Transcription of Audio Recording
Tom: The event this evening is around predictive analytics. This is an interesting topic. It is in and of itself something of a buzzword. It’s got lots of buzzwords that surround it. But it’s quite a fascinating technology that really has only become even possible with the advent of big data, and large datasets, and the cloud storage for large data. And then the analytic tools that go along with that. So we have a wonderful special guest this evening. His name is Greg Leflar. He is from our Houston office, the vice-president in Houston. And he is something of a predictive analytics expert.
Greg: So, to Tom, thanks for the introduction. I’m pretty sure that as we go through this, what I want to do is to try to kind of lay a foundation for what predictive analytics is, what it’s not. And then talk about some of the uses and the ways that people in different industries are taking advantage of these technologies to choose. All right. This is such an academic and heavy sounding topic. I thought it would be good to start with a quick story that would make it real and tangible to those of us in the business world.
Several years ago, a major telecom company decided to launch a very typical marketing campaign, direct mail marketing campaign. They intended to solicit a million people at a cost of $2 million. They expected to get a 2% response rate from that campaign. And then you see, at the end of the day, the profitability of doing that’s about 5%. This is, if you’re in marketing, this is a very typical exercise. But before they took action on this campaign, they elected to engage an analytics firm to build a predictive model to try to better understand the population of prospects that they were looking at, to see if they could improve the profitability of the marketing campaign.
So what that firm did was they took those million customers, they gathered all available data about them, they looked for patterns and behaviors and how customers like those prospects, had interacted with marketing this type in the past. And they ultimately put each of those people into one of four quadrants. So you see on the left side, it shows whether or not they are likely to buy if they do receive the direct mail. And across the bottom, what they would do if they don’t receive direct mail. So this group is obvious. These are the customers that you’re gonna get no matter what you do. So it’s a waste of resources to be sending these people mail.
Just the opposite of that. there’s another group that no matter what you do they’re not gonna respond. Also a waste of resources. those two are very simple. The third one is kind of interesting. So here, these are customers that you are actually more likely to get if you don’t mail them. A lot of times, these are existing customers. People call these “sleeping dog” customers. Who if you send them, particularly if it’s a subscription-based service, if you send them a marketing message, you may actually remind them that their contract is up or that it’s time to go shopping. So clearly this is a group that you want to stay away from. And that allowed them to really focus on this last group. This is the group that they perceive to be most persuadable and worth the investment of spending the time and energy to market to those people.
So let’s look at what that does to the financials of this particular case study. You see that we reduce the number of people that we mail to by 75%. And that brings the cost to the campaign down dramatically. We also increased because we are soliciting people who are most likely to respond positively, we increase that number just from 2% to 3%. It’s not a huge difference. But because we also brought the cost down at the same time, look what happens to the profitability, from 5% to 58%.
Now, that is one use case in one industry. But I want to try to open your eyes to the implications of what that story tells. The real opportunity here is that this can revolutionize the way that businesses make decisions and allocate resources. And this is the real advantage that is made possible by predictive analytics.
So before we go any further, as Tom alluded to, if you spend any time in industry conferences or reading trade publications, you’re probably overwhelmed with the number of buzzwords that are floating around in this in this field. I’m gonna try to give you a simple basic definition of what predictive analytics is, and then try to make sense of some of these buzzwords so that you can understand how these things are related to each other and how they’re different.
So let’s begin with just a really simple definition. Predictive analytics is a field of analytics where you take historical data and try to use that data to identify the likelihood of future outcomes. Now, you may be saying, “We’ve had analytics for a long time. How is this different?” If you follow this evolution of data and analytics from where we started, it really begins with reporting. I think everybody here is probably a consumer, or creator of some sort of reporting. This is purely a look in the rearview mirror about what happened in the past.
We advance from there to begin to try to understand why things happened, not just that they did, looking for correlations between variables, and trying to get smarter about what was happening in our businesses. Through primarily technological advancements, we then already know, essentially, what’s happening right now. Now, this collectively is what you would refer to as descriptive analytics or more commonly, business intelligence. So the key difference here when we moved to predictive analytics is that we crossed the threshold of the present. We’re no longer looking at what’s happening right now or in recent history, we’re trying to make some assertions about what is most likely to happen in the future. And that’s the fundamental difference between descriptive and predictive analytics.
Now as soon as I say that, you may also be thinking, that sounds a lot like forecasting. We’ve had forecast for quite some time. Let me use a quick example to try to explain the difference. So with forecasting, what you’re trying to do is use trend analysis to try to make a projection of what a group of people is likely to do. A Very common example is what we see from all the major news, the media, news outlets during a presidential election. They’ll go through and try to forecast which state is gonna favor which candidate during the election.
By contrast, with predictive analytics, we’re trying to understand an individual’s behavior or likely behavior. And by doing so, put ourselves in the position not just to be aware of what’s likely to happen in the future, but to be able to influence those future outcomes. So an example here in the 2012 presidential election, the Obama campaign hired a team of 40 data scientists to try to do something very similar to that, the analysis that we talked about with the ad campaign. They wanted to understand of likely voters. We have limited resources to work with. We can’t approach everybody. We want to know the people that we are most likely to move in our candidate’s favor if we go and knock on those people’s doors. And clearly, they were able to bring home a win for their candidate.
I also want to talk about the relationship with big data. And I use the word “relationship” too often. Big data and predictive analytics are used interchangeably and that’s not really a fair way to use these two words together. Simply put, you can think of data, whether it’s big or little, as the food that feeds predictive analytics. So, I’ll talk about the size of the data set here in a second but you have some sort of data feeds into your predictive analytics processes and in the output of that is some sort of predictive model that gives you, in this case, a decision tree about the percentage likelihood of particular behaviors.
It’s also worth mentioning that people always believed that you need tons of data in order to do this type of modeling. And that’s not necessarily true. You do need the certain amount to be able to build a model that has any degree of accuracy, but there’s a decreasing marginal value of additional data. So if you’re thinking that you don’t have terabytes of data to work with, that’s okay. You can actually get some pretty accurate models with medium-sized datasets.
Machine learning is certainly a very hot buzzword. So I wanna to try to touch on that a little bit also. I’m gonna compare this to try to explain why it’s so such a hot topic today and why people are so excited about it. I’m gonna compare it to what we’ve typically done with classical analytics. So in most cases, you have an educated expert who knows, I’m using another sales example here, somebody who knows enough about the industry to be able to build a hypothesis that says, in this case, advertising increases sales. They know this from their own personal history. And they can perform various analytical techniques to determine that if you increase advertising by 10%, then sales will increase by some other numbers, say, 15%. And there’s a feedback loop that comes back and they get better and smarter about that particular hypothesis.
What’s interesting about machine learning, and one of the reasons that it is so powerful is that it completely turns this process upside-down. You actually begin with the desired outcome. In this case, I wanna increase sales. A machine learning algorithm will then look at not one variable which in this case was advertising but all variables, all available variables. And using a massive volume of calculations, look at the correlations between each one of these variables and each one to increasing sales. And then, ultimately, output the desired action at the end which in this case is variables D, G, and H are most highly correlated to increased sales. So, therefore, whatever those are, invest in those.
And when you play with these algorithms you quickly find, what blows some people’s minds is that the features or the variables you are working with, it doesn’t matter what they are. The algorithms look at all of them and output some sort of information that helps you understand which of those that you should try to manipulate.
Another thing that we hear quite a bit about is the role of the data scientist. I hear people trying to determine what is a data scientist? Where do they fit in the organization? And I’ve heard many different definitions of what a data scientist actually is. So this is my best attempt to describe the different attributes that somebody who can call himself a data scientist would have to have.
So I think the first part is obvious, a quantitative analyst, or a quant. This is somebody who’s skilled in Math, Physics, probably has an advanced degree in those fields. There’s also certainly an element of being a researcher. The process of building predictive models and using predictive analytics is very different than most other software fields. That very much requires someone to be familiar with building hypothesis, testing those, and working in the scientific method.
The third component I’ve called being a hacker. You don’t necessarily need to be a software engineer or a computer science-oriented person, but you certainly have to be familiar enough with the tools and the technology to be able to use them in creative and non-structured ways to attack these types of problems. And then the last piece is, there is a need or a value for having domain expertise. What I mean by this is, if the person building these models is not familiar with the business context of what they’re doing, this is where you have people who emerge from a deep study of a particular model and tell a client something like, “Crops grow more when it rains, or sales will increase when stores are open,” things like that are total nonsense to anybody who’s got experience in the field.
And if you put those four things together, that’s why you’re about as likely to find a unicorn as you are at data scientist, and when you do find them, simple supply and demand dictates that they are quite expensive. The good news here is that the advancements in technology and some of the tooling that is available today has really brought this data science discipline onto a lower shelf to where you don’t have to have these advanced degrees and these sorts of things. In a lot and on other cases the tool and technology are doing some of this work for you.
I wanna to touch on this one. This is a little bit more futuristic looking but I hear more about this every day. So there this notion of a digital twin. There’s a guy named John Smart who is a researcher he runs a research firm. And He, in an interview, gave this quote which says, “When you die, your kids are not going to go to your tombstones. They’re gonna fire up your digital twin and talk to them.” That’s a little bit outlandish and a whole lot creepy but there are real tangible applications of this type of technique that we see customers using today.
On the simpler side of the scale, we’ve got customers who are collecting all of the information that they can about their customers so that they can model that person’s decision making to do things like product recommendations. If you’ve shopped on Amazon or liked a song on Pandora, it’s using this type of thing to understand the full context of you, comparing you to others, and making recommendations based on that.
The more advanced end of the spectrum is where we see that collection of data becoming so rich that it begins to represent a person’s values. So one of the examples that I saw recently was, to use the product recommendation example, is that I might not just say that people who bought this… You bought this product. People who bought that also like this. It may actually go deeper to know that you have particular beliefs, and values, and support, certain social causes, and recommend products from companies who also support those social causes.
There’s also an industrial application to this. I read a story recently about a manufacturer of the giant wind turbines that run in wind farms. They’re beginning to look at building digital twins of those so that they can model preventative maintenance, and experiment with different scenarios and how they operate their wind farms using the idea of the digital twin.
Okay. So that’s the last piece of kind of the lexicon and cut through some of the buzzwords. Hopefully, that gives you some clarity. I wanna spend the rest of the time talking about how people are using this. And there’re some that cut across all industries. And it’s also true that some industries are adopting more quickly than others. I got a retail here first because quite frankly, this is where the earliest adopters were seen. What we see in the retail space is going around now. We already talked about marketing effectiveness that’s actually a pretty mature field and there are software packages out there that help people do that.
But some of the other things that we see for subscription or service-based companies is a lot to focus on retention. And there are only two components to that. One is understanding which customers are most likely to defect or churn so that, again, you can focus your resources on those individuals. But the second component is what can we do that is most likely to encourage the person to stay. So we have historically worked on just kind of an average assumption that people like $5 Starbucks gift cards or whatever it is you’re gonna do, some special offer, maybe it’s a coupon code. In reality, everybody behaves differently and is likely to respond differently. So the other part of this predictive model is these are the people that you should focus on and here’s how you should approach them.
Product recommendations, I think we’ve talked about. Customized product pricing is another area where we’re seeing people break free from having to deal with the fact that they’ve built their pricing model, their business models on averages. We talked about the aggregate versus the individual with the presidential election example. Same thing here. So when you go to apply for a service or buy a product, you’re getting a price that has assumptions built into it based on averages of large groups of people. Using predictive analytics, companies are beginning to calculate a lifetime value of a customer and then build the prices dynamically based on the desired margin for every customer.
Moving into the world of finance. Now, obviously, if these models were perfect, we wouldn’t be here. We would be playing the stock market to predict what’s gonna happen tomorrow. But more reasonably, there’s a lot of activity around…in the area of risk. So determining the credit worthiness of the potential customer, what interest rate are you going to give them? Are you gonna require a deposit for a customer? All those can be delivered through a predictive model that is custom tailored for that individual.
In healthcare. So whereas the first two industries that I picked out we’re talking about making more money here, we’re trying to save people’s lives, we are largely still in a place where we treat people when they’re sick. Hospitals, doctors, research labs are beginning to use predictive analytics to try to change that, and try to understand, largely by studying genomic data, what type of preventative treatments could be prescribed that prevent people from getting sick. And by understanding the patterns between certain genetic traits and lifestyles and all these other things, what types of illnesses or ailments people are likely to have to deal with?
One kind of sub-section of that that’s really getting a lot of focus is with cancer treatment. By and large, again, this goes back to how we’ve historically dealt with averages. When people are prescribed a cancer treatment program, too often, it’s done somewhat blind. They’re given kind of the gamut of all the different things that have worked for other people. In reality, everybody response very differently to different types of treatments. Now that we can we can look at individual’s gene sequence and match that with other people who’ve tried different types of treatments, we can actually custom tailor a cancer treatment program that is most likely to work for you based on your particular genetics.
Real-time patient monitoring is another interesting area. So we’re working with a client here to capture vitals, data off of patients in real-time. And by monitoring those leading indicators, try to predict when those people are likely to have a heart attack, or a stroke, or some other negative event like that. And then one of the biggest cost drivers for hospitals right now is readmissions. Hospitals are beginning to look at predictive analytics to try to determine two things, when is the right time to discharge somebody, and what is the likelihood that they’re going to return. You wanna let them go from the hospital as quickly as you can for a lot of reasons, but you don’t want to do it so early that you run the risk of them showing back up the next day much sicker than they were when they left. So predictive analytics is being used here to try to determine the optimal point of when you discharge a patient.
Okay. So, collecting all these experiences, we have learned a couple of valuable lessons working with clients in the area of predictive analytics. And I thought it would be valuable to share with you all. Most of these lessons were learned with some bumps and bruises. One of them is that… You know, I talked earlier about, really, in the field of machine learning, technology has really made this accessible. It used to be the case that if you wanted to perform a high-end analytics, you needed a supercomputer and a bunch of PhDs. And a lot of that is going away. Particularly cloud computing and some of the surrounding tools have brought this idea of what people call utility supercomputing, Which means if you’ve got a credit card, you can essentially rent cycles on a supercomputer to run all these types of models I just described with very little upfront investment. And that’s removed this huge barrier to entry for a lot of companies that are now choosing to move into the space.
Unfortunately, the technology has not yet made it simple. If any of you have experimented with Hadoop or machine learning or any other popular tools out there, they’re still very complicated. It’s a quickly evolving field. It changes constantly. We’ve got several different client engagements where we’ve seen the technology shift underneath our feet while we’re trying to build something. And that creates a lot of headaches as you’re trying to implement these things.
More bad news. Data quality is still a major challenge. With all of the magic of machine learning and some of the things that I’ve described, it is still true that if your data is garbage, the model that it puts out is also going to be garbage. So oftentimes, clients come to us and say, “Hey we wanna talk about machine learning. We’re gonna do predictive analytics.” And we have to start with the far less sexy conversation of, “How good is your data? Where are you getting your data?” So, that’s still an important foundational element.
It’s also true that when you look at the overall solution, one of the hardest parts is just the ingestion of data. If you follow the discussions around big data, they talk about the 3 Vs of big data, volume, velocity, variety. We have just breathtaking growth in data every year. Sometime last year, I don’t know what the number is now, someone said to me that in the past two years 90% of the data in the world was created. And it’s not a straight line. This is an exponentially growing volume of data that were dealing with. And just capturing all that and figuring how to store it at the speed that it’s coming in can sometimes be a challenge. So when we talk about solutions with predictive analytics, a data ingestion component is often one of the first things that we’ll focus on.
Something else that we learned, and I kind to touched on this with the data scientists saying that they needed to have a research bent to them. The first couple of times that we approach these solutions, a bunch of software engineers, we tried to architect the solution like any good software engineer would. And it didn’t work out very well. What we found is that this field lends itself much better to using an experimentation based approach. So instead of knowing upfront what type of approach or model is gonna work best, we try five or six in parallel. And we test them against to each other and then refocus our efforts on the ones that produce the best results.
And then last and probably most important, more than any other field that I have worked in, there is a heightened need to understand the users and understand the outcomes that they care about. I’ve used a lot of examples with sales and if your output is going to be delivered to sales people it needs to make sense to them. Nobody in a business role wants to be given a science project. They wanna be given data and guidance that they can work with to make better decisions. And they don’t wanna know how it was created. There was a picture here a minute ago that showed a decision tree of… I think that snapshot was from a mortgage application predictive model. Any business person can read that and follow this decision tree that says, “Based on these features of a potential customer, what percentage likelihood do you have of failure to pay or prepayment.”
So there is an old Danish proverb that says that “Prediction is difficult especially when dealing with the future”. And I wanna leave you with this thought that this is still true. Predictive analytics is not a crystal ball that gives you a view into the future with any degree of certainty. It’s about incrementally improving the way that your companies make decisions. It’s about being better than guessing because that’s what most of your competition is doing. And predictive analytics, because it has become so accessible through the volume of data that’s available and then the tools and technologies that are out there, this is your opportunity to get a competitive advantage over your competition.