Sramana Mitra: Joe, I’d like to do a couple of use cases because a lot of the things that we are talking about here are fairly complicated. The best thing to do is to just go through a couple of use cases. You can pick whatever customers best illustrate the concepts that we have discussed.
Joe Shamir: There is a huge investment today by all the consumer goods manufacturers in promotions. We are talking about business in the order of up to 30% of the gross revenue of these companies. It is a very large investment. It’s growing because their investment in media advertisement is growing. On the other hand, the impact that these promotions is generating has a very low accuracy and as a consequence, there are a very limited number of players and processes that are able to afford a reasonable return on investment.
What we have done is use a large amount of data with high number of attributes of these promotions – much higher than the normal that is used. As a consequence, we were able to obtain, with the support of a learning machine, a higher precision in processing that big data and turning it into a more reliable uplift. This lets us create reliable ROI that supports both trade marketing, when they are planning these promotions on a mid-term level on the annual budget, as well as operations when sales people actually have to choose the promotions to do in the short term.
Afterwards, we use the same model in order to extract the data from the past in order to analyze the ROI from the past. Here comes the new trend that we’ve been leveraging very early on with very good results. It’s the ability to use the new generation of machine learning. In this specific case, it’s a very sophisticated one that puts us in a situation to leverage these large amounts of data both in terms of scalability as well as high number of attributes.
Sramana Mitra: I’m trying to get much more specific and visceral. I’d like you to take a customer and step us through their process and give me some specifics. You said 30% of revenue is going into marketing spend. What tools do you have to be able to predict anything based on the marketing spend? You’re working in the supply chain. How does that relate?
Joe Shamir: To be more specific, we are collecting the data directly from the channels. We are collecting the actual demand from the channel at a very high level of detail, including all the attributes that characterize specific promotions on specific dates of the year, which includes the discount strategy, the description of the promotion, and some other attributes.
For example, if flyers are going to be distributed. This fact, among others, is being collected and synchronized with the specific promotion that is taking place in a specific channel. The statistical engine that I mentioned before that we have developed is able to automatically develop the baseline. What is the predicted demand on a daily level at different locations in that channel without such a promotion?
As a consequence, on top of that baseline that we are able to generate automatically based on our statistical models of the demand, we are able to use the machine learning in order to build on top of it. We learn from those promotions from the past. I’m able to extract the uplift that those promotions have created on top of the baseline and then insert it and allow the machine to learn from those events in the past.