9.21.26 Treasury Price Discovery; Balerion’s Naren Krishna on AI Underwriting; Economic Calendar Ahead

Robbie ChrismanWelcome to the Chrisman Commentary, daily mortgage news podcast. I'm your host, Robbie Chrisman. Topics on today's episode include further takeaways from the Fed's tightening regime, why rate volatility weighs on mortgages in general, and my interview with Balerion's Naren Krishna on the company's AI-driven approach to mortgage lending, market needs, partnerships, and plans for growth. Here, take a listen, do a little preview. Robbie ChrismanI told you that I spoke at an event today and they they asked me what parts of the mortgage process are most ripe for disruption. And I thought my answer was obviously underwriting, but I thought there's multiple companies trying to disrupt every single stage of the process through application, through post-closing. And so I want to ask you why underwriting. Why did you view that as the opportunity that you wanted to sink your teeth into? Naren KrishnaYeah. Sort of starts with what we think is the meat and potatoes, what AI is really good at. And I think what we think it's really good at, and what our team's done in terms of you know proving that is you take a process that is partially an art form, but has some guiding text to go with it. So you have these kind of guidelines that are reasonably clearly delineated, but there's definitely an art form of there's ambiguity. You can solve some of that ambiguity by looking at historical data. There's a process element of every individual underwriter does things slightly differently. And if you think about what underwriting is, it fits into a lot of elements of what different types of lenders do. I would argue that a correspondent who's buying and onboarding a bunch of loans does an underwrite. They're just comparing their underwrite against the existing underwrite of the closed loan file. Uh the retail lenders are doing an underwrite. To a certain degree, even folks buying loans for servicing are sort of determining the risk profile of that loan. And you could argue that's underwriting in a form as well. So it's this kind of unique text-heavy thing that happens a whole bunch of times uh by different players across the lifecycle of a loan. I also think that it's the core blocker for what enables you to get a loan as a retail lender, for instance, you know, and and sell it. You know, you can get a bunch of inbound, but it doesn't matter how much inbound you have if you're underwriters and only being able to underwrite two and a half loans a day. And so that to me was was kind of why it was an interesting space of you could package it, you know, the same product from an AI perspective up to so many different kinds of customers, but also solve this core problem.SpeakerThanks to this week's podcast sponsor, SpringEQ, the home equity expert. See why SpringEQ is the clear choice in home equity, helping over 150,000 homeowners access almost $15 billion in equity by visiting mortgage.springeq.com slash equity. The Federal Open Market Committee, chaired by Chair Kevin Walsh, voted to raise the benchmark federal funds rate by a quarter percentage point to a target range of 3.75 to 4% last week, marking the first rate hike since 2023, as inflation remained stubbornly above target. In the bond market, the benchmark ten year treasury climbed to its highest level in nearly two decades, as resurfacing inflation peers and rising oil prices weighed on investor sentiment. The Treasury market is undergoing some renewed price discovery after last week's Fed rate hike with two competing forces. Renewed confidence in the Fed's inflation fighting credibility that could lower inflation expectations and support lower nominal yields, and persistent fiscal and treasury issuance concerns that demand higher real yields and term premium further out the curve. The unanimously hawkish statement, especially the dot plot showing sixteen of eighteen officials expecting at least one more hike this year, has pushed markets to price a more aggressive tightening path. In the short term, the two-year yield remains anchored by near-term policy expectations, making the yield curve more likely to bull flatten when Fed credibility dominates and bears steepen when fiscal concerns intensify. Two-year treasury yields rose ten basis points over the course of last week to 4.74%, and ten-year treasuries rose above five percent multiple times. Additional Fed hikes will depend heavily on whether policymakers are responding primarily to sticky core inflation or renewed energy-driven inflation, particularly if oil remains above $100 a barrel, and September inflation data proves less benign. For today's interview, I wanted to welcome to the show Balerions and Naren Krishna to talk about the company's AI-driven approach to mortgage lending, market needs, partnerships, and plans for growth. He's CEO and co-founder of Balerion AI, building intelligence before underwriting for mortgage loan manufacturing. Prior to co-founding Balerion, he led Time Series data analytics for financial services at Snowflake and was an ML engineer at two early stage startups. His background spans AI research, machine learning infrastructure, and human interaction with a focus on building systems that augment human expertise and complex high-stakes workflows. Robbie ChrismanAI for the last couple years in the mortgage industry was a lot of here's what it can do. And about a year ago, we started to see here's what AI really is doing versus the promise or the speculation of what it can do. Now we're into agentics. And so when we think about agentics, especially as it pertains to the underwriting space in mortgage banking, just where are we? What is the latest and greatest terms capabilities? Naren KrishnaI guarantee you my answer is going to be different six months to a year from now than it is going to be now, just because of how quickly the space is evolving. But I kind of work backwards from where the frontier models are at. The frontier models being you know predominantly for reasoning tasks, not just simple classification extraction, but but true reasoning would really be Anthropic and OpenAI. Those models are quite good at parsing massive quantities of text and being able to synthesize and summarize them. What they're not very good at is being able to do that with high accuracy and map ambiguity. So one of the things we've focused on is building kind of the mechanism to solve ambiguity in a model. The way we've done that is by looking at historical data. So I would say the models are capable of doing almost anything, but it's it's a matter of how much data are you able to feed it. And there's a boundary of if you feed it too much, then maybe it'll hallucinate because it it just sees too much context and doesn't know what to do with it. But I would say the task that it's incredibly good at summarization, computation, uh in terms of you know DTI, QMI, LTV, uh, and then being able to answer questions, question answering around, hey, there's this thing I see in the loan file. Here's this guideline. What do I do with this? Uh so it's able to resolve open-ended tasks, but doing it with very high specificity is still quite hard, you know, out of the box. Robbie ChrismanThat's a very good point you bring up. I've started to notice AI. It sounds like AI has too much data to figure out what to actually tell me in terms of there's like a new age hallucination with God, there's so many things out there that Robbie's not getting the response he wants based on his horrible prompt writing. Naren KrishnaYeah, the the technical term for this is basically the context that you provide the model. So think of it as effectively how much it can store in its brain. You could, you know, I use the human analogy of let's say you're going in for a class and you have a small amount of information you're absorbing or you know, over or over the course of an hour you're absorbing this. You'd probably be able to, at the end of that class, make some reasonably good judgment given kind of that domain. Versus if you do an eight-hour marathon session without breaks and you're expected to sort of take a quiz after the end of those eight hours. What you saw in hour one or hour two is probably not going to perform really well by hour eight. And so I think the challenge is figuring out how do you limit the context that you're providing this model so it performs better, while also maximizing that context from the standpoint of wanting to provide it as much information that's relevant as possible. Robbie ChrismanYeah, that's definitely a good tidbit rather than me saying one's computational power and chips going to be strong enough that this won't be an issue anymore. Naren KrishnaI think actually less so than the chip the chips are quite pretty strong. Um I mean, the reason the evidence I have for that is you have a bunch of companies that are working on edge computing, right? So not just cloud, but how do I get this on a mobile device, on robotics, on you know, autonomous vehicles without internet hookup and whatever. But but there is a massive supply chain issue with with TSMC and ASML and being able to fab stuff. But that's a different topic for another. Robbie ChrismanYeah, let's let's talk Balerion here, which was named after the largest and most powerful dragon in the history of Westeros, House Targaryen. Naren KrishnaOh, god, I'm not it is clearly a massive Game of Thrones fan, Rob. Robbie ChrismanYeah, massive. I actually I think I told this to you previously. I went to Dubrovnik, Croatia, which is the walled city they have in it. They have King's Landing there in Game of Thrones, and and we might have had a couple drinks. We basically took a kayak in the middle of the night out in King's Landing, thinking we were part of the show. That's about as close to uh understanding anything that goes on there as I get. Regardless, I have been known to say Balerion is the sexiest name in mortgage. I will stand by that assertion. Beyond the memorable name, what are you building? And I saw the tagline of intelligence before underwriting. What's that mean in practice? Naren KrishnaI think of underwriting as uh the most expensive asset a lender has in the sense that they're the most expensive part of the loan manufacturing process, or at least one of them. Uh TPR is another big one. But when we say intelligence before underwriting, I think underwriters are constantly burdened with problems that should never reach their desk. Uh, if they're missing files in a loan document that need to be provided by the borrower, if there are additional documentation, or if the borrower isn't qualified because the LO calculated income wrong or something like that. What we're trying to solve is poor loan quality submissions to underwriting. So it's things that really should be resolved upstream at the borrower or the loan officer level. So that way when the underwriters do get a clean file, they're able to do it while minimizing the back and forth in the process between the processor, the LO, the underwriter, the borrower. So that's the problem we're we're focused on solving. The way we do that is by presenting a really simple red, amber, green status. Red being issues that need to be resolved by the borrower, amber being, you know, issues where the AI, I think, you know, says, hey, this could be an issue, this could not be an issue, I'm actually not really sure. And I think it's really important to be able to surface that um and having a barrier threshold where the AI can say, hey, I don't really know, let's let's give this to a human. And then green being issues that that they don't have to you know look at at all, and and the AI is able to clear with high fidelity. Um so red, amber, green, really simple color-coded status. Um, and that's how we solve that. Robbie ChrismanWhat are you hearing in the market right now in terms of where teams are feeling the greatest pressure to improve? You obviously spend a lot of time listening to lenders and mortgage operators trying to figure out just what they need, how to make your product better. What are you hearing? Naren KrishnaI would say our economic buyer, the the people we sell into, are an ops persona. Um so that's a a COO or CLO, depending, depending on the shop. But for them, I think it all comes down to cost. Um, at the end of the day, cost of manufacturers obviously still quite high. But I think more interestingly, how they break down that cost is kind of an interesting exercise. For a lot of lenders, I think there's a cost of maintaining the 20, 30 vendors that they're using today. And then there's the opportunity cost of if I'm locked into a two, three-year deal with one of those vendors and I find a better technology that's out there, how am I able to get my hands on that without paying twice? And so really focusing on that problem of how can we kind of build really an orchestrator that'll allow us to build the right integrations to take us from kind of the start of a loan to sort of closed. And for us, starting with that meaty part of underwriting and then being able to shift either right or left from there is sort of the approach we've taken. But yeah, I mean, I would say the number one thing we hear a lot about is how do I take advantage of AI without blowing up my cost per loan dramatically in the short term? Robbie ChrismanCompanies are certainly believing in what you're building. You recently announced partnerships with FM Home Loans and Alameda Mortgage, both move pretty quickly. What created that urgency? Maybe more importantly for listeners, what use cases are you solving together? And I know you enact this building together approach, which can often differ from traditional vendor relationships. Maybe you can speak about that briefly too. Naren KrishnaYeah. So a couple things there. So to answer your use case questions, really simply closed loan QC and underwriting. Um, I think functionally it's a very, very similar type of AI on the back end that needs to be used for it. But the but the interface that in in terms of what you expose and how configurable it needs to be is slightly different, obviously, in the scale, of course. In terms of how we partner with companies, uh one thing I pride myself on and our team on is really creating kind of a more forward-deployed engineering motion, where it's not just, hey, let's drop this back end, let's drop this API, let's drop this UI in your lap and go figure it out what to do with it and pray and hope that your operating costs go down. It's much more so let us first, you know, spend a week or two when we start engaging with those companies to understand their business model, to understand what their bottlenecks are today, um, to do that in a quantitative way by hooking into their LOS and really observing their process. And then from there trying to tailor our product to best fit their use case. Um, I kind of think of it as there's always going to be an 80% overlap if we do our jobs and it's the right partnership of what we're building and what they're building. But that remaining 20%, that customization that comes into play uh is really key and important. So focusing on that is is kind of one aspect. The second aspect is ensuring that our system is as accurate as we say it is. So actually doing, you know, trial testing and ensuring that we pass several hundred loan files through the system and do shadow underwrites and compare our underwrite with with you know actual underwriters, looking at historical loan data, you know, examples like that. And so that's kind of how how we partner with those customers. In terms of how it moved fast or why it moved fast, I think one, there's incredible excitement in the market right now around what we're building and and the cut the capabilities of AI. I think number two, it really boils down to us having a scalable architecture, a Berkeley AI research team that you know is worked with mortgage banks before, and then you know, building that team in tandem with mortgage operators like Patrick Harkins and Devin Daly, who who are sort of on our team. Robbie ChrismanYeah. Let me dive into that a little bit because it would seem to me there's a clear demarcation or line in the sand between people going, I'm going to disrupt mortgage, and people going, we've done it this way forever. It's going to stay that way. You seem to have blended that between these Berkeley AI researchers and engineers with experienced mortgage operators. You talk about the combination, why it's so important to you, and how it differentiates Balerion. Maybe also who are you hiring as you continue to grow? I won't say do you have room for a podcast house, but it's certainly an interesting profile when it comes to expansion. Naren KrishnaOne is folks who have really been, you know, looking at reinforcement learning, post-training, foundation models, who know how to build and scale agents, uh, I think is is really important to have in-house. And so in order to recruit them, honestly, it's it's folks who have been doing this for years in the context of different other industries. Some of them come from the construction world, the insurance world, both of which are highly regulated, very tax-heavy domains. Um so there's a lot of application to mortgage that you can take from that. Mortgage, as I'm sure you know, it's his own little niche, it's his own little world where you know you're either in the clique or you're out of it. And if you're in it, you're very much in it. If you're out of it, you're very much out of it. And so needing folks who've really experienced that domain, both from you know, opening doors from a sales perspective, but more importantly, having a deep understanding of why there is so much complexity in the domain. Where not everybody who comes in is just going to be a standard W-2 borrower with, you know, a little bit of overtime income and you know, pretty clean on the right. It's those tail cases where you know you have folks who come in with seasonal income and borrowers with weird scenarios where putting them into a loan product may be difficult. Now all of a sudden, you know, you you have borrowers that are showing their income in non-traditional ways, which is why you have debt service loans and non-QM is sort of you know propping up so often. Really marrying domain experts who've seen the evolution of technology and mortgage and understand what the gaps are with the people who've solved those gaps in other industries and trying to meld it to. Robbie ChrismanSeems like a winning formula, at least from my perspective. And I mean that sincerely. So let's close with this. MBA annual is right around the corner. What conversations are you hoping to have in Chicago? And as you close out the year and look toward 2027, biggest focus for Balerion? Naren KrishnaI would say in Chicago, you're definitely going to find Patrick Harkins and me probably around the conference floor. We have a big booth at MBA, and then, you know, I'm sure I'll have various bar conversations with folks. I would say the conversations we're looking to have is just meeting more lenders, trying to understand, you know, the domain that they they see themselves in and seeing if the problems I talked about here sort of resonate with what they're seeing from an operations perspective inside of their shops. Going into next year, I think our big focuses are going to be, you know, number one, we want to actually publish what I think will be Mortgage's first lending evaluation. So like 200 odd, you know, fully labeled loan files and just really looking at, you know, where different players in the market are. How does Anthropics models do and OpenAI models do? How do you know these open source models sort of do on that on that framework and publish that? And then, you know, also focusing a lot with you know, getting in front of closed loan QC customers and seeing if we can partner with some of the bigger lenders that are working on that task. Robbie ChrismanWell, I wish you the best of luck. Either you're in the clique or you're out of the clique. You're in my clique. Hopefully I'm in your clique. Uh looking forward to seeing you at MBA Annual, and hopefully this is the first of many interviews we do. Naren KrishnaYeah, likewise. Thanks, Robbie. See you at MBA.SpeakerRate volatility has weighed on mortgages, which widened in four or five sessions last week and remained tactically cheap on spread, but buyers are waiting for treasuries to establish a more stable trading range before adding risk. The choppy rate environment also made mortgage hedging particularly difficult, while the deeply special Fannie 6.5 roll increased extension and financing costs for pipeline hedgers. Overall, last week reminded us that the Fed appears increasingly determined to regain control of inflation. Oil is adding to the pressure, growth has yet to weaken materially or meaningfully, and two-year yields now look more justified. For mortgages, realized volatility needs to settle before that cheapness can translate into cleaner performance. This week is light on the data front, with the focus being Thursday's new home sales report. Sales are expected to partially recover in August, rising 2.6% to a 623,000 pace after a sharp decline in July. Other highlights this week include treasury auctions of $69 billion two years, $70 billion five years, and $44 billion of seven years. With nothing of note on today's economic calendar, we begin the day with agency MBS prices, better by an eighth to a quarter from Friday's close, the two-year yielding 4.72, and the 10-year yielding 4.95 after closing Friday at 5.00%. Let's wrap up with a joke and some housekeeping. I told my wife I want to be cremated. So she made an appointment for Wednesday. Thanks again to SpringEQ for sponsoring this week's podcast. SpringEQ is the clear choice in home equity, helping over 150,000 homeowners access almost fifteen billion dollars in equity. To learn more, visit mortgage.springeq.com slash equity.