00:06:41 · Ying Hua
It's mostly just as a dirty work too. I just don't think they are interested in doing that dirty work. But the second thing is I think from first principle, one needs to understand limitation of language model, right? Language model, at least the bark architecture today, you know, it's next token prediction. It means, language is really a coding too. It's a coding language. It's like we put information in our language and based on that information you try to predict what is next piece of information that can likely come through. It is by definition backward looking, it's by definition pattern matching, it's by definition not true understanding, meaning it's just kind of filling the pattern. The analogy I will do is think about quant investing. All right, quant investing it's kind of like similar to that, it's just like information coded in numbers. You're trying to predict what's the next number that comes up. In this case, it's information encoded in language, you figure out what's next word that comes up. Right, so when you have that kind of market model architecture, by definition you're going to run into a few things. Number one is it's not true understanding. It's not true reasoning either. It's more just how do you give it enough information in your language in that context window, so it can do that prediction more accurately. That's all it is, right? That's number one. Number two is can never figure out things that hasn't happened before. It's a pattern matching. Quants backward looking, right? It's always backward looking. When regime change happen, when pattern changes, it cannot identify it. Same thing with LLM. If information you give up till that point, if something changes, it's not going to be able to predict that. So those are limitations. Think based on that, I think there's still a lot of holes to fill from each vertical. And I think that's kind of where we come in.
00:08:36 · Khe Hy
[00:08:36] S2: Can I jump in here, on that, on the pattern matching question, Yin? I completely agree with you that there's tons of backwards looking pattern matching. Isn't there an argument to be made, though, that the sheer volume of patterns that are able to be extracted is, in a way, its own set of knowledge, is, in a way, its own set of understanding. How would you think about that?
00:09:00 · Ying Hua
[00:09:00] S0: Yes. I think that's correct. But I think it depends on the input of those pattern metrics. So let me give you an example. How do you think language model train based on, earnings result? Remember like early in the days like even when Bloomberg was training his own model like we were talking about like all the financial trading, right? The idea of training you just shove it a lot of documents. 10 ks, 10 ks, transcript, blah blah blah. Okay. I think, if you guys ever read the earnings release, okay, management always say we had a great quarter. We had a great quarter. It doesn't matter how good actually good or bad a quarter is. We had a great quarter. When you feed that transcript into LLM, LLM will tell you they had a great quarter. Right? So my point is, number one is, I agree, if you can derive the right pattern, then the pattern itself is enough. But I think there is a limitation on how these models are trained because they're fed in raw data. And the other question is like, I covered the financials so I'm much more acute to this, but every time I read like a Wall Street Journal read about how does AIG do, It's really never the right read because they're really looking at the wrong metrics. There's something that they didn't know. It's because everyone's kind of like someone who doesn't understand industry. When you read the surface level reports they will have one interpretation. But when you have more domain knowledge, like depth of domain knowledge, you're going to reach a different conclusion. So my view is I don't think these models were ever trained in the proper way of ingesting knowledge other than it's just the raw data. So that's kind of my view as well. But I agree with you. If we can figure out a way to train it in the right way, causal way, right? Know snowflake is up 35% today because blah blah blah blah versus expectation. Then if you train model, with this, I think they can actually mimic the right type of reasoning. Meaning can... Next time I see something similar, it'd be like, okay, I think it's gonna react this way. But models are not trained that way. Model are trained period on basic, data.
00:11:12 · Brett Caughran
[00:11:12] S1: Can you talk to us about just the data challenge? I mean, part of the reason that I didn't really conclude that LLMs were institutional grade in 2025 was so much of the sort of data that informed outputs was from the open web. Training corpus is sort of fundamentally sort of compressed web scrape. And I think the open web may be great for coding sort of tool, but in terms of like institutional grade finance workflows and data, not so much. How have you sort of tried to crack that problem? And how have you seen the ecosystem evolve to sort of handle that institutional grade data problem?
00:11:59 · Ying Hua
[00:11:59] S0: Yeah, I think data is an easier problem to solve now, right? Because it's interesting. Like, remember we started the company in 2023, and within a few months, we just saw a bunch of companies starting. And then my co founder and I were like, it's interesting a lot of companies that our founders don't really have investment background why would they start a company like that? Then they hit us like oh a CC document. There is a large corpus of free data out there that you can just a put feed into LLM and then it spits out some answer. It's a very natural reaction to go. So number one is I think source of data. You want to go to sources, go to SEC, go to company websites, go on to go to transcripts And that's number one. And number two is we have taken a slightly different approach. Like we don't buy a lot of data. We actually process all the data ourselves. We do all the pre processing ourselves. And then, to give you an example, I think every competitor out there more or less buys transcript data. We actually do our own live transcription and we got a lot of questions like why would you do that? By the way it's about not even that it's not that easy to build. There's a lot of busy work right? A lot of not so glorious work that we needed to do in that. But the reason why we do it is number one is we get to control how data is processed. Number two we get to control when that data comes in. Like So for example, our live transcript, it's live, it's sub one second delay. As soon as the earnings hit, it can come through. Number three, it allows us to gather raw data. We're doing that raw data. So if you talk to most of quant that buys data, they're like, Oh, how is sentiment on this topic from management? They're using clean transcript. How can you tell sentiment on a topic if you don't include all the filler words? The raw transcript, the ums, uhs, the spaces in between sentences. You can't really do that. So we wanted to find a way to have the most raw data that could be processed as quickly as possible and control the full pre processing step. So that's number one, I think it's something that we did. Number two is very early on we thought that the qualitative data sets are more or less obvious to people but there's a lot of quantitative data sets that people are not including in there. And that has two folds.
00:14:18 · Ying Hua
Number one is a lot of questions should not be pulled from qualitative data. So very simple, this is like a very common trial question, Poll apples trying to sell over the last twenty quarters. Like before, a lot of people were doing this as a sample question. That question should not be pulled from 20 different transcripts. That should be a structured data set. That should be a Bloomberg poll. I have joked about that's a terrible question because a lot of people can pull that data a lot faster using Bloomberg code than you typing out that whole sentence about like you know what to pull, that's number one. So I think some data should be structured to begin with and I think that goes back to some of data quality stuff, doing models and stuff like that that should be structured. Second thing is, institutional grade is also we try to find data that people would have wanted to get but has a lot of hurdle in getting. So what do I mean by that? Some of the alternative data. Not alternative data like credit card reports or whatever stuff everyone every hedge fund uses. There are a lot of data out there that only works for one or two companies. No one really is scraping that data because it's not a viable business to scrape that data. Who are you going to sell to? It's only for work for two companies. Who's not going to buy it? It's not enough data to back test. So we very early on have helped clients say hey you can actually use us a tool to script those data we can help you script those data and everything's integrated. So we fill in the the data layer by one going to the sources we try to do a lot of pre processing ourselves two we're filling a lot of gap with a structured data set to one improve the quality but also give more scope to the availability of our data sets.