Feng Ji is the founder and CEO of Baiont, a top-performing quant fund in China that uses artificial intelligence to develop trading strategies.
He argues quant trading is fundamentally a pc science task and predicts that quant fund managers failing to embrace AI is not going to last one other three years.
In this conversation with the Financial Times’ Asia Technology Correspondent Zijing Wu, Ji talks about how his team of young computer scientists with no finance background is disrupting the quant trading sector in China and has ambition to go global. He says quant trading is attracting the very best AI talents and offers fertile ground for start-ups, resembling DeepSeek.
Zijing Wu: Quant trading remains to be relatively recent in China compared with the US and Europe. Can you describe the present landscape in China?
Feng Ji: The first wave of quant trading here began with some very talented Chinese traders getting back from Wall Street. Around 2013, regulations modified to permit quant trading, and more hedging tools were introduced within the Chinese market, which created a fertile ground for this primary generation of Chinese quant traders. They did thoroughly and remain leaders of the most important funds today.
We are the second generation and really different. We come from “out of the circle” with zero finance background. We imagine that quantitative trading is similar as many other tasks of knowledge mining and evaluation. There is nothing special about it. We regard it as a pure AI task, so our team consists of only computer scientists and engineers.
ZW: How do you apply AI in quant trading, and what’s the difference between what you do versus the old-school quant trading?
FJ: AI technology has made significant progress previously 10 years, especially in time series data modelling. Whether it’s language or multimedia AI models, fundamentally it’s all about modelling time series data. For example, the core task of ChatGPT is to predict the subsequent word. It’s essentially the identical with quant trading. Instead of predicting the subsequent word we predict the rise and fall of costs in the subsequent time interval.
A standard quant fund would divide up its team in to several functions specializing in different stages of the pipeline, mainly factor finding, signal generation, modelling and strategising. These functions are independent and somewhat isolated from one another.
We see all these stages essentially as the identical machine-learning task and approach it holistically with the identical foundation model. This has a far-reaching impact on operations. It’s like before ChatGPT, language processing corporations also had team divisions which focused on word separation, tagging, evaluation etc. Now ChatGPT can do all of them at the identical time with the identical model.
AI Exchange
This spin-off from our popular Tech Exchange series of dialogues examines the advantages, risks and ethics of using artificial intelligence, by talking to those on the centre of its development.
ZW: Why is your holistic approach higher than the standard division of labour?
FJ: First of all, you’ll be able to predict and plan the upgrades of a system based on machine learning. Like when ChatGPT launched its first generation model, you principally have an idea what the second generation could be like and the way long it might take to get there. The ability to repeatedly upgrade in a scientific way is essential to a quant fund manager.
The second advantage is cost effectiveness. Instead of hiring 50 people to seek out aspects, we use 100 GPUs and one one that writes the algorithm for factor finding. The result’s even higher and far faster. The same applies to all stages.
ZW: How big is your team and the way much are your assets under management?
FJ: We currently manage near Rmb7bn ($970mn) and our team has only about 30 people. Two-thirds are doing research while the remainder concentrate on operational work. Our research is especially about improving the algorithm and our own foundation model.
ZW: Does the industry see you as disruptive?
FJ: When we first began doing it, about 4 years ago, many individuals thought it was not possible. How can a bunch of computer scientists understand business and the markets? The truth is — we don’t and we don’t have to. Actually none of us did any trading before this. We see this as a pure machine learning task and one which’s totally doable.
Now only a few people doubt us any more. Instead everyone seems to be asking us frantically how they’ll higher use AI.
So my prediction is that in three years, quant managers who don’t complete their AI transformation will likely be eliminated by the market. Because the space is getting increasingly competitive, and machine learning will turn out to be an important tool. There isn’t any reason why one shouldn’t adopt it.
ZW: Do you construct your personal model from scratch and might you give us an idea of how it really works in trading?
FJ: Yes we built all of it by ourselves. Because market data and behavior may be very different from, for instance, language data. What we cope with is lots more complex and we’d like to construct specialised models for it.
Typically we concentrate on short term trading, from minutes to hours. This is what AI is best at. It’s like predicting the weather. If you’ve got to predict the weather in a month it might not be so accurate, but in case you predict in five minutes, the accuracy may be very high, because you’ll be able to catch many signals. Short-term signals are relatively predictable and we have now analysed enough data to make quality predictions.
We will comprehensively evaluate the prediction of various signals from minutes to hours in real time. Then make a comprehensive rating of those predictions, and based on such scores we construct a dynamic combination of trades.
ZW: Does it mean you don’t care concerning the fundamentals in any respect?
FJ: Basically yes. Fundamental aspects and alternative data aspects change little or no throughout the day. We mainly depend on trading data. The core of short-term price fluctuation is driven by trading data.
ZW: Why did you and your team, coming from a machine learning background, determine to get into quant trading as an alternative of the more popular AI start-ups focused on large language models for instance?
FJ: After graduating with my PhD in machine learning, I spent a couple of 12 months taking a look at various directions during which machine learning and AI can have a very disruptive impact, somewhat than an easy upgrade of the present tools.
The second factor I considered on the time was whether it may bring a superb money flow. I realised on the time many of the AI unicorns don’t earn money. They could also be doing precious things however it’s difficult to sustain. Also for lots of them, their success depends significantly on capability of sales, not technology because there’s limited differentiation of their core tech. I felt like being a brilliant nerd, I’m not curious about anything that’s heavily sales driven.
Then I discovered quant trading, which ticks all boxes. It’s an industry we will redo throughout with AI. It’s not only a standard linear model, but with the potential to create a neural network or a random forest. It’s a challenge I’m enthusiastic about. And it’s disruptive. It’s like designing a brand new electric vehicle factory, totally disrupting the old automotive manufacturing.
The other benefit of quant trading is it’s easily verifiable. You discover straight away in case you are on the best path or not by doing greater than a thousand trades in sooner or later.
It’s also almost purely technology driven. Most quant corporations are led by individuals with a tech background. Because you’ll be able to’t manage a team of nerds and geniuses in case you don’t understand the technology yourself.
ZW: What sort of nerds and geniuses are we talking about here?
FJ: Our team, myself included, got here from a pc science competition background. Out of our 30 people we have now 13 gold medallists. Our team’s gold medal “density” might be higher than any tech giant on the market. Quant trading is an industry where you see the very best proportion of geniuses. It’s the identical within the US. I imagine the highest machine learning talents are 80 per cent in Wall Street and 20 per cent in Silicon Valley.
ZW: Is this why DeepSeek got here out of High-Flyer, one in all China’s biggest quant funds?
FJ: Indeed. I used to be not surprised in any respect about that. (The) key contributions to LLM (that) DeepSeek made was to scale back engineering costs and improve communication efficiency between GPUs. This comes naturally to quant traders, because how we quantify time is in nanoseconds to microseconds, while traditional web corporations have a timescale of seconds, or milliseconds at best.
For example, an enormous tech platform with a billion users online at the identical time desires to ensure that there’s no lag and human response is between 50 to 150 milliseconds. It’s high-quality if you’ve got a ten millisecond delay. But in quants trading one millisecond is endlessly.
Quants trading can also be where you’ve got very healthy money flows to draw the highest talents. Ten years ago it attracted the neatest people from maths and physics, because they’ll transfer their data analytical skills to finance. But today it’s steadily taken over by computer scientists. Because we don’t even have to transfer skills — machine learning is actually the identical in designing the very best tool to analyse data. It doesn’t matter whether it’s data from finance or another area.
Making numerous money also means the team has the luxurious to branch out to do things they’re more curious about pursuing. I call this a technology spillover.
When you’ve got a considerable amount of geniuses and ample resources, they can spin off some unrelated technologies based on similar core skills.
It’s happened over and over in history. For example, (hedge fund) DE Shaw’s founder created a big scientific research centre to make use of self-developed super computers for chemistry. It has nothing to do with quants trading but applying similar core skills.
ZW: Just like DeepSeek, your team is all from a Chinese education background. How do you compare the young talents in China and the US?
FJ: There’s little or no gap lately. We are competing on principally the identical level. And China has a bigger pool of such talents due to our education system with a stronger concentrate on science and technology. We are particularly strong in engineering capability and algorithm innovation.
In the past decade, smart young people from in every single place on this planet can freely communicate with, learn from and work along with one another on open source AI platforms. This has given our generation of Chinese coders an ideal opportunity to meet up with the world’s leading technology on this area.
The other thing about this young generation of talents in China is, unlike their parents, they grew up in mostly middle class families where they didn’t should do things they didn’t like with the intention to make a living.
Most of our team are of their twenties. I’m 37 and the oldest by far. Their primary priority is to rejoice. So as an alternative of going to big tech corporations where they may probably should cope with politics a technique or one other, they’d much somewhat come to a smaller research-oriented team like ours, where they work with similarly smart colleagues and a manager who speaks their language.
Having grown up in well-off environment also means this generation of Chinese young talents are more idealistic than their parents. You see more going into research as an alternative of finance for quick money. We actually need to do something to alter the world.
ZW: What’s a each day work schedule like on your team?
FJ: It’s principally like a research institute. No dress code — shorts and slippers are essentially the most common. We arrive before markets open, start programming and discuss our work together, and review the performance before the market closes. Run a number of more experiments, read and discuss some papers, and go home. The difference between us and a research institute is we have now higher resources. We construct our own computing power. The more compute you’ve got, the faster you get the outcomes and the more efficient you’re. It’s very essential.
ZW: What’s the final word goal for you and your team?
FJ: In the midterm we would like to construct a world leading AI-native quants fund from China. We mainly trade within the Chinese markets now and we want to expand into the important thing overseas markets. When people discuss quants funds all of them think concerning the Wall Street top firms, few knew concerning the Chinese funds. While the primary generation of Chinese quants funds used the methodology learned from Wall Street, we will differentiate higher by being AI-native early enough. We have a likelihood to compete with the worldwide leaders.
In the long term we would love to construct a computing company. There are many potential areas we’re enthusiastic about, where we could spill over our technology. LLM isn’t necessarily the very best use of AI.
