Blogs

Why Synthetic Control Method (SCM) says Yes when the right answer is a Big Fat NO Honest surgeons don’t advise for surgery when not necessary or when the risk overweighs the benefits. Honest marketing measurement vendors similarly should ideally tell their clients that sometimes a Geo test is not feasible if a real control market can’t be identified. But does that happen in practice? Almost never. Instead, many marketing measurement vendors choose revenue over rigor. How do they do it ? Synthetic Control Method (SCM). I believe SCM was invented to be a cash cow for the clients. Here is how it plays out. Typical Scenario: Client: “We want to

MMM is a Game of Subjective Refraction Couple of days ago, I went for my regular eye check up and the subsequent power correction. If you wear glasses, you know the drill. The optometrist tries lens after lens, each one slightly different and asks you “Is the image clearer now or not?” If you add too less or underpowered lens, the image looks blurred. But the moment you add a few more lens (with right power), the image magically becomes clearer. This process is called subjective refraction. 📌 MMM parallel with Subjective Refraction While the optometrist was trying many lenses and kept asking me “Better with this or that?”, My

How to Validate The Interaction Terms in MMM In my last post, I wrote about why one should add interaction effects and how many one can add. In this post, I will cover how you can validate whether the Interaction terms you added are accurate or just inflating noise. One of the most common mistakes in MMM is : – Adding interaction terms because they sound right – And then Keeping them because they improve R squared value ! But neither tells you if the interaction is actually real. Yes domain knowledge can inform but statistical proof to go with it will be even better. 📌 Enter : Likelihood Ratio

Why add Interaction effects in MMM and How many interaction terms to add? Last week I got a very interesting comment “How many interaction terms to include in MMM”. This is a very pertinent question. But lets step back for a minute and understand why add interaction effects? 📌 Why Add Interaction Effects? We have been asked this questions in our courses and seminars too. “Why add the interaction effect in MMM? Wouldn’t the variables already in the model suffice?” Y = β1X1+ β2X2 + e The above is a simplistic representation of Multi linear regression or MMM (pls assume adstock transformed). In this, your X’s are independent variables. Think

Getting to 90% accurate MMM Model – Just Crown The Highest Spend Channel as Hero. I am about to tell a hack which perhaps some people already know. You can get to a 90% accurate MMM Model (both statistically and business wise) just by predicting the highest spend channel as Hero – The biggest contributor to your KPI. But the real skill of a MMM vendor is to identify and rank the other channels. They help you answer questions like : – Which mid-tier channel is actually pulling weight? – Which low-spend channel is punching above its size? – Where are the hidden synergies? – Where is saturation killing incremental

Brand Building Channels’ Saturation – Diminishing Returns, Not Diminishing Importance !! One of the most common reactions after looking at MMM saturation curves is this: “Oh look, We have hit saturation on this channel. Let’s dial this down.” This may be the right call but there always exceptions. One of the major exception include – Brand Building Channels. 📌 Saturation ≠ Stop Signal In MMM, saturation curves are meant to show diminishing marginal returns, not zero returns. At the saturation point or zone: Incremental ROI is lower, but absolute contribution is often still high. Cutting spend here is like shutting off your most consistent volume driver just because it’s not

Good MMMs are not built to predict the future, they are built to change the future!! Every good MMM is destined to be proven wrong. This may sound counterintuitive, but a successful Marketing Mix Model (MMM) is actually designed to be proven wrong. Let me elaborate. 📌 Old Policy Regime vs New Policy Regime An MMM produces a prediction of the future based on historical relationships between marketing inputs and business outcomes. But the moment you use the MMM to make a decision – change budgets, shift channels, alter media flighting patterns etc, you are intervening in the very system that generated the historical data. And when you intervene, the

Why we built a Product for each failure mode of MMM? Every MMM project fails primarily due to the following 5 reasons: 1) Not enough data / bad data 2) Model built but no trust in it 3) Model built but no way to use its insights or plan budgets (no software / dashboard) 4) Model is built but how do we prove its efficacy and causality? 5) Model is built but there is huge information asymmetry (only few teams have access to reports) We studied these failure modes or choke points over many years. In a recent pitch meeting (that lasted 2 hrs !!), the client said “you guys

Still Hesitating to Adopt Marketing Mix Model (MMM)? Don’t. Happy New Year All, In 2017, I made a prediction that MMM will be huge in the coming years. I even wrote my now famous post “MMM 101” in 2017. However, at that time many told me I am dead wrong and that MMM will be dead soon too !! Fast forward now I think my prediction was more than right 😊 I am going to make one more prediction today – MMM will be in demand for at least next 5 years and more. Why? Because accurate market measurement and attribution will only matter more in the coming years. 📌 MMM

Gen AI and The Need for Galapagos Island 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