Feng Ji is the creator and CEO of Baiont, a top-performing quant fund in China that uses artificial intelligence to create trading approaches.
He argues quant trading is basically a computer science job and anticipates that quant fund managers stopping working to welcome AI will certainly not last one more 3 years.
In this conversation with the Financial Times’ Asia Modern Technology Correspondent Zijing Wu, Ji talks about exactly how his group of young computer scientists with no financing background is interrupting the quant trading industry in China and has aspiration to go international. He claims quant trading is attracting the best AI abilities and uses productive ground for startups, such as DeepSeek.
Zijing Wu: Quant trading is still relatively brand-new in China compared with the United States and Europe. Can you define the current landscape in China?
Feng Ji: The first wave of quant trading here started with some really gifted Chinese investors returning from Wall Street. Around 2013, regulations changed to permit quant trading, and even more hedging devices were introduced in the Chinese market, which developed a productive ground for this very first generation of Chinese quant investors. They did effectively and stay leaders of the greatest funds today.
We are the 2nd generation and really various. We originate from “out of the circle” with zero money background. We believe that quantitative trading coincides as lots of various other tasks of information mining and evaluation. There is second best concerning it. We regard it as a pure AI task, so our team is composed of only computer scientists and designers.
ZW: Exactly how do you apply AI in quant trading, and what’s the difference in between what you do versus the old-school quant trading?
FJ: AI modern technology has actually made significant development in the previous 10 years, specifically in time series data modelling. Whether it’s language or multimedia AI designs, essentially it’s everything about modelling time collection information. For instance, the core task of ChatGPT is to anticipate the following word. It’s basically the exact same with quant trading. Instead of anticipating the next word we anticipate the rise and fall of rates in the following time interval.
A conventional quant fund would separate up its group in to numerous features focusing on various phases of the pipe, primarily factor finding, signal generation, modelling and strategising. These features are independent and somewhat isolated from each other.
We see all these stages basically as the exact same machine-learning job and strategy it holistically with the same foundation design. This has a far-reaching impact on procedures. It resembles prior to ChatGPT, language processing business likewise had group divisions which concentrated on word separation, tagging, analysis etc. Currently ChatGPT can do all of them at the same time with the very same version.
AI Exchange
This spin-off from our popular Technology Exchange series of dialogues checks out the advantages, threats and principles of utilizing artificial intelligence, by talking to those at the centre of its development.
See the rest of the meetings below
ZW: Why is your holistic technique far better than the typical division of work?
FJ: Firstly, you can predict and prepare the upgrades of a system based on machine learning. Like when ChatGPT launched its first generation design, you generally have a concept what the 2nd generation would resemble and for how long it would take to get there. The capacity to constantly update in an organized way is crucial to a quant fund supervisor.
The second benefit is cost efficiency. Rather than hiring 50 individuals to locate elements, we utilize 100 GPUs and someone that composes the formula for factor finding. The result is even much better and much quicker. The exact same puts on all stages.
ZW: Just how big is your team and how much are your properties under monitoring?
FJ: We currently take care of near to Rmb 7 bn ($ 970 mn) and our group has only about 30 individuals. Two-thirds are studying while the remainder focus on functional work. Our research is mostly regarding enhancing the formula and our very own structure version.
ZW: Does the sector see you as turbulent?
FJ: When we initially began doing it, regarding 4 years back, many individuals assumed it was difficult. Exactly how can a number of computer researchers understand business and the markets? The reality is– we don’t and we don’t need to. In fact none of us did any kind of trading before this. We see this as a pure device learning job and one that’s absolutely practical.
Now very few individuals doubt us anymore. Rather everyone is asking us frantically just how they can much better utilize AI.
So my prediction is that in 3 years, quant managers who do not finish their AI makeover will be eliminated by the market. Due to the fact that the space is getting an increasing number of competitive, and artificial intelligence will certainly come to be an essential device. There is no reason that one should not adopt it.
ZW: Do you build your own model from scratch and can you offer us a concept of how it works in trading?
FJ: Yes we built everything by ourselves. Since market information and behaviour is extremely various from, for example, language data. What we take care of is a lot much more complicated and we require to develop specialised designs for it.
Usually we concentrate on short-term trading, from minutes to hours. This is what AI is best at. It’s like forecasting the climate. If you need to anticipate the climate in a month it would certainly not be so accurate, yet if you predict in five mins, the accuracy is very high, because you can capture many signals. Short-term signals are relatively predictable and we have actually analysed enough data to make top quality forecasts.
We will thoroughly evaluate the forecast of different signals from mins to hours in genuine time. Then make a detailed score of these forecasts, and based upon such scores we build a vibrant mix of professions.
ZW: Does it suggest you uncommitted about the basics at all?
FJ: Primarily yes. Basic aspects and different data aspects change very little throughout the day. We primarily rely upon trading data. The core of temporary cost variation is driven by trading information.
ZW: Why did you and your group, originating from an equipment learning background, decide to enter into quant trading as opposed to the a lot more popular AI start-ups concentrated on large language designs as an example?
FJ: After finishing with my PhD in machine learning, I spent concerning a year taking a look at different instructions in which artificial intelligence and AI can have an absolutely disruptive influence, rather than a simple upgrade of the existing devices.
The second element I took into consideration at the time was whether it can bring an excellent cash flow. I understood at the time a lot of the AI unicorns do not earn money. They might be doing important points yet it’s hard to endure. Also for much of them, their success depends substantially on ability of sales, not technology because there’s restricted differentiation in their core tech. I felt like being a super nerd, I’m not interested in anything that’s heavily sales driven.
After that I discovered quant trading, which ticks all boxes. It’s a sector we can remodel all over with AI. It’s not simply a conventional direct version, yet with the potential to develop a semantic network or a random forest. It’s a challenge I’m fired up regarding. And it’s turbulent. It’s like making a brand-new electrical car manufacturing facility, absolutely disrupting the old cars and truck production.
The other advantage concerning quant trading is it’s easily verifiable. You find out immediately if you get on the ideal path or not by doing greater than a thousand trades in eventually.
It’s additionally almost totally modern technology driven. Many quant firms are led by people with a tech background. Because you can’t handle a group of nerds and brilliants if you do not understand the innovation on your own.
ZW: What sort of geeks and brilliants are we talking about right here?
FJ: Our team, myself consisted of, came from a computer technology competition background. Out of our 30 individuals we have 13 gold medallists. Our team’s gold medal “density” is most likely greater than any type of technology titan out there. Quant trading is a market where you see the greatest proportion of wizards. It’s the same in the US. I think the top machine learning abilities are 80 percent in Wall surface Street and 20 percent in Silicon Valley.
ZW: Is this why DeepSeek came out of High-Flyer, among China’s greatest quant funds?
FJ: Indeed. I was not shocked at all about that. [The] key contributions to LLM [that] DeepSeek made was to decrease engineering expenses and improve communication performance in between GPUs. This comes normally to quant traders, because just how we measure time remains in split seconds to microseconds, while conventional web companies have a timescale of seconds, or nanoseconds at ideal.
For instance, a huge technology platform with a billion individuals online at the very same time wants to ensure there’s no lag and human response is in between 50 to 150 milliseconds. It’s great if you have a 10 millisecond hold-up. Yet in quants trading one nanosecond is forever.
Quants trading is likewise where you have very healthy and balanced cash flows to bring in the leading skills. 10 years ago it attracted the smartest people from mathematics and physics, since they can move their information analytical abilities to fund. Yet today it’s slowly taken over by computer system researchers. Due to the fact that we do not also need to transfer abilities– machine learning is essentially the same in making the best tool to evaluate information. It does not matter whether it’s information from finance or any kind of various other area.
Making a great deal of money likewise suggests the group has the deluxe to branch off to do things they are much more curious about going after. I call this an innovation spillover.
When you have a huge amount of geniuses and adequate sources, they are able to dilate some unassociated modern technologies based on comparable core abilities.
It’s occurred sometimes in background. For example, [hedge fund] DE Shaw’s creator developed a huge scientific study centre to make use of self-developed super computers for chemistry. It has nothing to do with quants trading yet using similar core skills.
ZW: Much like DeepSeek, your group is all from a Chinese education and learning background. Just how do you contrast the young talents in China and the US?
FJ: There’s really little gap these days. We are competing on generally the very same level. And China has a bigger swimming pool of such talents many thanks to our education and learning system with a stronger focus on scientific research and innovation. We are especially strong in design capacity and formula technology.
In the previous decade, smart youngsters from all over in the world can easily connect with, learn from and work together with each other on open resource AI platforms. This has given our generation of Chinese coders a wonderful opportunity to catch up with the globe’s leading technology around.
The various other feature of this young generation of skills in China is, unlike their moms and dads, they matured in mostly middle course households where they didn’t have to do things they didn’t like in order to make a living.
A lot of our team remain in their twenties. I’m 37 and the oldest without a doubt. Their top concern is to enjoy. So as opposed to mosting likely to huge technology firms where they will probably have to deal with politics somehow, they would a lot rather involve a smaller sized research-oriented team like ours, where they work with in a similar way wise associates and a supervisor who speaks their language.
Having actually matured in rich atmosphere also suggests this generation of Chinese young talents are more radical than their parents. You see even more going into study as opposed to financing for fast cash. We really want to do something to alter the world.
ZW: What’s an everyday job routine like for your group?
FJ: It’s generally like a research study institute. No dress code– shorts and sandals are the most common. We get here prior to markets open, begin shows and discuss our collaborate, and examine the performance prior to the market closes. Run a couple of even more experiments, check out and discuss some papers, and go home. The distinction in between us and a research institute is we have much better resources. We build our own computing power. The more compute you have, the faster you get the outcomes and the extra reliable you are. It’s extremely crucial.
ZW: What’s the ultimate goal for you and your team?
FJ: In the midterm we intend to construct a world leading AI-native quants fund from China. We generally trade in the Chinese markets currently and we are looking to expand into the crucial abroad markets. When individuals discuss quants funds they all think of the Wall surface Street leading firms, few understood about the Chinese funds. While the very first generation of Chinese quants funds used the technique gained from Wall Street, we can separate better by being AI-native early enough. We have an opportunity to compete with the international leaders.
In the future we wish to develop a computing business. There are lots of potential locations we are thrilled around, where we might overflow our technology. LLM isn’t necessarily the best use AI.
This transcript has actually been edited for brevity and clarity.
