In this episode, David Herman shares his insights on:
- How can financial service companies create value using the emerging integration between behavioral neuroscience and financial data science (the right way and the wrong way)
- How the consumers win when these two fields interact
- How to win consumer trust when building advanced AI products or campaigns, especially catering to finance industry.
David Herman packed some powerful concepts and insights in quick 30 min conversation and it is definitely worth a listen or two.
Dave’s been busy at Payoff, but his youtube channel has some great videos as well.
For additional questions, you may reach out to David on his linked or his email dh.herman “at” gmail.com
Transcripts may contain grammatical error because we don’t speak the same way we write 🙂
Welcome to Data Leaders Podcast.
I am Dhaval Bhatt and this where I sit down with top data thought leaders and extract really meaningful insights and actions that you can apply to your life and work to help you achieve your dreams and goals. This is where data and intuition meets decision-making.
Dhaval Bhatt: In today’s Podcast, we have David Herman joining us. After receiving his Ph.D in Neuroscience from University of Southern California, Dave Herman delved into data science at Payoff. Over the course of his career, Dave has developed a passion for understanding human behavior at mass scale and leveraging those insights to build truly transformative experiences that improve as many lives as possible.
Specifically, Dave develops algorithms and control systems that are focused on product needs. He is also an expert at building analytical teams that make ideas emerge at scale.
Dave has over 12 years of experience, in researching and managing data and algorithms in the space of psychometrics, human-machine interaction, time-series algorithms, recommendation and control systems, standard and alternative risk models, financial forecasting, cost-optimization models, and iterative Bayesian methods.
A little fun fact about Dave is that he is an avid gamer and he has been one since he was a very little kid. He played everything from chess, poker to competitive games.
Dave created a Bayesian ninja fighting with a quail on his YouTube channel during his college days. In his ideal world, Dave would like to hear Einstein and Bruce Lee discuss their world views with each other.
Dave, thank you for joining the show. How are you doing?
David: Doing great! Thanks for having me.
Dhaval: Awesome. I heard you speak at one of the founder’s conferences, you brought up the topic of psychometric measurement and it was really an interesting talk so I figured it will be great to have you talk to our audience. One thing I find very interesting about you Dave is that you are at the center of emerging field of Financial Data Science and Behavioral Neuroscience. How do these fields interact with each other and what are its implications?
David: Yeah. I think that’s a really an important question especially the last part, the implications. So I think that if you take the history of kinda financial modeling and the world of behavioral neuroscience, you can kinda see how the two come together in this burgeoning field, so one thing actually I learned a lot coming… in working in the field of financial modeling and this is more about credit underwriting, insurance modeling… not about the real-time trade of Wall Street, I think that has a lot of really advanced things going on there, but that there’s a rather deep history of using pretty reasonably advanced algorithms in the financial sector. I worked with some people from Capital One and here they’re dumping like markup models, they’ve used tools just progression in all of their systems and what I think is that the more modern day of science with these ideas of deep learners and whatnot is definitely finding its way into the users for these systems and things like fraud analytics, real-time update, real-time learners. But also there’s this real strong history of rigor behind like a logistic regression model.
How do you assess it? How do you measure the things performing right? You know, when a company’s, you know, bread and butter is wrap around this one algorithm, this one belief system, a lot of test and a lot of analysis used to go on around it, not just production eyes look at the ROC and you’re good to go. And I really appreciate that rigor that’s required and from that environment, I think the two are coming together going,
“Hey, these are really cool, nice new things.”
“Hey, these are real good history.”
Or how do you be rational and be predictably stable over time from the world of financial modeling. I think there’s a lot of synergy there and I really appreciate kinda entering the scene within that kinda construct of being very rigorous with our model.
From behavioral neuroscience, what I think is really interesting is that, you know, a long time ago, you could kind of basically more a qualitative theory development kinda science brought up empiricism validating those theories but not explicit quantitative modeling coming out and over time we’ve seen a lot of that come around.
Generally in this notions of reinforcement scheduling to more modern notions today of these models of novelties, salient piece of prize, predicting human behavior, predicting the way we want to understand information in a qualitative way. And so those new fields, you know, financial data and for me, psychological data is starting to become more available and behavioral data on the internet in real-time, lots of data filling the data point. And so now we’re seeing the opportunity for neuroscience… behavioral neuroscience to take that vast amount of literature and understanding and empiricism into this realm of the modern tech world, combine this into industries, in this case, financial modeling and really start to see real fast learning and real fast application to basically helping people understand their financial situation.
I think the implication is very interesting because, if we look historically, if you look at the idea of reinforcement scheduling, that exist since like slot machines and a lot of other things. So it’s not the case of these tools are inherently positive or negative but it definitely being used to drive behavior in some good ways and some bad ways. And I think that the real goal here is that if we wanna have people trust data science in the future, we need to make these models useful, transparent and directional in a positive way. So we need to build products that help people and we need to provide analysis and science in understanding about what those models believe about a person right back to them so they can be useful for them.
I think that one example is FICO. FICO tries to provide this averse action and while FICO’s still kind of a black box, it has some elements of it of transparency and an intensity clear about it and I really appreciate that and I think with these new world of “black box” algorithms we need to be very, very open about what we’re using with the data for and I think that’s honestly not some kind of like compliance issue, that’s just better science. That’s better startup and better engagement with the customer. And I think that the average Joe should be very… I would say selfish in the way they engage with this new democratization of products in a case of financial industries and all these different products. They need to be going,
“Who’s the actually has my best interest in mind?”
“Who wants to share with the learning about me and not just sell it off without telling me anything?”, you know,
“Who wants to engage with me in a way that beneficial to me in the long run?”
And I think that average Joe needs to, you know, getting an education after about what data science is and how financial industries are using it. I think it’s a really great opportunity for the masses to get a sense of what this is about and pick accordingly.
Dhaval: Very interesting. I almost stop in the parallel realm, I think of quantified self movement that is growing here in, you know, our world and it’s a… the whole quantified self where you incorporate data acquisition of your personal life especially in the context of this conversation of financial data. A lot of… you hit a nerve there, where a lot of companies are acquiring this data point and then keeps using that to commercialize it and selling it to other vendors and use it for advertisements and whatnot but a lot of times their loop is never closed on providing value back with the consumers.
David: I think that… like there’s… I don’t see anything from them something wrong with sharing data between companies, you know. It’s the point… the whole point I would assume is that that’s better information so that that other person can drive the better product to that person and that’s fine. It’s just that when it’s… when people have this… you know, when you already feel like you don’t know what someone’s doing with the data, you tend to be just trusting of it. And in the end it’s not just a bunch of data point. But we really want AI to be this thing that we live with. You know, in all these movies, it’s robots in action with people. Well, that’s way pass the point of us accepting it. You know, they never show you that… if they do, it’s like some terminator like movie of over fighting the machines but it’s really this how do we get comfortable with this kind of human machine interaction. And big part of that is trust and how do you get trust? Transparency is a bit part of trust.
Dhaval: Yeah that’s a great point. How do you establish that transparency when you’re dealing with people who don’t necessarily understand the technology at a very deep level? What are your experiences in terms of establishing that trust, Dave?
David: Yeah, so I think for example when we talk about… say for example psychological assessment, that’s rather a complicated initiative and when we do them, you know, there’s all this history of how do you score people, how do you compute, how do you give this outcome? I think that it’s not so much the case of people we need to understand like,
“Oh, you squared this.”
or this is a regression model versus a gradient boost or something like that. But more about, you know,
“What is this thing saying about me and how are you using it?”
and that’s usually very explainable. So for example when I work on things and we do an assessment, that ends permission is given back to the person going,
“Hey, this is what the system thinks. This thinks you’re very extroverted.”
Here’s stuff about the extroverted person and here’s a way in which you engage with your finances from what our data tells us. That’s like being very open with it without saying, you know,
“Here’s all the dirty details of the math.”
You know, I think that’s really the way to do it. It’s to say, you know, in the end I deeply believe that every math equation is a long sentence or a story. It’s a story about how you move this number, divided by this and it’s always some narrative, it’s just compressed and really formal but in the end, it’s saying something and has a belief structure in it.
Frankly, when I build like any model, like a logistic regression, I treat that as a kind of small organism. It has this belief structures in it and it has ways in which it interacts with the data to learn. Now we could argue that,
“Oh, people can say this is a black box system.”
But, you know, that would almost say that neuroscience is a mute effort because the brain is a rather complicated machine learning organ yet we believe we can understand it and explain it. I think that when it comes to things like random forest and gradient boosters and all these things, they’re actually deeply explainable. It’s just a question of the way you explain it. There’s many ways to describe what a big decision tree looks like however you can always throw data through it with Monte Carlo simulations. You can always see what are the due under the circumstances and you can always say what is it thinking about this person under different scenarios. And so ultimately you can always get some kind conceptualization at an individual level about what the model “believes” about them.
Dhaval: Yeah, you are touching on my very next topic. One of my very favorite quotes from you, Dave and I’m gonna paraphrase this is that the objective of data science and machine learning is to go beyond predicting the outcome. And it really do help explore the underlining beliefs that led to that outcome. That’s something I learned from you when you were… when you were speaking on stage and I, you know, jotted it down and I would made circles around it and I thought about it and I was like,
“This is golden right there.”
My question for you based on this, how can financial services, growth marketers or business leaders use this specific insights to better serve their customers which is, you know, the insights once again is, it’s going beyond predicting the outcome, it’s actually helping them understand underlying beliefs, helping the consumers understand their underlying beliefs in outcomes?
David: Yeah. So I think I actually… I indirectly stole a quote from Patch Adams and I think Patch Adams… it was just… it could not even be Patch Adams, it might just be Robin Williams in the movie but the claim was that… is that,
“You treat a disease, you win or lose. You treat a person, you always win.”
And that really hit me, even though it’s just like a trailer and I think it’s piece exactly into what data science has in front of it that, you know, we like to think about all of these model predictive, what’s your oral C score and whatnot and that is extremely important for driving certain KPIs and performances in your company and that’s a very strong use case for it. And I think that there’s a bigger use case additional to that where, you know, that model was trained up using rather sophisticated analysis and rather large data set and has all this intuitions inside of it. The weight on different variables is descriptive and it’s a shame that we might treat that as something that we leave in the black box rather than utilize this, tell the person about himself, tell them why this thing thinks the way it is and how to make that change in their lives.
And I think that’s the critical difference here. I think data science could be much more than a predictive system, it can be an engagement model and I would consider more on what we’re trying to do is data driven design, not just data science for this predictive outcome.
So in the context of the psychological kind of component… Yes, people can use psychological model to try to predict like churn rate for companies and things like that and they’ll have some power there, but why not also say,
“Hey here’s your psychology. Here’s the way you think.”
Help managers understand how to engage with them. Help them understand their selves like,
“Hey, here’s your strong point… Here’s your weak point.”
Choose people like you on how they handle this situation and instead of just prediction of outcomes, you’re actually driving outcome and you’re helping people change themselves and go to a better place. And I think that’s really a big opportunity. If you look in their say, their credit space, most of the time it’s this kind of rear view mirror kinda style where you come, you get underwritten and we’re done. You’ve been underwritten and we give you the loan and you go for it and we hope you don’t default, right?
And that’s just… I think a very limited world view on using a technology just in the small scope in which it exists and like,
“Hey, that model has a lot of information in it.”
Let’s provide value for the individual so that they’re less likely to default on their own. And now, you’re driving outcomes to better people’s lives and you’re using that data to help your business model stay efficient. So I think there’s a lot of added value… I wouldn’t say it’s antagonizing, it’s using the model for more than just this black box solution.
Dhaval: Very, very interesting. So, you know my, next question is what are some of the business results or outcomes that you have seen because of this deeper, more of an engagement models style approach instead of the traditional parasitic approach towards consumer financial behavior?
David: Yeah, and I, you know, I thought about this a lot and what really… so this kind of idea of where’s the proof in the pudding and I honestly can’t find too many examples where this is fully flashed out. I think that this idea of… cause what are the things you need for it? You need repeated interaction with individuals. You need to be watching them, working with them over time. There needs to be outcomes and data related person outside some kind of very spurious spars matrices of the clicked an ad one time in the last three years. And I think that a lot of companies are starting to think about this but I don’t think a lot of them have done much with it but I would say that, you know, you look at burgeoning of this AI chat box trying to bring a humanity to the equation and it’s something that I’ve seen a lot in this screen tech conferences and whatnot is this claim of what is it mean to really develop a technology? It means to bring back this humanity to the system.
So if we look for example the notion of the uncanny value in robotics that when you build a system that’s not very sophisticated, it’s just a smiley face, people will quote that because it’s not that great but if you’re trying to make it look like a human but you don’t do a good job, it really creeps people out and so it’s really not what people want and if you really go over the chasm and you end up with things like the TV show Westworld or it look just like a person and now you have these weird beliefs about it but it seems very real to you. Well, I think all technologies had that problem where they get scaled before they… with scale they’re losing humanity. So for example, typically loans were given and you know between people like they come and apply for a loan and there’s just a measure of character and whatnot. And with the internet age, people don’t get a chance to show that part about themselves and what I think is that by adding in these components of psychology, interaction, transparency, you’re bringing back the human quality to the system and there’s a lot of research in computer human interaction research, how do people interact with machines in more optimal ways and again and again it’s been seen that the more XY growth we wanna have as an experience, the more we appreciate it.
So take for example Pokémon Go, all they did was put it out in the physical world and it was a massive effect. People loved it. Now, admittedly there’s a lot of game things that people have an issue with and there seems some kinda churn on that but in the end it put it out in the world, it made it integrated into our normal human experience and it was massive really engaging and surprising. And I think that we’re gonna see more and more of that and a big part of that is the psychology of how we treat and trust these machines and yes so I think that we’ll see a lot more evidence for this as you go forward.
Dhaval: Yeah, that’s fascinating, bringing humanity to AI. That’s a whole new field that’s gonna grow massively as we more and more into using human computer interactions on a day to day basis.
David: You know what… what’s interesting to me is that, you know, I think it always happens that when a new technology comes in and maybe because there’s a lot of funding for those kinds of things is that we think that deep learning is the solution to some of this AI problem and that I do agree in a lot of ways it is. But if you say, for example, you know, I have a relationship with a plant. I have a relationship with cats, dogs, all… you know, really smart people to think they don’t even actually have a well definable nervous systems. And what did that say about what is the meaningful interaction to me, right? I think that we have to be very… I think, creative and aware about what it means to really have a good value adding interaction for people and how much of that is about the intelligence of the system and how much of it is about the dynamic that sits into the humanity we would expect or want.
Dhaval: Nice, I’m gonna make that a quotable quote. Hey, switching gears a little bit, what are some of the projects that you’re most excited about that you’re gonna be working over the next few weeks or months?
David: Yeah, so I would say that what we’re… here at Payoff, what we’re really interested in is building a good interaction with people that’s long term, that’s sustainable, that’s isn’t a one-off experience. So we’re starting to develop some technologies and what the way to describe exactly what those are but we’re trying to create ways in which we create more than just a coach, more than just advisements but understanding their psychology, drive towards change and really building out that technology. That’s what I’m really excited about right now.
And what’s part of that, what’s really interesting to me is using psychology in ways that it hasn’t typically been used. So there’s a large deep field of psychometric that really has modeled out how do you quantify as a measure human decision making, human expression, right? So they have Likert Scales to disagree or agree. But there’s a lot of math on how to assess and deal with that data from notions of your standard kind of through scale analysis to item response theory and then transforming those into the realm of machine learning where we just start to develop this normaling your ways of interpreting those scores and mapping those in other characteristics. And that’s where… a very exciting to me of the merging of machine learning with classical psychological theory into really getting that as much information as possible out of that interaction. I mean, if you take network for example, the stars, that’s a Likert Scale. All these things out there, some up and down, those are dichotomous assessment. All these things are known tools in a psychometric literature but I don’t always see a lot of the right math where you see appropriate analysis for those things coming out of it. And I’m sure network has an awesome system but in general there’s a lot of ways to leverage psychological theory into that.
Dhaval: Yeah. Now we’re gonna move into my favorite portion of the podcast which is a round of rapid fire questions.
Dhaval: Are you ready for it Dave?
David: I drink some coffee so I’m still on pretty good.
Dhaval: Yeah, okay alright. So, we’ll start off with this question, which is if you could put one data science related billboard anywhere and it can say anything, what would it say and where would it be?
David: Is that the physical? Can I put it on Google front page?
Dhaval: You can put it there. Yeah it could be a banner as well.
David: Okay, I’ll put a banner on a Google front page saying or maybe in San Francisco saying that “Data science is not equivalent to machine learning. Let’s put the S back in data science.”
Dhaval: I like it. I like it. I like it. Alright, next question,
“What’s one data or decision product that you have purchased in the past 6 months for under a hundred bucks that had a positive or negative impact on your life?”
David: Does a cold brewed coffee count? No? Or whisky? No, actually Safari books online when we got that here at the company and I really like their video tutorials and their stuff’s real solid. I really like it. It’s super, super good. Yeah.
Dhaval: Yeah, I watch their videos all the time. I love it. Next is…
David: No, no vested interest with them, I just like it.
Dhaval: Same here. No I have no relationship with them. I just love the product. It’s like I can learn the stuff forever and ever. So next question is,
“When you think of data person being successful, who first comes to mind, and why?”
David: So when I think about that question, it makes me wanna say… look I have this impulse to say Nate Silver, right? I’m just impulse to say cause I’m deeply jealous of how good he could visualize information and just how effective he is at his stuff. It’s just really awesome. I’m a major fan. But what I also wanna say, you know, I really like Einstein. I like Galileo. I like the, you know, the people who use data to change the world at scale, right? Even though maybe it took awhile to get that scale but I think that it’s important for us not to forget the algorithms and the science that is used… maybe it’s not so connected to start up and may sound maybe some of the will-be, maybe Einstein’s unintentionally critics from start up, right? But I think that… those are my heroes. People who have kind of really challenge the status quo with data driving their belief structures. I really appreciate that and I think that that’s a trait that doesn’t have to just be that of a data scientist. It could be a trait of anybody who’s developing something and I’ll say monetarily for companies, this is the strongest opportunity to make that real for them because data affects a lot of things. I just appreciate anybody who’s really solving problems with the data driven point of view. I once sat in this class of a high school in the summer next to the sixth grader and with the geometric problem that no one could solve and he solved it with this analogy of hamburgers and he was just like shockingly brilliant. And I really appreciate that. People who are willing to let data influence and drive decision making… I mean not exclusively and let that, you know, have opportunity to change people in big ways.
Dhaval: Very cool. So next is in your data related book collection, which book to you often go back to the most?
David: Yeah, so I’d say it’s very few and I guess it’s because the internet and that’s unfortunate the way it can kinda fragment our learning but the two books… I had a professor named Roger Kirk at Baylor, it’s called ‘Experimental Design‘, this giant red book from that… if you saw it, you’d walk right passed it and you basically are forced to derive every single aspect of classical statistics and I could be… I can’t be more honest, that has been the more fundamental, almost habit forming tool around math that I’ve ever had. And then the other book is ‘Probabilistic Robotics‘ by Thrun. I think its Sebastian Thrun, I’m not sure. And the reason I like that is cause that’s kind of a book that got me into machine learning itself. I’m a very example driven person and robotics is all about examples cause in applied use of a lot of interesting models like Kalman filters, Particle filters, Bayesian integration model and it’s just a great, great book that walks very clearly through this algorithms.
Dhaval: Well, thank you for sharing that. I can’t wait to check them out. This is my favorite question of the whole podcast actually which is,
“Which story do you most often use when you’re communicating to the C suite about machine learning or data science?”
David: Yeah, and this… I think connects to… I mean the reason I even made that YouTube video in the beginnings is… the YouTube channel… was that I was really frustrated that as things got more complicated, the discretion in education around it got equally complicated. People tease the easy things easy and the hard things hard and it just boggles my mind. And I deeply believe that nothing is truly complicated, it’s just a lot of pieces and that makes it complicated but there’s still conceptualization and relationship between the information that is communicable. If I can’t explain the equation, there’s something… there’s something deeply wrong about that in my opinion. And so with the data science, I’m using a lot of analogy because our brain is a big, nice, deep learning machine and I just speak to intuition about those rather easily. So for example, when we talk about the probabilistic of nature of classification I go,
“Well, think about it.”
Everyone right now are thinking of a cat or think of a dog. You all have kind of your own exemplar in your head and you use that as a model to assign different pictures of different animals. If I show you this kind of raggedy, fluffy cat or show you Garfield or I’ll show you, you know, a really tiny cat or baby cat or adult cat, you don’t really have much trouble classifying it yet you don’t have this one idea of a cat that everything fits in really nicely too. And I could give you this cat like dog pictures and as soon as I put those up at slide, well people don’t know how to classify it. And I’m like,
“That’s exactly it.”
Classification isn’t perfect and because… cause the objects in the world aren’t really that well defined either but we develop these examples or this template that allows to classify information. And that’s a big part of machine learning tools and classification and I think that… speaking of those kind of intuitions, our really get at that the beliefs people have and going,
“Hey, okay I get it. There’s such things I understand.” Like for example, another one…
Dhaval: Yeah go ahead.
David: … another one say, cluster analysis… I like giving analogy of trees and bushes and ferns or whatnot saying that,
“You know, a lot of things have the same attributes.”
Things are… like trees have leaves, bushes have leaves, they have barks, they have stems, they have trunks but you know there’s… that you can tell a big difference between a bush and a tree pretty quickly. Why? Because the relationship of those attributes are very different. Bushes have small leaves and many little branches and not big trunks, whereas trees have big trunks, fewer branches, fewer leaves but bigger leaves. And those are related traits and through that when we use a lot of machine learning, we can automatically discover those population. You know and I think that’s just related to the way we normally understand information, has a lot of opportunities to really gel with the intuitions people have.
Dhaval: Very well, very interesting. I can’t wait to incorporate some of that on… some of those things in my own communication. Thanks for sharing that Dave. You know, I really enjoyed this conversation. You are a high octane, high energy dude and I look forward to continue to stay in touch with you and bring you on board with your new stories, with new adventures sometimes in the future but in the meantime if people wanna get in touch with you to get your advice, to get your info on something that you said here, what would be the best way?
David: Yeah. By the way I mean I really appreciate it. I love this opportunity… well there’s chance to talk about it. Thank you for having me. I think through LinkedIn or you know David Herman or you have my email dh.herman “at” gmail.com and I have my YouTube channel called Student Dave. And I promise to make a new video. I’ve just been rather busy for the last couple of years.
Dhaval: I just saw that channel and it was pretty awesome.
David: Some of them comes from a very dark place like my old lab in the dark corners of the room.
Dhaval: And that’s the beauty, that’s the beauty of creating videos, right? It’s like a data point in the time series of your life and you can go back to it and you can evaluate your qualitative properties around that timeframe. And reminisce about it, enjoy that.
David: Yeah, well one of the videos was made in my old apartment that I’m pretty sure it’s condemned now so it was fun.
Dhaval: Well, it was great to have you on the show. Thank you so much for being our guest and I’m looking forward to get this out to the world and share it.
David: Really appreciate it, thank you.