Sramana Mitra: You are saying that the number and complexity of input variables are going to go up, creating greater accuracy.
Saed Syad: Yes. If you have more variables, you have a chance to better understand the problem. But at the same time, traditional modeling can not support this type of change because of the amount of data and the speed at which this data is generated. It is not easy. It is about real-time predictive modeling. If somebody else wants to do something about this, they cannot use traditional modeling techniques.
SM: Then comes the question of being able to handle that many variables at speed and scale.
SS: Yes. This is an issue in any RTB platform. It is not just about the data, it is about the speed of the scoring, building, and updating the model. All the aspects of predictive modeling need to be changed and have some sort of scalability. Those features should be regarded in all aspects. Building a model, creating a model, updating a model, changing the variables, etc. It is a completely different game.
SM: You are excited about 1.5% click-through at the moment. Presumably, in the next eight to ten years, we can increase that to 10% to 20% with better targeting.
SS: That can happen if you have very focused segments because there is a limit for the clicks.
SM: What do you mean by that?
SS: If the rate of the clicks per ad is 5% in total, then moving beyond those 5% means you need to do a lot of work to improve the ad and the time in which you show that ad. Moving from 2% to 20% is not just about the predictive modeling issue. It is about the whole system in place – the type of model, the category of the campaign, etc.
SM: So you are saying that this level of performance will require optimization at various different levels, not just at the RTB level.
SS: Exactly. The way you design your campaign is going to affect the outcome.
SM: How much performance do you think the predictive modeling and the real-time bidding parts are going to be able to deliver in the next 10 years?
SS: We look at the uplift. If we don’t have the predictive model, we are getting x and y amount. If we use the predictive model we are getting [another amount]. We saw uplifts up to 10 times better than before. Once we have a better and optimized system, the campaign is going to perform better eventually – without using predictive models – because you have a better campaign architecture. At the same time, the predictive models can put an uplift to that base amount by 4% to 5%.
SM: Thank you, Saed. This has been very insightful.
SS: Thank you.