CEO of DeepSeek's parent, High-Flyer Quant, spoke in 2020
Lu Zhengzhe talks about a quantitative investment model fully reliant on AI. It is from there that DeepSeek was born.
As I’m typing this, Nvidia is down 13.5%. Hundreds of billions are literally at stake because of DeepSeek, the Chinese AI company. However, despite the enormous opinions about it on Twitter, very little is known about the company.
As far as I’m aware, and I’m like an authority on this sort of thing, the company has remained absolutely radio silent in the past month. Neither international mainstream media nor Chinese media, including China’s top state-run media outlets, has been able to crack it open.
So far, there are perhaps only two first-person accounts from DeepSeek, in two separate interviews given by the company’s founder. They were conducted by 于丽丽 Yu Lili, a female reporter for 暗涌 Waves, a media brand within 36Kr, published on May 24, 2023, and July 17, 2024.
Jordan Schneider’s ChinaTalk has translated each [2023] [2024] to English.
And that’s it.
I think that’s certainly not enough. So below is an interview in 2020 given by 陆政哲Lu Zhengzhe, CEO of High-Flyer Quant, the parent and funder of DeepSeek. (He is NOT the founder of DeepSeek, though.)
Lu obtained his bachelor’s degree from Zhejiang University and master’s from the London School of Economics and Political Science (LSE). In the 2020 interview with 朱昂Zhu Ang, available in Zhu’s opinion column and elsewhere, Lu shared insights into the company’s AI-driven investment strategies, the evolving quantitative investing landscape in China, and his vision for the future.
High-Flyer Quant’s investment approach is rooted in AI which underpins every stage of its operation—from data collection and analysis to portfolio construction and trade execution. The firm relies on deep neural networks and integrates market, macroeconomic, and alternative data to predict stock price changes, enabling it to generate excess returns through systematic, automated processes. This AI-centric methodology differentiates High-Flyer from traditional hedge funds in China, positioning it as a leader in China’s quantitative investment sector.
Addressing the debate on market efficiency, Lu highlighted that China’s A-share market, dominated by retail investors, is far from fully efficient due to emotional trading behavior. This presents unique opportunities for quantitative strategies that capitalize on price and volume fluctuations. By analyzing historical data and human behavioral patterns, High-Flyer aims to identify and exploit market inefficiencies. Unlike fundamentals-based active investing, quantitative strategies focus on shorter-term price corrections, with High-Flyer relying on high-frequency and mid-frequency trading to generate returns.
Lu acknowledged that China’s relatively nascent market structure provides greater alpha opportunities compared to mature markets like the U.S. However, he stressed that High-Flyer’s competitive edge lies not in specific strategies—which have limited lifespans—but in its robust research and development (R&D) ecosystem. With a team of 120 specialists, advanced AI infrastructure, and a significant annual reinvestment in R&D, the firm maintains a rapid cycle of strategy development, testing, and iteration. Its cutting-edge computing capabilities allow for the training of complex neural networks, enabling the firm to process vast amounts of data and refine its models for future growth.
Despite these advantages, Lu recognized the capacity constraints inherent in high-frequency trading. As market turnover increases, the firm has diversified its strategies across a broader frequency spectrum to manage over 50 billion yuan in assets without overloading the market. He emphasized the importance of scaling efficiently and adapting strategies to maintain performance as market conditions evolve.
Looking ahead, Lu predicted a decline in industry-wide excess returns as China’s market becomes more institutionalized, with retail investors giving way to professional participants. This trend will intensify competition, creating a “Matthew Effect” where top players consolidate their advantages through reinvestment in R&D.
Before going into the full interview, which gives a comprehensive, in-depth introduction to DeepSeek’s parent and also the prototype of DeepSeek long before its recent rocketing to international fame, I think I should mention one more background, and that is the Chinese government has been cracking down on quantitative trading in its domestic stock market, to the applause of numerous Chinese retail investors, in the past year, which the Wall Street Journal terms as “broke the quant” in March 2024
Quant funds, rushing to sell their holdings of small-cap stocks, had to accept heavy losses. Their need to sell had greater urgency because some of the quant strategies had taken on leverage, borrowing from brokers to increase the amount they could invest.
“Everyone was trying to sell at the same time,” said Hao Hong, chief economist at Grow Investment Group. “The exit door was too small, there was a stampede going on, and the market just collapsed.”
Some quant funds caught in the meltdown saw their year-to-date returns plunge by 30 percentage points, according to Rayliant. Some smaller funds have closed.
Mingxi Capital, a Shanghai-based quant fund, issued a statement on Feb. 8 telling investors that the market data and the patterns it had previously used to trade were no longer a good guide. As a result, many quant funds’ algorithms—including its own—went from “getting it right to being repeatedly wrong.”
Minghong Stock Selection Fund I, a flagship strategy launched by Minghong Investment, is down 8.8% so far this year, according to fund distributor Simuwang. It made a double-digit return last year. High-Flyer Quant, another firm, has lost an average of 9.5% across its funds in 2024. Since it was established in 2016, it has returned an annualized 13%.
The rapid change of fortunes for China’s quant funds led to criticism about their strategies and use of leverage. But the bigger problem for the industry was still to come.
Frozen out
On Feb. 19, the Shanghai and Shenzhen exchanges froze the accounts of Ningbo Lingjun for three days. The reason: The quant fund sold around $360 million of shares in less than a minute, at a time when Beijing expected market participants to play their part in boosting the stock market. The exchanges said they had initiated “public condemnation” of the fund.
“The central government could not tolerate having the national team buying only to find that the quant funds were selling,” said Colin Liang, head of China research at Redwheel, a global asset manager. “That was against the broader market and national interests. That was the red line.”
After Nvidia’s enormous success, some in China blamed the government’s crackdown on its gaming industry for the country's inability to nurture a similar giant and cede the Graphics Cards business to the U.S. These reflections tried to underline that no central planner would have foreseen that computer game hardware, which in many Chinese minds is electronic heroin, would one day become strategic assets.
The likelihood is, though, that both the China hawks in the West and the Chinese government will now say that the Chinese government has been cultivating and supporting High-Flyer Quant and DeepSeek.
LOL
幻方投资陆政哲:完全依靠人工智能的量化投资模式
Lu Zhengzhe of High-Flyer Quant: A quantitative investment model fully reliant on artificial intelligence
June 10, 2020 by 朱昂 Zhu Ang
Zhu Ang: Could you share High-Flyer Quant’s investment strategy?
Lu Zhengzhe: Since its establishment, High-Flyer Quant’s investment strategy has undergone significant changes, but it has always been centered around a large-scale AI framework. The core of our approach lies in training models using deep neural networks, which is why we define ourselves as a hedge fund that fully relies on artificial intelligence for investment.
From data collection, processing, and analysis to building AI models with neural networks, portfolio generation, and executing trades with refined algorithmic systems, every step of our process has been fully automated through an AI platform.
Our model draws on three main data sources: first, market data, which includes a series of price and volume indicators; second, macroeconomic fundamental data covering the entire market; and third, alternative data that has been structurally processed. These three categories of data are input into our models to predict price changes for stocks over specific periods, which then serves as the foundation for constructing our portfolios.
If our stock selection model is effective, the selected stocks will outperform the broader market, generating what we call “excess returns.”
Zhu Ang: Active managers view short-term market fluctuations as random walks, but quantitative investing captures patterns within these fluctuations. Do you believe these random walks have discernible rules?
Lu Zhengzhe: The concept of “random walks” has long been debated in both the financial and academic worlds. Theoretically, in a perfectly efficient market, short-term fluctuations should resemble random walks. However, in reality—especially in China—where there is a high proportion of retail investors, trading often involves emotional factors, which is essentially within the domain of behavioral finance. Over the past decade, the Chinese market has experienced three major waves of volatility: rapid surges followed by sharp declines. In contrast, the U.S. market has shown long-term stable growth. This partly reflects various factors influencing China’s economic growth and partly the relatively nascent structure of its market participants.
As such, China’s market is not fully efficient, making investor behavior and decision-making patterns valuable research topics. Market information is ultimately reflected in price and volume systems, such as price changes, trading volume fluctuations, and technical movements. Behind these metrics lie human preferences and behavioral characteristics. We strive to identify the underlying connections by analyzing extensive historical data and modeling them to predict price movements.
Over the years, whether it’s active or quantitative investing, the essence has been about correcting price deviations and then trading based on those deviations to generate returns. Fundamentals-based investing focuses more on longer-term value correction, whereas quantitative investing emphasizes shorter-term price fluctuation trading.
Competitive Edge: R&D and Computing Power
Zhu Ang: Is it fair to say that China’s investor structure gives quantitative investing funds a higher alpha compared to overseas markets because domestic counterparts are weaker?
Lu Zhengzhe: Absolutely, that’s a clear reality. In the U.S. equity market, a hedge fund generating an annualized return of 8%-10% is already considered exceptional because the market is extremely competitive, highly institutionalized, and heavily concentrated in top stocks. Simply outperforming the index is already challenging. In contrast, the A-share market still offers abundant alpha opportunities, although it is rapidly becoming more institutionalized.
Zhu Ang: High-Flyer Quant is a leading player in China with strong market recognition. What do you think distinguishes High-Flyer’s alpha or competitive advantage from other quantitative funds?
Lu Zhengzhe: From 2016 to 2020, the domestic quant industry experienced explosive growth. When breaking down the strategies that drove this growth, we found that the core drivers were strategies based on price and volume signals, which are relatively high-frequency technical analyses. They were not focused on fundamental factors. High-Flyer Quant, as one of the major players in price-volume-based stock selection, grew alongside this trend. The largest source of High-Flyer’s alpha is trading opportunities created by market volatility.
When it comes to competitive advantages, we don’t define them by specific strategies because every strategy has a lifecycle, and that lifecycle is getting shorter globally. Instead, High-Flyer’s strength lies in its ability to continuously develop new strategies.
First, we have an exceptionally strong development team with 120 people, which is rare in China’s private equity investment industry. More than half of our team specializes in AI, IT, or research, supporting a comprehensive strategy development system. Unlike many institutions, we don’t have traditional portfolio managers. Our strategy development system is a highly collaborative, team-based production model. From factor exploration, signal discovery, and data processing to training neural network models, portfolio optimization, risk management, IT infrastructure, and platform architecture, all of these processes are executed through vertical team collaboration. This allows us to develop, iterate, upgrade, and test strategies at an incredibly fast pace.
Second, we have the most competitive computing power system in the industry. Our hardware and software infrastructure is currently the most advanced in China, with nearly 200 million yuan invested over the past three years. Of that, more than 100 million has been allocated to hardware alone, enabling us to build a supercomputing cluster that supports the training of our complex neural network models. While this investment may not yield immediate results, it provides robust long-term support for our research and development efforts.
Zhu Ang: Will high-frequency strategies face capacity bottlenecks in the long term?
Lu Zhengzhe: It’s important to clarify what we mean by high-frequency strategies. According to international standards, pure high-frequency trading does not exist in China’s stock market due to limitations in market structure, trading rules, and signal speed. What is often referred to as “high-frequency trading” domestically is not the same as its overseas counterpart.
At High-Flyer Quant, we have extensive experience with high-frequency strategies, but our primary focus is on mid- to high-frequency strategies. Quantitative investing naturally faces a capacity constraint: our trading volume cannot exceed a certain percentage of the market’s total trading volume. As frequency increases, capacity limitations become more pronounced.
Based on our estimates, achieving satisfactory returns in high-frequency trading requires acknowledging these capacity limits. Therefore, we develop strategies across a broader frequency spectrum. Our current strategy framework can manage over 50 billion yuan in assets, which is feasible given the current market turnover of 500 to 600 billion yuan. With our multi-strategy approach, we expect to further expand capacity.
Zhu Ang: You mentioned supercomputing capabilities earlier. How does your AI Lab’s computing power contribute to High-Flyer Quant’s investment strategies?
Lu Zhengzhe: Before we established AI Lab, we relied on general-purpose computing power to support our platform. These were sufficient when our models were relatively simple. However, as our models became more complex and required exponentially more computing power, general-purpose systems could no longer meet our needs.
Our AI Lab’s advanced computing power has opened up new possibilities for the richness of our models. The data training and model development we now perform were previously unachievable. While these advancements don’t immediately reflect in our net asset value, they accelerate strategy development. Additionally, this computing power enables us to handle more complex models and massive amounts of parameters. Finally, it improves our ability to process large-scale real-time information, paving the way for higher-frequency strategies and execution in the future.
One of the Few Fully Localized Quantitative Teams
Zhu Ang: Could you introduce High-Flyer’s product lineup?
Lu Zhengzhe: High-Flyer Quant is one of the few firms in the industry that has developed a relatively standardized product lineup.
Based on our assessment of different clients’ risk preferences, we divide our product lines into two main categories:
1. Index Enhancement Products with a Long Bias: These products are benchmarked against various indices but are essentially long-only strategies. They cater to clients with higher risk appetites.
2. Hedging Strategy Products: These products are designed for clients with lower risk tolerance.
Both product categories share the same underlying stock selection model. In fact, as long as a single stock selection model is optimized to the extreme, it can support nearly all of our front-end products. The primary difference between the two categories lies in their structure:
• The first category, index enhancement products, is fully invested and thus influenced by market beta.
• The second category, hedging strategy products, deploys a portion of its capital in derivatives to hedge market risk, providing relatively stable alpha returns.
While the operational mechanisms of these two product types are the same, the key distinction lies in the risk preferences of their investors. The foundational strategy, architecture, and position ratios across the product series are quite similar. Over the long term, the performance of products within this system tends to be consistent.
Zhu Ang: Can you tell us about High-Flyer Quant’s development journey?
Lu Zhengzhe: It has been over a decade since we began trading. As early as 2008, when the founding team was still in university, we started exploring quantitative trading. At that time, very few people in the market were aware of quantitative trading. Because our team came from engineering and computer science backgrounds, we sought to explore the market systematically and programmatically. Between 2008 and 2015, before the company was formally established, we were in a period of self-guided experimentation.
Unlike most domestic quantitative private equity investing firms, whose founders or technical teams often have backgrounds in mature overseas institutions, High-Flyer’s team is fully localized. From the beginning, we pursued quantitative investment purely out of interest.
By 2015, our exploration had matured. We had established our trading system and developed highly effective strategies, which led to the creation of the High-Flyer Quant brand. By 2017, we had built a comprehensive multi-factor strategy framework, launched actively managed products, and began entering the asset management business.
From 2017 to 2020, during a period of rapid industry growth, High-Flyer was fortunate to be a part of this expansion. During this time, our assets under management exceeded 20 billion yuan. As early as 2015, we had actively experimented with machine learning and deep learning models for live data testing. By 2018, we had fully integrated our product line and asset management into a core strategy development system driven by artificial intelligence.
Between 2018 and 2020, we focused on two major initiatives:
1. Improving the foundational structure of our team.
2. Aggressively recruiting AI talent with expertise in research and development.
We also established an AI Lab to push further along the technological development path. At the time, almost no other quantitative investor in China had ventured down this road. While we faced many uncertainties along the way, we proved that it was a viable path.
After building our first supercomputer and high-performance computing cluster, we rapidly transitioned to production, testing new strategies. Today, we reinvest approximately 70% of our annual revenue into research and development.
Zhu Ang: Many quant funds in China rely on T+0 strategies to generate alpha. What’s your perspective on this?
Lu Zhengzhe: T+0 strategies are common in high-frequency trading systems in China. They involve building a basket of stocks based on an index and trading within the day based on price movements. If paired with index futures for hedging, these strategies produce smooth, upward-sloping returns. Without hedging, they resemble enhanced index strategies. T+0 trading provides stable returns and plays a supplementary role in our overall strategy framework.
Machine-based T+0 has the advantage of monitoring the entire market in real-time, handling massive data volumes, and executing trades on a large scale. While individual trades yield small profits, the high frequency ensures consistent returns. Human-based T+0, on the other hand, can capture larger profit opportunities but is limited by the number of stocks a single trader can manage. Overall, machine-based T+0 is more advantageous, though top traders can occasionally outperform machines.
Quantitative Investing Will Also Face the Matthew Effect
Zhu Ang: What are your thoughts on the future of fundamental-based quantitative investing in China?
Lu Zhengzhe: Quantitative investing in China began exploring fundamental-based strategies quite early and has a relatively long history. The first generation of quantitative investors in China used fundamental factors for stock selection. However, after 2014 and 2015, there was a significant shift in market trends, and many previously effective alpha factors began to fail. In recent years, the most successful strategies have not been based on fundamental factors but rather on price-volume, technical quantitative strategies.
That said, we still believe fundamental quantitative investing has substantial potential. Fundamentally, it aligns with the logic of active investing, even though the scope of evaluation differs. As long as active investing continues to generate good returns in China, there will still be significant opportunities for fundamental-based quantitative strategies.
Years ago, fundamental-based investing primarily relied on financial statements and basic data, which was challenging for two main reasons: first, the quality of financial data from domestic companies was not very reliable; second, the high volatility of the Chinese market created additional difficulties for fundamental quantitative strategies. However, with the widespread application of Internet technologies, we now have access to a growing range of alternative data and relatively high-frequency data, offering new avenues for fundamental-based quantitative investing. This allows us to perform better than before.
In addition, High-Flyer Quant is working to refine models that predict financial indicators at the individual stock level within various industries. We aim to integrate these predictive data with trading strategies in the future. Therefore, fundamental quantitative investing still has significant room for growth in China.
Zhu Ang: Quantitative investing essentially profits from market inefficiencies. In less mature markets, like China’s A-share market, it’s easier to generate returns. As the A-share market becomes more mature, will it become increasingly difficult for quantitative investing to succeed in China?
Lu Zhengzhe: Over the long term, the industry-wide excess returns for quantitative investing are gradually declining, and this is a natural trend. As retail investors exit the market and institutionalization progresses, both quantitative and active investors will face challenges, and the share of the pie available to everyone will shrink.
1. Increased Competition: With more players entering the market and the growing institutionalization of the market, emotional, momentum-driven investors who chase gains or cut losses will eventually exit. As a result, the space for alpha will become increasingly smaller.
2. Decline in Excess Returns: The decrease in excess returns is not limited to the quantitative investing sector—it’s also happening in the active investing space. Some of the most successful value-oriented fund managers still have their place because value investing in China is relatively rare. It’s precisely because value investing is scarce that it generates alpha in the current Chinese market.
Another trend is the Matthew Effect, where success becomes increasingly concentrated among top players. In recent years, both active and quantitative investing have shown this trend: the better the performance of top institutions, the larger their scale becomes. As their scale increases, a significant portion of their revenue is reinvested into R&D systems, further enhancing their competitive advantage—especially for quantitative institutions. We believe that in five years, the market structure will become much clearer.
Zhu Ang: High-Flyer is already a leading quantitive investor in China. What are your expectations for its future direction?
Lu Zhengzhe: I prefer to call this our “vision.” We have two primary aspirations:
1. Longevity: This is a very practical goal. Looking at both domestic and international predecessors, very few institutions have maintained a leading position for more than 10 years. Most rise rapidly, scale up quickly, and then collapse just as fast. This seems to be a vicious cycle in the quantitative investment industry, like a Sword of Damocles hanging over our heads. Because we want to break this cycle, we invest heavily each year in strategy development systems, as well as hardware, IT, and technological advancements. These investments are not just for today’s returns—they are meant to prepare us for the increasingly intense competition we anticipate in the next 2–3 years and beyond. We aim to ensure that we’re not left behind and can thrive for the long term.
2. Using Technology to Invest: We want to demonstrate that this is a viable path. Many institutions overseas have already proven this, but in China, there has not yet been a firm that has rigorously followed this technology route. Our attempt comes with foreseeable challenges, but we firmly believe that it is a path worth taking—and one that can succeed.