How we reduced MMM’s In-sample MAPE by 4.6 percentage points through RBF technique.
Recently we replaced an existing MMM vendor for a prominent client. One of the complaints the client had with the earlier vendor was that – they were not able to capture key events and promotional events accurately.
The client had clear data on which weeks they ran the promotional events and which weeks they conducted the key events.
However despite the data, the earlier vendor could not capture it comprehensively.
We dug into their code and found that they had used the archaic ‘dummy encoding’ to capture the events.
📌 The problem with Dummy Variable (one hot) Encoding
A simple dummy variable assumes that the entire promotional event impact happens on one isolated day or week.
But anyone who has looked at real market behavior knows that this is not true.
The influence starts building on the day / week of launch, peaks after some days and decays over the next several days.
When we compress this entire behavioral arc into a single spike (like in Timeline 1), we end up with:
❌ Mis attributed channel contributions
❌ High in-sample MAPE
❌ Wrong budget recommendations
This is one of the hidden reasons why MMM sometimes “feels off” during key promotional events. If you see a lot of daylight between your predicted and ground truth chart, that is not a healthy sign for your model.
📌 How Aryma Labs tackled this through Radial Basis Function (RBF) Encoding
We didn’t believe that the effect of the promotion or events lasts only at the exact time of launch. We believed that its effect slightly lingers.
So how do we capture this?
Applying RBF can be thought of as placing a ‘water cup’ or ‘standard Normal Distribution’ at the point of campaign or promotion launch.
Instead of a crude one-day or one week spike, RBF encoding allows us to model any event as a smooth curve with a rise, a peak and a decline.
📌 The Result:
After applying the In-sample MAPE reduced from 10.6% to 6.0%. A remarkable 4.6 percentage point reduction or approximately 43 % reduction !!
Overall, RBF application leads to:
✅ Correct separation between true media impact vs seasonal lift
✅ More trustworthy contribution splits
✅ Better model fit and improvement in In-sample MAPE
✅ Better downstream budget decisions
If you are still using Dummy variable encoding, you are underfitting reality.
▪️Any promotion or event is not a date. It is a distribution of activity.
Model it like one and your MMM will immediately start telling a more truthful story.
If MMM is not feasible you can use techniques like ITSA or Information theory based Incrementality tests. With the latter, you can also model Creative Fatigue !! Check out the links for Aryma Delta and Adstock ITSA in comments.
