Yet instead of treating the observed world as realized world, Bayesian MMM introduces a distribution of possible parameter realities around it.

A sort of posterior universe of possibilities. Different contribution / coefficient realities.

Marketing data is not in quantum superposition.

- The sales already happened.

- The tv spends already happened.

- The Meta / TikTok/ Google impressions already happened.

Once the outcome is realized, there is no unresolved state left to observe.

The customer did not simultaneously buy and not buy the product across parallel universes.

Even the Bayesian statistical interpretation feels wrong.

Let's take beam splitter experiment for analogy, the electron ultimately takes one realized path when measured.

A Frequentist confidence interval behaves similarly. Either the interval captures the true parameter or it does not.

There is one realized parameter value in the world.

The uncertainty lies in our estimation procedure across repeated samples, not in the parameter existing simultaneously across multiple realities.

By contrast, Bayesian credible intervals often get interpreted as if the parameter itself exists probabilistically across many possible states.

And that is precisely why I often question the philosophical motivation for taking a Bayesian approach in the first place.

📌 Frequentist MMM is better

In Frequentist MMM, we ask “What is the single best explanation of the realized world we observed?”

Not: “What are all the plausible worlds that could have generated this?” This part is same budget optimization scenario.

I also hence believe goodness of fit matter more than uncertainty theater.

Instead of probing posterior universes, I would rather ask:

▪️Does the attribution align with business reality?

▪️Are coefficient signs stable under bootstrap?

▪️Does the model generalize under perturbation?

The task is not exploring infinite possible worlds. The task is explaining this world correctly.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

We recently got the opportunity to demo MMM Synapse to a seasoned MMM expert. The person was very impressed and asked a important question:

“How is MMM Synapse different from Copilot or similar AI tools?”

It is a fantastic question.

Firstly I believe, the generation part is still not completely solved, especially the visual imagery and video aspect of it. The generation part is currently solved only for code generation and language. You can see why many companies are now pivoting to the Anthropic's playbook.

My belief is that general purpose LLMs will eventually dominate generic tasks.

The real edge for startups now is going niche and deeply understanding domains.

With this background, let me explain how MMM synapse is different from copilot.

My mental model is:

Microsoft Copilot = General Productivity AI

MMM Synapse = Specialized Marketing Measurement Intelligence Memory System

In PPT context, copilot is extremely useful for:

- Generating slides

- Summarizing decks

- Editing images

But it fundamentally operates on a broad generic intelligence layer.

MMM Synapse is very different. It is vertically specialized for:

▪️Marketing Mix Modeling

▪️Multi-year deck interpretation

▪️Institutional measurement recall

▪️Strategic reasoning

Unlike Copilot, it is not trying to be a universal workplace assistant.

It is solving a very specific enterprise problem:

Marketing organizations repeatedly forget what previous MMMs already discovered.

Most enterprises run MMMs monthly/quarterly.

But then:

▪️Analysts leave

▪️Agencies change

▪️Decks get buried

▪️Learnings disappear

▪️Nobody remembers why a recommendation was made 2 years ago

So every new MMM engagement starts almost from zero.

MMM Synapse was built specifically to solve this institutional memory collapse.

📌 The Domain Intelligence Layer

This is the biggest differentiator in my opinion.

MMM Synapse has a deep measurement native intelligence layer.

For example, it can reason across:

▪️Why TikTok contribution increased over time

▪️Whether brand equity is collapsing

And with our built-in decision history scoreboard, things become even more powerful because the system remembers:

▪️Why a decision was taken, what was the business impact

▪️Whether the same strategy should be pursued again !!

📌 Temporal Intelligence

One unique capability we built into MMM Synapse is temporal reasoning.

It can answer questions like:

▪️How did the mix change from 2024 to 2025?

▪️Why did Meta ROI collapse in a certain year?

📌 The Strategy Sparring Partner

Since MMM Synapse is trained on a decade of MMM consulting notes across industries, it has seen many strategic patterns repeatedly play out.

It can act as a genuine strategy sparring partner.

Unlike generic LLMs, the responses are grounded specifically in marketing measurement logic and historical MMM context.

To see how MMM Synapse can unlock value, feel free to schedule a demo

Check out our website here - https://www.aryma.ai

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://www.aryma.ai

Everyone is suddenly racing to add AI into MMM.

But I increasingly feel many vendors are solving the easy problems.

Creating MCP servers to expose MMM outputs to Claude is not some massive breakthrough. You can vibe code it nowadays.

The real problems in MMM were never about “connecting to AI”.

The real problems are:

• Why do organizations forget MMM learnings every year?

• Why do MMM outputs live in silos?

• Why do budget optimizers recommend mathematically optimal but business impossible plans?

These are the problems we have been obsessed with solving at Aryma Labs.

📌 MMM Synapse

One of the hidden problems in MMM is organizational memory decay.

- Teams change.

- Agencies change.

- Decks get buried.

- Insights disappear.

A new CMO comes in and the same questions get asked again.

MMM Synapse was built to solve this exact issue.

It acts as a memory layer for marketing measurement - connecting past MMMs, experiments, decks, decisions and business context into a searchable intelligence system.

More than connecting MMM outputs to external AI, connecting and establishing intelligent logics within many MMM reports is the harder task.

And we are proud to have cracked this.

📌 MMM Singularity

Another major problem is MMM outputs are disparate and analysts fail to connect the dots into one coherent narrative.

MMM Singularity was built to solve this.

It acts as an intelligence layer that connects the various outputs of MMM (Contribution chart, predicted and actual charts, effect share and spend share, budget optimization) into a unified reasoning system.

📌 Aryma Nebula

Most budget optimizers fail because they optimize mathematics instead of reality.

They ignore:

• Interaction effects within channels

• Spend floors

• Risk tolerance of the brand

Aryma Nebula was designed very differently.

The goal was not just optimization. The goal was decision making grounded in business reality.

A mathematically perfect answer that cannot be implemented is useless. That is why our multi agentic approach helps. One checks the mathematics and the other checks the business relevance.

📌 The reason for our success

Honestly, I think a big reason we could build all this is because both Venkat and I are still extremely hands on.

- We still talk to analysts, create and debug MMM models.

- Still sit in client discussions.

Venkat Raman is perhaps the only founder in the marketing measurement circle who has engineered and built NLP products (yes the word that was cool before Gen AI :) )

Further we realized the importance of having a team of AI experts very early on. AI for MMM is serious business. It can't be vibe coded. Since 2022, we have invested heavily in our AI team and R&D. And Recently launched our AI vertical Aryma AI

And the results are now paying off.

Schedule a demo with us (https://aryma-ai.arymalabs.com/) and you will be amazed by what we’ve built and how it works on real data.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Can AI build MMM models in mins?

Yes.

But would it be accurate and reliable?

Not all.

AI maturity still has a long way to get to that level of accuracy.

At Aryma Labs, the core nucleus 'The MMM model' is still built by us (humans). But we leverage AI innovatively to generate insights, disseminate insights, perform deterministic tasks like budget optimization. We call it the Peripheral Agentic MMM.

Recently we demo-ed our products MMM Synapse, Nebula, Singularity and the yet to be publicly released 'PPT add-in' (this is gonna be real game changer) to a group of CMOs, Brand and Media managers. Stay tuned for this release.

One of the CMO was so impressed that he wanted to explore strategic investments in us !!

While we were flattered at the offer, he asked us what makes your AI solutions so accurate, so on point and so rich in insights?

Well the answer is - Our MMM Knowledge.

Yes, our Moat is not the LLM. Soon all LLMs will become a commodity.

In terms of AI products/solutions, what will separate good MMM vendors from mediocre ones is not which advance LLM they use, but on what data the intelligence layer was trained on and plus the prompt engineering skills.

LLM is only one part, the other part is 'How well do you prompt?'

Consider the following scenarios:

• Sales and media spends trend together

• There are multiple overlapping promotions, strong seasonality

• Saturation curves demonstrating diminishing returns.

AI cannot magically know how to connect the dots and form a coherent accurate narrative.

These overarching rules must be explicitly told to AI in the form of Prompt templates.

For example:

"Check the budget optimization scenario and verify whether the spends are taken out of channel that was demonstrating diminishing returns"

As you can notice, the above is a MMM skill. There is a link between budget optimization and saturation curves.

📌 Good MMM Prompts Come From Good MMM Analysts

At Aryma Labs, both Ridhima and I have taken a new role "Chief Prompt Supervisor". Our excellent MMM analysts provide prompts grounded in MMM knowledge. We then supervise them and edit it to account for edge cases and additional logics.

The quality of an MMM AI system is not determined only by the LLM.

It is determined by the quality of the domain expertise embedded into the prompts, workflows, diagnostics and reasoning chains.

The better the MMM analyst behind the system, the better the AI performs.

📌 Good MMM Analysts - The force multiplier of AI

A lot of people think AI itself is the force multiplier. That is only partially true.

In MMM, AI becomes powerful only when guided by strong MMM analysts.

Because good MMM analysts define:

• The guardrails

• The edge cases

• The statistical checks

• The business constraints

• The causal assumptions

• The operational logic behind the model

In many ways, the first force multiplier is not AI.

It is the MMM analyst who shapes the intelligence behind the AI.

Interested in our AI solutions ?

Check out our website here - https://aryma-ai.arymalabs.com/ and feel free to schedule a demo.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Marketers / analysts however apply the same measurement philosophy for retargeting as they do for other non-retargeting campaigns.

This in my opinion is not the right approach.

In case of Retargeting, most marketers think they are measuring overall impact. They are not. They are measuring impact on a very specific, pre-selected group.

Let me briefly explain the causal estimands of ATE and ATT.

ATE (Average Treatment Effect) -> Effect of treatment on the entire population

ATT (Average Treatment Effect on the Treated) -> The Average Treatment Effect on the Treated (ATT) specifically measures the average effect of the treatment on those individuals who actually received it.

In a clean randomized controlled world, ATE is what you aim for. But we all know the marketing reality, it is messy. RCTs are not possible in marketing.

With the definition behind us, lets understand the Retargeting measurement nuance.

📌 Retargeting campaigns

Retargeting campaigns are not random.

They are explicitly designed to target:

- People who visited your site

- People who already saw your ad.

- People who added to cart but didn't purchase

- People who showed some form of intent

This is in essence is selection by design.

📌 The change of Estimand ATE to ATT

The moment a marketer implements retargeting campaigns, the question changes from "What is the impact on everyone?" to "What is the impact on this 'high-intent' or already 'exposed to treatment' group.

Hence the estimand changes from ATE to ATT

📌 The counterfactual change

One tricky part is, when you go from ATE to ATT, the counterfactual also changes.

You can’t meaningfully ask: “What if a low-intent or non exposed user saw this retargeting ad?” Your counterfactual group gets restricted to the people who already received the ads (treatment).

📌 The generalization of ATT mistake in Retargeting

Marketers often take retargeting results and say:

“This channel has 5x ROAS, let's scale it massively”

But mind you, this result is not for the whole population. This is just applicable for 'high intent' or 'already exposed' group.

📌 Final Takeaway

Appreciating the nuance of how in retargeting, the ATE changes to ATT leads to correct inferences, better decision making in the form reduced ad spend wastage.

One should not confuse Precision targeting with Population-level impact. Retargeting doesn’t just change who you target. It changes what you are estimating.

P.S: If you are interested in learning causality and causal experiments for marketing, we have an extensive course on the same. Link here: https://mmm-courses.arymalabs.com/causality-and-causal-experiments/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Most discussions on MCP focus on the obvious benefit: AI can now talk directly to marketing platforms.

But I think a bigger opportunity lies elsewhere. Here is my crazy idea :)

📌 The Real Opportunity: Ad Set Level Measurement

Today, most MMMs operate at a relatively aggregated level - Channel level.

MCP potentially opens a future where MMM vendors can access highly granular advertising data directly from platforms.

Imagine building MMMs at the level of:

• Ad Sets

• Audiences

• Creatives

• Placements

• Objectives

Aryma Labs is the first company to crack this granularity puzzle as early as 2024 through our patent pending GTA approach. We already model at such granular levels !!

Some of our clients instead of asking : "How much should I spend on Paid Social?"

They ask:

Which 20 ad sets should receive the next $100,000?

This has already been a game changer.

📌 MCP and Creation of Premier League

Now imagine a large brand working with multiple MMM vendors simultaneously on the same measurement problem - a MMM.

Each vendor's MMM recommendations are exposed through MCP and connected to execution platforms.

The brand then chooses:

• Implement Vendor A only

• Implement Vendor B only

• Implement a blended strategy

Then observe the changes (lift or decline) in performance through in-platform metrics.

For the first time, competing MMM models can be tested against each other in production !!

📌 MMM Becomes a Competitive Sport

Today, many MMM vendors make optimization recommendations.

Very few are systematically scored on whether those recommendations actually worked.

Here is my Future workflow:

Step 1:

Vendor submits budget and bidding recommendations.

Step 2:

Recommendations are executed.

Step 3:

Incremental outcomes are measured.

Step 4:

Performance scoreboards are generated.

Step 5:

Top performers stay. Bottom performers get relegated.

📌 What Gets Measured Gets Improved

I propose a scoreboard tracking of sorts that details out:

✅ Revenue lift generated

✅ Incremental ROAS achieved

✅ Budget efficiency gains

✅ Prediction accuracy

✅ Recommendation hit rate

Every vendor would know their recommendations are being continuously evaluated.

For decades, marketing measurement has largely measured marketers.

Soon, we may start measuring the measurement companies themselves !!

And when that happens, the industry may finally discover which modeling paradigm really works - Frequentist or Bayesian, Does Hierarchical modeling works or not, Which vendors' MMM algorithm really moves the needle.

MCP may not revolutionize MMM because it connects systems.

It may revolutionize MMM because it may creates accountability.

We are building some cool products in this regard. Stay tuned 🚀

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

For the uninitiated, MCP stands for Model Context Protocol. And here are some good resources to get started on it.

https://www.reddit.com/r/mcp/comments/1mb8xm6/explain_mcp_to_me_like_i_am_5/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

It is not some massive technical breakthrough. But it does provide benefits under certain assumptions.

📌 Why the MCP play?

If you have seen Breaking Bad, you know one thing: distribution is everything 🙂. In other domains too distribution is everything.

Most measurement vendors are betting that usage of Claude, ChatGPT and similar AI interfaces will explode in the coming years. They are not wrong.

But the big question is :

- What fraction of those people will actually use those AI chatbots to ask about MMM and Incrementality experiments?

- And whether those interfaces will do justice to the requirements of MMM?

📌 What MCP can and can't do

MCP solves a connectivity problem. It allows Claude or ChatGPT to access your measurement platform through a standardized interface. Useful? Yes.

But MMM never really suffered from a connectivity problem. MMM suffers from a visualization problem, interpretation problem and implementation problem.

In MMM, seeing is believing.

When a marketer asks: "Why did Paid Search contribution decline?"

The answer is rarely a paragraph of text. The answer is usually a combination of:

- Contribution charts

- Saturation curves

- Comparison of previous spend share effect share vs now

The visual itself often carries more information than the explanation (at least for a seasoned veteran).

A generic MCP call may retrieve the data.

- But can it automatically identify the most relevant charts needed to support the conclusion?

- Can it know when a saturation curve matters more than a contribution chart?

That is a much harder problem.

As MMM outputs become richer, the amount of context required also explodes: This quickly leads to token maxing and inflated AI costs.

Which raises another practical question:

Will organizations really want teams spending all day inside premium AI subscriptions for MMM workflows, when a purpose built SaaS +AI platform may provide a cleaner experience?

📌 Why we didn't take the MCP route for MMM Synapse

MMM Synapse is not just a connector between AI and MMM. It is a memory and reasoning layer specifically designed for MMM. When a user asks a question, MMM Synapse does not simply retrieve data.

It retrieves:

- Relevant outputs and supporting charts

- Historical discussions, business context and previous recommendations

And reasons about them like a seasoned MMM practitioner. It is after all trained on 10+ years of our consulting notes.

In MMM, evidence matters as much as the answer.

In many ways: MCP = Access Layer; MMM Synapse = Understanding Layer

The future of MMM is not simply connecting models to AI.

Because ultimately, MMM is not just about generating answers. It is about generating confidence.

Check out our website here - https://aryma-ai.arymalabs.com/

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

Well, You Should.

In Marketing Mix Modeling (MMM), almost all vendors adstock transform their media variables. This is done because the ad displayed today has a lingering impact into the future and more airing of ad sometimes stops having a incremental effect (the saturation).

However when it comes to assessing the impact of a promotional effect or holiday, no carryover or lagged effect is taken into consideration.

In most MMMs, holidays and events are treated like switches. Example:

Black Friday = 1

All other days = 0

This form of encoding is also known as 'Dummy encoding'.

But this is to simplistic and wrong. Because human behavior doesn’t follow calendar boundaries.

📌 Events also behave like media

Let's take an example:

A Black Friday deal creates anticipation -> browsing -> delayed purchase

A World Cup or Super Bowl doesn’t just spike viewership on match day alone. It builds momentum, conversation and residual engagement.

Hence all of these are not a 'point-in-time' effect.

What is perplexing to me is, we model media with memory but we model events as memoryless !!

This is clearly not the right way to go about.

We should ideally model carryover + lag for events and holidays too.

In fact we too didn't until Nov 2024. Through our breakthrough RBF paper we solved this problem. More on this shortly.

📌 Traditional MMM with Dummy Encoding - The wrong approach

When you use dummy variables:

- You assume instant impact

- You assume zero pre-effect

- You assume zero post-effect

But in Reality:

- There is build-up effect (anticipation)

- There is peak or crescendo(event window)

- There is decay (after-effect)

A single binary variable cannot capture this shape.

So what happens?

▪️You over-credit the event week

▪️You miss pre-event demand creation

▪️You misattribute post-event sales to media or base

📌 How we started measuring effect of Events accurately through RBF.

Radial Basis Functions (RBFs) are built for exactly the above problem.

At their core, they model effects as distance-based smooth curves around a center point.

Instead of: Event = 1 or 0

We get a continuous influence curve around the event:

If you don’t model event's carryover or lag properly:

You will inflate media ROI during event period

You will make wrong budget decisions post-event

And worse, You will think your model is “working”. Because hey R squared will still look good.

📌 The Success Story

Since Jan 2025, we have incorporated this technique for all our clients. Roughly 50 models. In all our models we saw a remarkable improvement in MMM metrics.

The clients also expressed a greater satisfaction in having measured the effect more effectively.

We also had multiple MMM analysts from different companies write to us sayin that RBF technique is really a gamechanger and it improved their models too. We open sourced our research and code for anybody to try.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

With the launch of Meta Ads MCP + CLI, AI agents can now directly interact with Meta Ads infrastructure. All through natural language !!

This is a pretty significant shift.

Until now, AI mostly sat outside the execution layer. It generated ad copy, summarized reports or suggested ideas.

Now the AI can actually touch the media buying system itself !!

But there is an even bigger opportunity here.

📌 What if MMM becomes the intelligence layer before the MCP layer?

The architecture could be something like this

MMM (or other marketing measurement like Geo tests) -> Decision Layer -> MCP -> Meta Ads Execution

Instead of letting the AI blindly optimize based only on platform signals, the AI could first consume:

▪️MMM contributions

▪️Saturation curves

▪️Marginal ROAS

▪️Carryover effects

▪️Interaction effects

▪️Budget constraints

▪️Creative fatigue diagnostics

or if it experiments

▪️Lift%

▪️ATE

▪️ATT

Then use the MCP connector to execute changes inside Meta Ads.

Example:

"Reduce spend in Meta Bid cap campaign because MMM shows saturation beyond $50k/week."

"Shift budget to retargeting campaigns because marginal returns are higher."

"Pause creatives with rising CPM but declining incremental contribution."

The execution layer becomes AI driven. But the guidance layer becomes Measurement guided or driven.

📌 The Most Important Question - Trust in Measurement Layer

Can you trust the MMM itself enough to allow autonomous execution?

Because the moment an AI agent starts changing bids, budgets and creatives based on MMM outputs, the cost of a wrong model becomes enormous.

A poorly specified MMM can now operationalize bad decisions at great speed.

And this is exactly why the measurement layer becomes the real bottleneck.

The bottlenecks are:

▪️Is the model causally reliable?

▪️Are coefficients stable?

▪️Are the saturation curves trustworthy?

▪️Did you validate incrementality externally?

▪️Are interaction effects modeled properly?

▪️Is the model robust under structural changes?

This is why we at Aryma Labs still believe in 'The Core should not be automated". The MMM or other experiments needs to be human made with statistical rigor and care.

📌 The future stack according to us looks like this:

Foundation Layer -> MMM + Experiments + Diagnostics

Intelligence Layer -> AI Reasoning Layer

Execution Layer -> MCP / CLI / APIs

Platform Layer -> Meta, Google, TikTok, Open AI

The irony is:

As advertising execution becomes easier through AI, trustworthy measurement becomes even more valuable.

But at Aryma Labs we are well prepared. We pay greater attention to measurement layer first and because of which our AI layer is trustworthy.

The more things change, more they remain the same :) Accurate measurement matters.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/

One interesting comment was "Why don't we include Macroeconomic variables as part of Base in MMM?"

It is a good question.

I believe they should not be included as part of the Base for the following reasons:

Macroeconomic variables are not structural demand

Firstly, the media and marketing events are assumed to have continuous positive effect on the KPI (every day, week, month). But macro events have intermittent and unpredictable effect (sometimes negative too) on the KPI. In some weeks or months they may have an impact and in others they may have no impact.

Second, the base is construed to be a 'structural demand' .

Macro variables, on the other hand, are explicitly included in the model to explain variation in demand.

GDP changes, inflation shocks, tariffs etc are not part of the baseline. They are technically explainers of deviation from baseline.

But a paradox is also that, Macroeconomic factors during peaceful periods are often very boring and remain static.

Take GDP of some countries for example. They remain static at 3.1% month over month.

How do you reduce this variable to zero? you can't.

But these variables have powers to explain the change in KPI over and above the media / marketing effects.

Predominantly they are good explainers of negative effects.

Therefore, though at first glance macroeconomic variables may seem to be a good contenders to be included in the base, they should not.

They don't capture structural demand. Veblen goods almost always buck the macroeconomic trends and also bare necessity goods.

Thanks for reading.

For help with MMM, Causal Marketing Experiments and Experimentation, get in touch with us.

We also build some pretty cool AI products to aid Marketing Measurements. Check out our products page to know more -

https://aryma-ai.arymalabs.com/