New Constructs Monthly Podcast - September 2025

Welcome. Let's get going.

Yep. Started off with a little quote here for the topic of the day. Our AI agent provides real, reliable insights. You may ask how? Because it is grounded in superior data and built on immutable core principles for intelligent investing. It's like our research. There'll be much more on that today. So we'll kick it off there. This is our monthly podcast. We titled this one Separating Fact from Fiction: See Our New AI Agent for Investing. I'm Kyle Guskey, and we are here with CEO David Trainer. The disclosures and disclaimers, mainly that we are not providing advice. We are publishing research. These will be in replay. You can check 'em out at your own time when you watch that replay. Going for the agenda today, off the top, I want to note the next one. Like I said, this is monthly. The next one will be on October 10th and talk a little bit about the latest research we've put out since the last time we were on the monthly podcast. I've had some long idea updates, danger zone picks, put out an updated IPO report on Klarna after they failed to IPO in, I think, April of this year, and they actually successfully IPO'd this past week. We put out an update on that before they went public, um, closed a few positions, updated some long idea danger zones, updated some focus list stocks as well, and put out a few different earnings watch parties. We love doing that during the earnings season to dive in into what's going on, you know, the day it happens, you know, the day after, taking a look at what the earnings numbers look like compared to what our model shows and how it kind of interacts with the DCF. And so it's a great way to get a hands-on look at our whole system. David goes through those. And then we updated the 2Q25 Footnotes Season.

We'll talk a little bit about that today.

And then the AI agent that Noah wants you to see, Training Replay, that is on our website as well. These links will all be available in these slides that we'll share on the community. But big items for today, we're talking Separating Fact from Fiction, our new AI agent for investing. We'll talk a little bit about the actively traded indexes, how they continue to outperform. Update on the footnotes season, some upcoming training sessions, posts on community, and then any questions. If you have any questions for us, please send those in the Q&A or the chat. We can get to those throughout the presentation today, or, you know, as it comes up in topics. If you have questions, be sure to send those in as well. But with further ado, we'll get into the main topic for today: separating fact from fiction on teaching AI to pick stocks. Started off, we actually posted, published an article recently titled exactly that, How to Teach AI to Pick Stocks. And it really dives into this idea that getting AI to pick stocks, it's the big elephant in the room. How can we get AI to pick stocks? But it appears it's a long way off using LLMs. The article goes into exactly why. Without explicit guidance, LLMs can't navigate the nuances of accounting. Studies showed that even the best LLMs achieved only 17% accuracy when linking extracted numbers to the correct concepts in US GAAP taxonomy. So you've got over 80% of, you know, the data and the numbers being incorrect or unable to be linked to US GAAP principles. That is a problem when you're trying to drive insights and especially reliable signals. And without accuracy in that data being mapped correctly, you can't generate reliable signals. Ultimately, broad LLM using unstructured, inconsistent data cannot develop accurate and reliable financial analysis capabilities. I think we've seen more and more evidence to that as people have started using LLMs and try to use them for everything and anything, and they run into problems when they try to go very specific.

Yeah, I think it's really important to, for people to understand that AI is not alchemy. You can't make gold without gold. You can't create intelligence without intelligence, and machines are not born with it. They have to be taught. The intelligence. And as a result, it's super important to understand that these machines are only gonna be as good as the data given them. The same thing that people have been saying about models for, for decades since models were born. Results are only as good as what you put into them.

Garbage in, garbage out, right?

Garbage in, garbage out. That has not changed and it's not gonna change for a long time. When it does change, I mean, then we got Skynet. That's like a singularity. That's a different world. And nothing we talk about now is probably gonna even matter in that point. But as long as we, we're still living in a world where humans matter and, and we have somewhat some control, at least perception of control, we have to understand the machines are only gonna be as good as what we teach them. And so it's gonna be really, really important for people to understand and focus on what do we teach the machines? The machine that you rely on, what has it been taught? We all know that with the general GPTs and Groks and things, They've been taught the internet. And that means that there are significant and severe limitations on the sophistication of the prompts and tasks you can ask them to do. So New Constructs is, I think, helping people go to the next level with AI by providing a dataset that matches a dataset and a set of rules and an ontology that's as good as, as what any expert can put together. And when you build an AI agent and a system on something like that, you get something special.

So we got the exciting news. We are pioneering that first AI agent to produce truly reliable results. We're actually going to do a live demo. I'll turn it over to David in the next slide here.

As always, any questions, send them over. We'd love for this to be as interactive as possible. But yeah, so we created this with some super smart engineers over at Google Cloud, and I'm excited to share some really cool stuff. that you can do with our AI that you can't do with any others, right? And let's just talk about, how about getting a warning about which stocks are most likely to miss earnings?

So let's ask, which stocks have the worst earnings distortion scores? And why don't we narrow it down to a sector so that we don't have to get a huge list? We'll say which tech stocks.

We'll watch this thing do its work. Ah, there aren't any tech stocks with an earnings dispersion score. So why don't we just set, make it a little easier?

We'll just say all stocks. We'll take out tech.

And while it's working on that, I'm going to write another prompt that says, tell me which stocks in my portfolio are to miss next quarter. And then you can put your portfolio in here.

Just say it's a, a, a, a, a. APZ. Here we go.

10 stocks with the worst earnings distortion scores. So we've set this thing up so it never gives away more than 10 stocks at a time because we're not giving away the whole database, but you get the idea, right? This can immediately tell you. You try this on Perplexity Finance or ChatGPT, you're not going to get anything like this. How about the stocks in my portfolio?

I'll go and get a couple other prompts here.

Put some suggested prompts in the Q&A or something like that, or in the chat if you guys are interested in other prompts. None of the stocks in your portfolio So you're clean, right? These all get the distortion score of B. Super cool, right? How about the best tech stocks?

Why don't we line up what happens when we ask this out of Perplexity Finance?

Please try ETFs as possible. We don't do earnings distortion scores on ETFs because they don't report earnings, but we do have ratings on earnings.

So why don't we ask which stocks in the tech Sector are most likely to miss earnings. Let's see what it comes up with. What are we going to go with? Here's some of the best stocks in the tech sector based on New Constructs ratings. For my portfolio, please give me the rating, earnings distortion score, ROI-free Cash flow. Let's put in AABZ. We don't have all the tickers in here, so that's why I'm just making some stuff up. What does Perplexity say?

Oh, big surprise.

So it really doesn't give you any idea in terms of what's going on in terms of missing. Which stocks have the worst ROIC over the last year? We could do that.

Here's the information for your portfolio.

Right? Bam.

Here's another super cool feature.

Boom. Immediately into Excel. Pretty cool, right? Which tech stocks had the worst ROIC over the last year? Hope it's not too painful watching me have to type.

So we'll let this one go here. Why don't we ask the same thing of Perplexity? Stocks most likely to at risk of missing earnings. Key sector concern for Risk factor, high valuation, slow growth. You know, I don't know if this adds some elevated valuations or slowing growth.

Yeah, it's not bad. Not bad.

Not based on anything telling you what's really going on with the numbers that might be funny.

See what it says. NMTC, Palisade Bio, other medtech. Let's see what how their ROIC compares to ours.

These are the ones we have. We're only looking at a subset of companies in this set, so it doesn't have the whole database yet here.

And let's see here. Let's see. So these guys, let's see.

We may not have this one in our system right now about Adobe.

These guys do not have negative ROIC for sure. Doesn't show the ROIC, just says declining. I wonder, we can actually, why don't we just look at our system? What is, what do we show for NMTC and PLI? That we're very—

we have either of those in coverage.

We don't cover either one of those anyway.

No.

All right.

Which— that first one, the PALI, is a $5 million company market cap.

Yeah. Which stocks have top quartile and 2-year free cash flow and declining number of shares? Let's see how that does.

Thank you, Forrest, for that. We've got a couple nice suggestions for prompts here. Love it.

Let's see what Perplexity can do with that. AppLovin, Nvidia, Apple, Microsoft, Meta. Interesting. 2-year free cash flow growth.

So let's see what we say for AppLovin in terms of ROIC.

I believe it's way different. 62% ROIC. These guys are pretty close.

ROIC for Nvidia says 30%.

Is that right? Not really. Not even close.

Apple, is that right?

35%? Not even close. Microsoft, 30%? I think this one will be closer. 24%. Meta, 36%.

So not so close. 2-year free cash flow, we didn't take a look at that.

We could do that at the same time. 2-year free cash flow, zero. Not looking so good, but this guy thinks it is good.

Couldn't find any stocks that meet all the specific criteria.

Very attractive ROIC rating, but Positive 2-year average free cash flow, and a declining number of both shares issued and shares outstanding over the last year.

We're going to be a lot more discerning. That's the benefit of new constructs. Those kinds of good stocks may not necessarily exist. Again, we're not dealing with the whole database, but at least you're seeing how you can get much deeper insights from this agent that we've put together in just a couple of months compared to what all of what Perplexity AI has put together.

in the last, I don't know, several years and several billion dollars later.

We can also ask it the simple stuff like, what's the rating for AAPL?

Show the distortion for AAPL over the last 5 years.

Apple got a neutral rating.

We're going to look at earnings distortion, compare Core earnings and net income for A over the last five years. Put in a in a line chart. There's the earnings distortion scores for Apple over the last five years.

Now we're going to go look at the data.

We're going to compare core earnings and net income, and put it into a line chart. Let's see what we get. Try this in perplexity.

See, a question here came in from Chris about how many stocks do we cover. We're right just over 3,400 right now in our active coverage list.

We don't have everything. This has been a test database. So there's—

oh wow, they have a core earnings number. Where's the source for that?

It's referencing us.

There it is. Annual adjustments for Apple. There is no line chart. Says there is. I wonder if that's anywhere close to right. Estimated quarter earnings adjustments, $289. So what do we have here? So let's look at the difference. Let's ask it for the difference and see how that compares. Show the difference between—

Got a question if this will replace the current screener at New Constructs eventually.

Yes, this will be a better way to—

it's a lot easier to use than the current screener, and this would probably be more like an institutional level service. So we'll would be a higher level because, you know, depending on the guardrails that we put on this, you're able to basically get the whole database, right? So we're going to have to figure out how we want to do that. So here's what we're showing, the adjustments. This would be the difference between core earnings and net income would be the adjustments. So we're saying $179 and $37 million. Perplexity's got very different numbers. But I do like that, hey, they're talking about us. This is where you need to go. If you want to get the core earnings numbers, they just don't know how to do it. They copied this from the paper, and you can get those categories, but you're not going to get the answer.

We get line chart. There's our line chart, core earnings and net income. Again, can download all this stuff straight into Excel, which I think is super cool. I don't think you get to do that at all in Perplexity. You're not able to— sources, answer, stuff like that.

So. Yeah, I hope you guys like this. Would you have to provide portfolio each question? Say if you mix of 20 stocks across multiple accounts? I'm not sure what that question means. Typically at New Constructs, we track things by portfolio. So, you know, you can look at multiple portfolios and get different ideas.

So, like, for example, here's the Dividend Aristocrats, and, and so it's broken up by Portfolio.

I think they were also referencing in the questions you asked, like, in my portfolio, you gave it tickers in the AI agent there.

Yeah, no, if you wanted to throw in all the tickers at one time, you could do that too. I was just giving a sample.

So if you wanted to put in a ton more tickers, it would work on a larger list.

I was just doing a sample because this is still a beta prototype. I think it's pretty awesome. And I think the, you know, the explanations you get in terms of how to do things are way better.

You know, it's interesting. Let me ask Perplexity, how do you calculate ROIC? What do they say here? This is the right one.

NOPAT. This is the shorthand way to get to NOPAT. And this is a very shorthand way to get to invested capital.

Pretty good from Investopedia. All right, who is the best measure? Which providers offer the best measure of ROIC? I did this on the new Google Finance and they mentioned New Constructs, so that was pretty cool. Interesting.

Accurate, consistent, well-adjusted calculations based on clean operating profit and appropriately defined invested capital. Among the top recommended sources are Clarify Capital.

Never heard of them. Never heard of them.

Morgan Stanley sell side. Doubt it. Investopedia.

Yeah.

And financial platforms like Yahoo. Yeah. No idea.

Interesting. Interesting. What about— see what happens when we ask ChatGPT.

Oh, look at that. They got new constructs. Bloomberg and FactSet Morningstar. Okay.

Interesting. That's pretty cool.

Widely cited for economic ROC detail forensic.

That's awesome.

Question about how our earnings distortion scores performed over time, showing welfare in terms of earnings. Beats and misses. I can do better than that. I can give you, show you an index that's based on core earnings and show you how it's performed compared to the S&P. How about year to date? We're up 20%. The market's up 12. Over 5 years, 140 to 98. I can show you when we look at an individual stock, we will show you the earnings distortion. Over time. So GAAP earnings and core earnings, as well as the earnings distortion compared to GAAP earnings. We show you that. We don't have it lined up company by company by stock price at this point in time. It's a good question. Good question. We do have that on our ratings in general.

If you go to the summary and you look at rating history, we'll show you this, and you can actually download a CSV that will show the rating as it is every single day. Going back to '06, I guess, when we first started covering BMY.

We had that there, but we don't have this. You know, in terms of predicting, good question, James. So I think the question always when it comes to predicting stocks, how do you do that? And I think we have to be careful that we don't assign magical capabilities to any AI, not even our AI. That kind of comes back to the data unless you really feel comfortable relying on a black box. Because if you did ask an AI, recommend stocks most likely to beat the market in the coming 12 months, what happens if we ask that? For me, I mean, I'm like, okay, great, great. You gave me some stocks.

How?

Unless I believe that Perplexity is some sort of supercomputer genius, maybe it is. I don't know how to answer that question.

Right.

But if for new constructs, At least you're going to get some kind of honest answer based on real data and a rating system that I think is pretty well proven. Speaking of ratings and predictions, let's take a look at how the live-traded index based on our rating system— how's it performing? Outperforms the S&P this year. It's up 14% versus 12%. This is an index based entirely on the very attractive stocks. So it just says, hey, we're going to hold the very attractive stocks over 5 years. 172 to 98, pretty good. That's what I'd want any system that I was going to ask that kind of question to be armed with. Well, here it is. These are— that's what it comes back with. I'm not going to say I'm recommending anything. I'm saying these are the stocks with the best ratings, most likely outperform the market in the coming 12 months. Again, this is based on data that's, I don't know, at this point, 3 or 4 months old. And it's not all the companies in our system. So it's not all the very attractive stocks. It's certainly not all the current very attractive stocks. But I think you get an answer. And I think that's where we have to move, James, as a society to, you know, a lot of people talk about how prompt engineering is like a new skill and all that. It's also understanding and having the right expectations about what an AI can do for you. If you expect it to cook your dinner and make your wife happy, you know, that's maybe too much. We have to be realistic. You expect it to predict the future. Well, that would never work, James, because everyone would just go to whichever oracle AI could predict the future for stocks, and they'd all buy those stocks, and that would immediately be arbitraged.

Right?

So we have to be careful that we don't expect AI to do unrealistic things. And I think you would want to lean toward or gravitate toward an AI That gives you more honest answers, doesn't pretend to be able to do impossible things. And so part of it's like, what kind of questions should you ask? What kind of questions should you expect an AI to answer? I think that's a big part of it. All right, we are running over time here. So Kyle, I'm gonna, I'm gonna stop sharing and let you take us home here.

That works.

Let's see.

Just wanted to touch on a few more items. We just looked at the performance there of the very attractive and Core Earnings Leaders Index. So I'll jump into the footnote season update. You know, this is the work that backs all of our data. And we spent in 2Q25, 2,800 filings that were parsed and updated models across the whole system. You know, we estimate that being worth $2.5 million in value that is provided for all new ConstructX members. This is the work that, you know, cleans, gets the data, Finds the adjustments, builds the models, and then gets the ratings to you guys through our website and email alerts and everything. So I always like to point out, you know, just how much work goes into all that data you just saw, you know, David on the AI agent there go through. It's not, you know, made up. It's not fake. It is taken directly from the source, adjusted to get the proper accurate earnings measures, free cash flow measures, ROIC. invested capital and everything else that goes into our model.

Yeah, just one second on this. I want to double tap on this because I think it's important to emphasize that one of the reasons why you don't hear much about people talking about data is because it's not sexy, right? GPUs that can power rockets to the moon, right? That's what dominates headlines. Billions of dollars spent on data centers, nuclear reactors to provide the electricity, Right? If it bleeds, it leads. Guess what? Data doesn't bleed. It's not that exciting, but it is important. It's boring. Like, some of the most important things in life are the most boring and monotonous things, but they're necessary. And so I just kind of want to make you guys understand, like, that what we're doing is unique. You don't hear about it because people aren't doing it, because it's not exciting. And I think there's a big moat around the new constructs business in that sense. Like, we've been out there doing the nerdy hard work diligence that's needed to drive real intelligence for years. And it perfected it despite all of the distraction to do otherwise. And I think that should hopefully give you some additional confidence that what we're doing after the papers, after the real market performance, all that's great. We're proving it's worked, proving it works. But you're also talking to people who've been willing to be in the trenches doing hard work for the right reasons for a long time. And I hope that means something.

I know it means something to us. That's why we keep doing it. We share all this research via our website, send out email updates in our online community. If you're not a member of the community, there'll be links in the replay as well. And we can go to newconstructs.com/society, also sign up there as well. Got members, you know, posting, sharing trade ideas on the community, talking about Western Union acquiring IMXI, some different trade ideas, some options discussions. So we'd love to see you guys out there just to get more minds on topics in the market, topics at New Constructs Research, things of that nature. And in addition to the community stuff, we also have upcoming trainings that we want to make you guys aware of. The first one is for our Pro and higher members. That'll be next week, September 16th. That is our Ideas Dinner. David will join the Pro members, institutional members, and talk long ideas, short ideas, anything that comes up in those meetings. And then we also have an upcoming training, Want Proven Stocks? The Super Stock Bundle is back with fresh picks. That'll be next September 17th. We'll walk through the exact methodology we use to find the super stocks and the secret criteria that separates true value from market noise.

Real quick here to the question, recommended stocks, that's still something that I think you need to rely on human experts to do. And whatever the AI is doing, it's going to be based on an expert, right? Just like I mean, the AI is going to be regurgitating what someone has taught it, whether it's logic on how to interpret information to derive an opinion about a stock, or whether it's just the actual picks themselves. Either way, there's some logic there, and you want to be able to understand the logic. You want to understand the integrity in the logic, understand the data driving the logic. And that's what we're putting together in New Constructs. And that's what the Super Stock Bundle is. This is based on experts in data, experts in stock picking finding even better ways to pick stocks. And by the way, there is no index, there is no ETF, there is no model portfolio that contains Super Stocks. This is a special private offering that we do on a— every several months, a couple times a year. And like, what can I say? These picks have killed it. They work really well, and it's based on criteria that we don't make readily available to anybody but institutional clients. So it's These are special.

Yeah, and the first registration email went out on that this morning. So be sure to check your inbox. We'll send out more before the event. So you've got a moment to sign up. And with that, I think we've answered a bunch of questions. I don't see any more at the moment. So we want to thank you guys for being here. Like I said, you can go to newconstructs.com/society to join our online community. And you can always reach us at support@newconstructs.com.

Real fast, I'll answer this question on tariffs. Because we get it a lot. So, in what area of your research do you see the effects of tariffs on your companies? Does it affect any of the ratings? How's it handled in a GAAP adjustment reconciliation? Look, tariffs, I think, are more about the future because we don't know yet. I mean, some companies are reporting it, but it depends on the business. And really, for that, I think what you want is sensitivity analysis on valuation. And by that, I mean, hey, can I actually forecast What the impact potentially will be on my stock that I'm interested in here. If I assume that tariffs are going to hurt revenues and margins or just margins, and our reverse discounted cash flow modeling system is designed exactly for that. You guys have seen that, especially any of you Pro members who've been to some of our idea dinners. We have enormous amounts of reverse DCF case studies. Kyle, do you mind popping a couple of those, looking a couple of those up and put them into the, into the Q&A or the chat? We have reverse DCF case studies where we go through and we put in all kinds of different scenarios for the future. And I think that, to me, that's the ultimate exercise as an investor, is to say, okay, if I think the stock is cheap, let me go in and put in the worst margins they've ever had and growth that's below consensus. And what does that imply? What is the stock worth if I see that? And if it's still higher than what the current stock price is trading today, then man, I've got some awesome margin of safety. But it's not just some amorphous margin of safety based on some shortcut calculation. Is it a margin of safety based on specific calculations in a specific scenario and the most robust DCF model in the business based on the best data in the business? That's how I think people should do business, and that's how you should do research. That's, that's how I'd answer that question with respect to tariffs.

Here we go.

Reverse DCF case studies. Check these out, everybody. These are awesome. That's a sample of one. We've got many others that you can see.

Yes, we've got a section of the blog on the live trainings, and then I also sent you the link of the one in the community as well.

Oh yeah.

So if you're a member of the community, this is probably the best space because this is a whole category in the community that does nothing but reverse DCFs, at least standalone.

You do a lot of them during the earnings watches.

So definitely if you want to see earnings watches, and for sure I do them during the long ideas dinners. That's for pro members and above.

Great.

Thanks, everybody. We're out.