Sramana Mitra: You’re basically using past promotions and modeling those past promotions in your machine learning algorithms to see how the behavior changes when you run certain kinds of promotions. When another promotion is being run in a similar geography by a similar channel partner, then you’re able to predict demand using those models.
Joe Shamir: Yes. One of the problems is to identify the baseline – on top of which you have to take away the uplift in order to analyze. One piece is how to use statistical models to create the correct baseline and then to take the uplift and use them in order to teach the machine how those profiles are correlated with all these attributes of the promotions. The machine learning creates a model and the model is able to tell you what profile you will have – what type of uplift if you are putting any type of combination of attributes. Not necessarily the same that has happened in the past, because a model has been created that can see all these attributes and variables that are influencing the outcome of these profiles.
If you are inventing a new promotion in the future at a discount that you never ran before, we’ll tell you the most probable profile that you will get out of it. We’re talking about simultaneous combination of a high number of variables. When it provides me that type of future uplift, I’m able as trade marketing for example, to plan my account and say, “I would like to make this promotion with Best Buy.” I can use it to place different types of promotions over the baseline that the system has generated. On that basis, I can get all the ROI calculations, given the cost parameters driven by the attributes as well as the uplift that has been generated in terms of additional volume and additional margin.
Another example is a large retailer with a large catalog. For example, he’s introducing about 5,000 new items every month. Of course, the ability to predict demand for these new items is very difficult. One of the problems is to identify – both in terms of where to put the inventories but also in terms of where to put the marketing money – who are going to be the winners. We have to follow closely and provide the right level of service as well as the right marketing investment.
We collect, once again, events in the past of product introduction. We collect web events or web traffic on the website prior to the launch and during the first week of activity both in terms of the sales as well as in terms of the activity. We are able to predict the winners that only constitute 4% of the total number of these 5,000 items a month. We are able to identify them with 80% accuracy. This is another example of how machine learning is able to basically make decisions automatically.
We’ve been through such processes in the past in different revolutions and this is the next revolution where you need Big Data in order to have the scale of the data and the number of attributes to allow machine learning and engines to extract knowledge from them in order to turn that into an automatic process. Since machines are replacing man, there is a very strong resistance to these processes. We have evolved from traditional forecasting and statistical modeling to the collection of data closer to the consumer. When you are also going to collect the attributes of the market-driven demand, then you are trying to turn that into an automatic process.
You’ll find the attitude of the organizations to be very cautious. There is a strong resistance. Despite the fact that it is required, I wouldn’t say that a very fast process is taking place. A lot of organizations resist it. As a result, we see the early adopters that are leveraging this to their benefit. Simultaneously, I think that for smaller and younger entrepreneurs, especially if they are operating on the Internet, this is a big opportunity to gain a big advantage over the larger organizations.