I love all the feedback and questions that I've gotten from doing my portion of this customer analytics program, by the way, keep it coming. If you haven't already done so, connect with me on LinkedIn, follow me on Twitter @Fader P. This is the stuff that I do every day, not just when I'm in teaching mode. The research, my consulting activities, my whole life is around these kinds of models. So if they intrigue you at all, then let's keep the conversation going. The most common question I get people say, this is interesting stuff, we've never seen anything like it before. How does it compare with machine learning? So let me spend a few minutes talking about that. It's not just the comparison, not just a conceptual, not just the empirical, but how it all fits together because it really does fit together. We talk about very different kinds of methods when we're flipping coins, heads I buy, tails I don't, this and that. So conceptually they seem very different each other and they are. But that's the beauty, that's the opportunity. You see machine learning, if you boil it all down again, I'm not going to educate you about that because there's a both other aspects of this program that start to put you in that direction, or plenty of other content out there on it. So I'm going to boil it all down. Maybe I'm going to oversimplify, but I actually think it's pretty fair characterization. Machine learning is good at two things. Only two things. Turns out, though, that there really, really important valuable things. Thing number one is, machine learning is really good at putting things in buckets. So it's like, is this customer good credit risk or not? Should we give him a credit card? Is this customer likely to churn in the next period or not? Will this customer make a purchase in the next year or not. So anytime we can take a managerial problem and break it down into a bucket type question, will they or won't they, okay? Or among a set of different brands, which one will they buy? Which option will they choose? You can't beat machine learning, you just can't. It will just do a better job of putting things in buckets of making these kinds of yes or no decisions. That's number one. Number two is giving you some explanation or at least telling you which X variables which covariates are going to be most associated with which kind of bucket? So we know from regression an machine learning that it's going to take in all of these Xs, it's going to help us understand what combination of X is our most associated with someone who is a good credit risk or not. So when it comes to putting things in buckets and explaining why machine learning is where it's at. And if you can frame a managerial question so that it is bucket worthy, you're good. So let's go back to it. If you can say, well, this customer churn in the next period or not. Will this person make a purchase the next period? Which is very similar to what I've been talking about and I'm telling you machine learning is the best way to do that. So what about our models? There are some questions that don't lend themselves to the, will they or won't they kind of thing that aren't quite as bucket worthy. So if you start asking questions about when, okay? So not will this person churn in the current period or not but when will this person churn? And if I care about the timing of it, okay? And if we're doing it over a long horizon, so it's not just will this person make a purchase with me in the next period? But how many purchases do I think they'll make over the next 15 years? So we're talking about when type questions, longitudinal questions, the how long, the how many, not just the will they or won't they. When we're talking about long horizons. And finally, if we're talking about a settings where it is longitudinal and a long horizon and we're not content to just simply say what will our sales be over that period. But how will those cells very for different kinds of customers? Based on what you've done in the past, what's my best guess of what you'll do in the future, and how often and for how long? Then our coin flipping type models are going to tend to be a little bit better, and so that's the complementarity between the two different approaches. Again, bucket type questions go with the machine learning, can't be beat. Longitudinal long horizon with granular inferences then these more probabilistic models are going to attend to work a little bit better. But there's a beautiful complementarity between the two. I was talking earlier about Zodiac and we look at those folks in the green bar on the right and say what's up with them and what distinguishes them? How do we find more customers like them? Well, a really big part of that is machine learning. One of our co founders Zodiac was a machine learning expert, and that was basically his job. This is the way it works, is going to run our probabilistic models or body and di model to come up with the overall behavior. But then we're going to use the machine learning in several different ways. Number one, to profile the customers who are high value versus low, goes back to the bucket thing. What makes the people in that green bar different from the other folks? And hey, we got a new customer coming along. Which bucket do we think that they'll belong to? So layering machine learning on top of these models is an excellent, excellent way to go. And there's another way to do it. I mentioned before the idea of covariates that if we know what covariates we want to bring in, we want to bring in the Christmas season. [COUGH] We want to bring in the anniversary sale. Then we can build those covariates directly into our body and di model. And it works pretty well, but sometimes there's a lot of different covariates were not even sure which ones we want to use. So here's another approach. Let's run are simple, simple body and di model exactly the kinds of models that we covered before. Then let's take the deviations from the models. Let's take the residuals from the models and then apply machine learning to that. So the machine learning can tell us why this one is an over forecast and this ones and under forecast and what are the X variables that tend to be associated with overs versus unders? So again it's a beautiful one two punch in order to run the basic probabilistic model to give us just the baseline behavior of what we think would be happening in the absence of other stuff going on. But then to use the machine learning to tell us about the other stuff going on. And so there's there's a very productive interplay between machine learning on the probabilistic models. And again, there can be some overlap. because there are some managerial questions that could be framed either way, as the, will this person churn in this period, or when will this person churn. And in some cases it's hard to decide which one to use. But my main point is you want to be skilled at both. You wanted to understand the kinds of questions that are going to lend themselves to the probabilistic models. And then the kinds of questions where the regression and the machine learning is going to be either the powerful way to go. Or the powerful companion to use in conjunction with the body die models in order to make better forecasts, better statements, better managerial decisions.