How AI Is Changing Investing— with David Trainer

0:00You know, there's a problem with AI. Like, you can ask Claude and regular Gemini and all these things, all kinds of questions like, hey, what are the best stocks in a sector? It's like, you know, they're giving me the most popular stocks, you know, or you get hallucinations. And where Google Cloud and new constructs really fit together was this belief that we can't have a generative AI that answers everything. The idea that you can just pour the internet into a large language model and all the answers will emerge. You're tuned in to the Investing for Beginners podcast. The show for the long-term investor.

0:36We cut through the noise to focus on what works. Compounding, discipline, and the conviction to buy wonderful businesses and stick with them. Your path to financial freedom. Start now. Welcome to the Investing for Beginners podcast. Got a special guest for you today, two-time guest, David Treanor, CEO of New Constructs. New Constructs is an independent research firm that helps investors with an AI tool, and we're going to talk all about AI. David was gracious enough to do this interview a little bit early, so by the time you guys listen to this, there might have been some things that have changed in AI, but uh thank you david for for joining and and being flexible around my schedule with with my baby coming and everything my pleasure andrew uh glad to do it and these are the most exciting times that pre-baby time your world's about to get rocked so uh enjoy yeah the nesting just started like last weekend and then like my wife and i are google calendar people so next weekend has a Google calendar event that has to do with further nesting.

1:51So it is an exciting time. And I'm just, I'm trying to get as much sleep as I can. Like I fell asleep watching the Lakers last night and I've been sleeping ever since. Cause I know I'm going to, I'm going to have quite the deficit pretty soon. Yeah. It's, you know, it's funny. You will, you will find having been through it three times myself, you'll find that you have resources, energy that you didn't expect. It'll come, it'll kind of come in droves. And it's part of the magical nature of it all is that it's, it's, you can do it. It's been done a long, for many, many years. So you'll find, you know, you'll find it's, it's, it's pretty magical and things will, things will work out really well.

2:31I appreciate that. Well, speaking of things that have been around for many, many years, analyzing financial statements and doing the right kind of fundamental analysis has been a key strategy for investors who are picking stocks and doing well over a long time period. With all that in mind, I would love if you can give us the backstory of how you came up with your AI tool. And specifically, I know you talked with the folks at Google and how they've helped inform some of the ways that your AI tool works today. So I thought that would be interesting place to start so people can kind of hear that evolution of how you're using AI to help investors.

3:15Yeah, I think evolution is a good term to use, Andrew. And I'll share a slide that I use in a lot of meetings I do about how AI is affecting investing. We've had an AI platform, I would say, for, I mean, had been building it for over 20 years. And my AI and the way I think about AI is not just a chatbot interface, which is great. We were talking a little before the show about how much better effectively just AI is than traditional search because you can actually put in a question and get an answer, not just a list of websites with similar keywords, right? And that's just like, that's a huge, almost quantum improvement over traditional search.

3:58But the real gap, I think, between where AI is now, where people want it to be, where you can trust it to do things more than just menial tasks or simple things that you can assign sometimes junior labor to do. The gap there is only gonna be filled when we're able to successfully endow machines with real deep subject matter expertise so that they're capable of making sophisticated suggestions and insights on par with human experts. I think we're a long way off from that. And the part that got Google and New Constructs together was, you know, I was with a friend of mine who runs professional services at Google Cloud.

4:44And I said, hey, so what are you guys doing with all this infrastructure? You've got more computing power than, you know, anything in the history of the world. And you're building and building and building. But, like, what are you seeding all this computing power and storage and chips? What are you seeding it with? like what is it using to make decisions what do you how's it going to do sophisticated things and he went through a couple of case studies including i think a huge project that google cloud did for a weather organization where they got tons and tons of data that's that's almost mathematical right and they've done some really super cool things with those guys but for more subjective elements of society where people you know really want advice they didn't have they don't have anything.

5:26And I explained what we were doing. And he immediately was like, yes, we want to build an agent based on your data. This is the secret sauce we need to show the power of our AI. Because if we can match our AI tools with truly reliable data sets, 100 % accurate data sets upon which real deterministic rules can be built, then we got something pretty powerful. And so last summer, we probably started talking in late June. And I mean, by the end of July, we had a prototype. These guys move so fast. Andrew, I've been working with huge, sophisticated financial services firms for decades. Google did more in a few weeks than I've ever, with the most efficient financial services firm that we got done in like three or four months.

6:15I mean, it was amazing. And they were really, really psyched to be able to use this data set and have an AI agent where you can go in and say, show me the company's most likely to miss earnings in my portfolio. And the answer you get is based on the same data that Harvard Business School and MIT Sloan wrote papers on. And now there's a live traded index from Bloomberg that beat the market last year 27 % to 18%, the market being the S &P 500. And it's all based on superior earnings measures. And so, you know, you're getting an AI that's based on research and a data set that's been proven to outperform the market statistically by academics.

7:02And then as in live markets, dramatically outperforming. So it's almost like you've got one of these, you know, if you could download the brain of a hedge fund guy who beats the market all the time to give answers, it's kind of like what our FinCites is what we call it. FinCites agent is able to do. And it's super powerful. If you want to understand profitability and valuation, we're giving you arguably the best answers in the world on the companies we cover, right? And vouching for that, again, Harvard Business School, MIT Sloan, Ernst & Young, and then live trading performance. And so we think it's a paradigm shift.

7:40And filling that gap I was saying before is a problem with AI. You can ask Claude and regular Gemini and all these things, all kinds of questions like, hey, what are the best stocks in a sector? The answers you get are, it's like, they're giving me the most popular stocks, or you get hallucinations. And where Google Cloud and new constructs really fit together was this belief that we can't have a generative AI that answers everything. The idea that you can just pour the internet into a large language model and all the answers will emerge. We clearly know that's not working. So what's the other strategy?

8:24I think it's pretty straightforward. We have to break the world down into small enough pieces or domains where we can build, we can trust that humans can build a perfect data set upon which reliable deterministic rules can be designed and built. And that powers an agent. You know, an agent, AI agent being focused on a very particular domain. And once you can trust that that agent and that data set are 100 % reliable, well, then you've got something special. And then you get the agentic workflows because you can stack up a bunch of agents that are all reliable, but now you can do some really cool things.

8:59And I believe that's the way it's going to work. We've kind of think about it more organically inside out. We've got to conquer different domains and then put that together into a larger decision making system. And that's where really AI will have the sort of Skynet singularity kind of effect that people love to talk about. I'm kind of thinking like walled garden. Your input is you keep the input non-garbage through a walled garden and then everything else can kind of compound and flow. Is that kind of the idea there? 100%. Like that's one of the things about our agent. Like we say in our sort of training materials or introductory materials, like this is not talking to the internet.

9:46Because when it did, it got bad answers. Like we had to like, you know, the way it works at Google Cloud is that, or Google in general is that they used to have a lot of teams building different AIs and they kind of all put it in one spot. So they all use the same product. And they had to do a lot of very specific things to say, do not go to the internet. the problem with the internet is that is that there's a lot of good stuff out there there's a lot of bad stuff but the machines don't know they have no way to discern oftentimes humans have no way to discern between what's good what's reliable what's popular what's been hyped what's been sponsored you know all these kinds of things and and so the internet is like everything in the world you get a lot of it and you don't know how much of it is good uh you know there there's no There's no regulation of it, which is a good thing.

10:32But in order for us to make decisions, we do need to regulate our data. We do need to have a walled garden that says, for the purposes of answering these kinds of questions, these will be the only data sources allowed. And that's not an easy thing to do. It's very difficult to find the one version of the truth about anything. But at some point, you've got to do something like that, right? You know, to live our lives, we have to have at least a narrower range of inputs that we allow to dictate our behavior. Otherwise, you know, we would be constantly paralyzed by overwhelming amount of information.

11:12And that's the role of discernment is to help us choose those inputs. And the same thing with people we allow into our lives. Like discernment is a part of being a successful human, and it's going to be a part of being successful AI. your ai roi is the sum of the five agentic agents closest to you how about that that's right it's a good way to put it it's a great way to put it uh yeah and i think that's often overlooked andrew i think this is the big thing that people you know look we always want to believe in in magical powers right like going back to the caveman days we wanted to be like yeah you know the sun has moved you know across the sky by a magical force, you know, or the gods are controlling this or whatever.

11:57And I think that's just an ingrained human bias that, you know, I don't really need to understand how it works. It just works. And that's great for me. And I think that's where people, a lot of people are getting suckered into sort of misuse of AI, you know, misunderstanding how to use AI, because it's no different than any other machine or model in the history of the world, which is it's those machines and models they're only as good as their inputs. You put bad fuel into a Formula F1 car, you're gonna have problems, right? And I think these AI models like Formula F1 models. But when you just kind of think about, and I remember my earliest lessons as a young analyst was garbage in, garbage out.

12:42The models don't do anything magical. They just organize the inputs better. So if you got bad quality inputs, you get bad quality outputs. There's no alchemy in models. They didn't have a way to make data better. And I think the other thing people misunderstand about data sets that's super important is that when you have a bad data set, you know your data is not 100 % reliable. It makes it that almost nothing is reliable in that data set. Because this is simpler than most people realize. Let's say your data set is 99 % good. okay well it would be 100 good if you knew which one percent was bad you just take out the bad stuff it's smaller but you know at least it's 100 good and that's what you need but if you don't know what of that 100 which one percent is bad how can you trust any of it with important decisions you know and so when it comes to movie recommendations or restaurant recommendations, not the end of the world.

13:48When it comes to medical recommendations or financial recommendations, you don't want to take the risk that you're getting that bad 1%. And by the way, I think I'm being generous about 99%. I think in some cases we're talking 50, 60, 70, 80. We're lucky if we're 90. And people don't have that discipline to say, you know what, either it's all good or it's all bad. And that's really the way you have to think about it. Like if you don't know which 1 % of the information you're going to get back is going to be based on the faulty data, you're taking some big risk. And that's something that has always kind of frustrated me for years.

14:22You know, we wrote a piece about AI and technology for investing. I wrote this back in 2018, and we're making the point that, look, man, data is everything. And you don't really clean data. right you don't you know how do you clean you take a hose pipe and like wash it off no i mean think about going through the process we may have talked about this last time too andrew cleaning data means well i got to go and validate it against the source otherwise how do i know it's right which is as much work as getting it from the source to begin with and you realize everyone's trying to avoid that work everyone's trying to avoid like oh going and harvesting data from the source and getting it right because by the way it's a difficult It's onerous.

15:08It's tedious. If you don't have the right taxonomy going in, you could potentially have wasted all your work before you know it. And so it's a tough thing to do, but it's absolutely 100 % essential. And so, you know, you don't clean data, you replace it. So when people say they spend a lot of time cleaning data sets, I'm like, well, you know, what does that mean? You can't either you replace it or you verify it. and you don't really have this idea that, oh, 90 % confidence in the data set. It's like, all right, well, then you just didn't check a whole bunch of it. So we don't know what we have there.

15:43It's really simple on that front. And I think the lack of clarity around that, a lack of forthright explanation of what's going on is a big red flag. If you don't know what's going on, you don't know what's powering your AI or your model, that's a big red flag. Don't trust black boxes. So what does your firm do to maintain the integrity of your data set? It started 20 plus years ago, Andrew, when I designed a system for really all around data integrity. I'd been on Wall Street and I'd built a bunch of these sophisticated earnings models that took into account footnotes and all the stuff you're supposed to do.

16:27I was sort of in the automation of models business back before the tech bubble. So I had a lot of experience going through a lot of filings. And one of the things that struck me during all that time was like, when you trust humans to do it, you're going to have problems. These are people I would train and check the work. But when you have humans doing the work, unless you're checking every single thing, there's no scalability. You don't know they're doing it right. You know if they're shortcutting. and and then when I as I as I stayed on longer on Wall Street I realized that very very few humans wanted to do this kind of work even you know senior analysts successful senior analysts making millions of dollars a year you know weren't interested in going through the footnotes they wanted to talk to clients they wanted to go on tv they wanted to you know they wanted to do more interesting things so in the beginning I designed a very simple system that connected the filings a parsing tool in the database all together so that we could have looked through transparency into everything that was going on, everything could be validated.

17:33That database was also connected to a model. So we were producing results. We were eating, cooking. A lot of the original data collection firms, the legacy data collection, financial data firms, you name them, all the big ones, they don't have any models. So like people are collecting data blindly without understanding that, oh, that number I collected for... for, uh, I missed the number for Sam's revenue at Walmart. And so the total revenue number was off by like 20 % and nobody noticed, right? Because they're like not using the data. That's a real use case. It happened a few years ago. And so we had that from the beginning and, and then we had clients checking the data early on too.

18:15I was really fortunate that Fidelity was a very early a client supporter of what we were doing. And then we had a lot of people looking over the data internally as well. But there are things that we could build while collecting data to ensure that the models were sort of self-verified. And some of it's easy stuff, like making sure the balance sheet balances, making sure the income statement adds up. But we took that approach to sort of every single data point. Operating lease tables, all those lease payments have a sum to a total. Now, a lot of things you can't do that for. But we developed over the years and having parsed several hundred thousand expertly parsed, humans parsed and modeled and checked, right, and now our signal manifests in the new constructs, Core Earnings Leaders Index, live publicly traded, you know, that's working, right?

19:11It was up 27 % last year versus 18 % for the S &P 500. So we've seen the signals validated, but over time we were just constantly checking. And then we started using, you call us an AI. We write a little algo that would identify any anomaly that we, for example, easier to put it in a real world situation. I get an email or call from a client saying, hey, what's this going on with this number? And I go look it up like, oh, we made a mistake. And so the team goes back and says, we're going to write an algo that will automatically identify any time this scenario occurs again. And it will alert an analyst to take a look.

19:49This seems like a problem. And eventually, the analyst fixes it enough times. We see it enough times that we can automate the fix. We see X, we do Y. The machine sees that pattern enough, boom, it does it. So we've been doing that kind of work for 20 years. while also having the data out there with a lot of really sophisticated clients, firms like Goldman Sachs Asset Management, Two Sigma, some of the best and most sophisticated folks in the world, and portfolio managers checking the data, checking us. And then we always were getting machines to do the work as much as possible. So kind of back to my original point, my AI, in the beginning, humans were going through all the filing themselves.

20:32Now, we had tools to automate a lot of the process. Like, you know, we could create little easy programs that would like virtually automatically get the income statement. Right. And so one of the other things also, Andrew, like there's no typing stuff in zero. Right. Everything was sort of selected pointer or, you know, knew how to grab a phrase or a number or understood that a phrase and a number and a table was something to be collected together. nothing was typed in it was all directly from the filings always so we could validate to the source and and so that was part of it and then the machine would track everything the humans were doing every action from the beginning we started doing this in 2003 and over time you build enough human actions that is a effectively a library of instructions for a machine to do stuff on its own.

21:25And so it starts with simple stuff like, hey, you know, you're going to parse income statement, balance sheet, cash flow statement. And then the humans don't spend any time on that. And then they can focus more on footnotes. And then the machines get better if the footnotes are in tables because we can validate those. Now, some footnotes are a nightmare anyway, like pension tables, because those are almost always different. You've got US, non-US, Then we get good enough to master those because my smart humans now have time to teach machines to deal with the anomalies that we see in certain kinds of tables.

22:00And then there's now there's time in the end, you know, even more now time to to just look for those unusual gains and losses and footnotes that are very idiosyncratic. And it's difficult to systematically get all of those all the time without humans, but we're getting close. I mean, we've probably got, you know, several million. Each one of those is a scenario that if the machine sees again, take it. But companies are always innovating. So we've always also had the rule that if the machine isn't 100 % sure, red flag to the human, fix it. And the human fixes it, and the machine can be sure about that.

22:48And so if that exact same scenario presents again, we got it. So it's kind of like a way of, you know, over time winnowing down the number of ways that companies can disclose and report things so that the machine does more and more and you got humans doing more and more sophisticated things. And then we can just cover a lot more companies as well.

23:09I'm wondering how excited you got when AI started to really start to evolve into the way it has now. Because to your point, you've been working, grinding, improving on this for like 20 years. And then now the box has opened and everybody's trying to kind of do what you guys have been doing. So I'm curious like what that was like for you to see all this innovation. And then you mentioned the agentic. I don't think we've talked on the podcast about what that even means, agentic agents. But if you could give us like, I'm a dreamer. So I like to dream and like think of what are the different businesses that could be built from stuff in the world.

23:55So like, could you paint us a picture of like what an agentic agent could do and how it could help investors and just kind of your overall thoughts on some of the advancements in AI as they relate to what you're doing? Yes, we were very excited about AI and machine learning, even when that first came along. Because to be honest, people didn't really understand what we've done. Especially a lot of traditional legacy investors have been around for a long time. They didn't really understand technology. They just didn't understand. They didn't get it. And now with machine learning and AI, people get it.

24:32Like, for example, my buddy at Google Cloud, like he got it immediately. He's like, oh, wait, you know, this is, you know, it's all about the data. And we're trying to show clients use cases how great our technology is. But if the data is not right, the machine is going to give a bad output and they're going to just discard the whole thing. The clients will. So having, you know, a truly reliable data set was a big deal. and yes, now people have a rising awareness around importance of data. I see it more and more, and that's helpful. I think it still needs to grow a lot. I think because we're fighting against this ingrained human tendency to want to believe, oh, it all just works.

25:10I don't even know how it works. I don't even look under the hood of my car. People used to at least look under the hood of cars before they bought them. I don't even look ever at all. Just drive into, you know, with valvoliness and oil change, they do it all. You go on. You don't want to have to look. Just works. And people want to believe that. And I don't blame them. I do too. But for things that are really sophisticated, we still have to be mindful of what kind of fuel we're allowing or what kind of information we're allowing into our lives. We do that with everything. I don't know about you.

25:39I don't watch every single newscast, right? I mean, I got to be selected. Mostly it's zero, but you get the idea. And so it's very helpful that people are now developing the language, the tools to understand what it means to have machines in our lives. Because that's a new thing, and we've got to get better at that. And I think that's related to the whole agentic workflow idea, Andrew. And really what I think the world means by agents is a recognition of the limitations of AI in that you can't just pour the Internet into a large language model and have it figure everything out. So what does that mean?

26:27We give it, you know, we pour less into it. What does that mean? Well, let's narrow, break the world into small enough pieces so that that one piece, we can be 100 % confident that everything is exactly right. Now, in the beginning, that's tough because it's like, well, I can only do that on a really small part of the world. And when you want to try to get into more complicated, sophisticated things, it becomes a really tough task. But that's what you got to have. You've got to have, you know, and just think about an agent is like a smaller, specific domain focused piece of AI. And so, you know, for that agent to be 100 % reliable, you've got to have 100 % good data into it or driving it.

27:09And that's the key. And that's why we're hearing about agents. It's saying, you know what, we're not going to pretend the large language model can do everything. And they don't put it to you like that. They just say, oh, there's going to be an agent that's really focused on this one thing, and you can use it for that. You can ask him questions about a particular subject matter. But that agent will only be as good as the subject matter expertise that's programmed it. So that's what an agent is. And the idea is if we can have several, lots of truly reliable agents, each trustworthy for their particular subject matter, well, then they can work together and do impressive things.

27:54Not dissimilar from what humans have done over time when we've gotten together with a lot of different experts to try and build something special. Like what we did with the first atom bomb. What do we do? We brought a ton of experts together and we need diversity of thought. We need different kinds of people to think up and be creative about how to do things that have never been done before. And so it's the same concept. It's just that, you know, the agents, you know, you got to make sure that they're real experts. Right. The last thing you'd want and in a, you know, a collaboration, if you were, you know, with a bunch of humans and you're on a special team to figure out how to solve some major problem, the last thing you'd want is somebody on your team in there just filling you with a bunch of misleading information, acting like it was true.

28:45It's about how counterproductive that would be. Right? And so you've got to have the same thing with agents. That's cool. So you mentioned earnings.