“Self-developed models, rushing toward the pinnacle of trading.”
Written by｜Sun Xiaowen, Hai Ruo Jing
Edited by｜Hai Ruo Jing
Before starting his AI finance model venture, Zichen was already earning a yearly salary of a million dollars as a “gold collar” professional. With his trading instinct honed at top investment banks, making money was not a difficult task for him.
The inspiration for entrepreneurship came from an AI trading competition called Alpha Arena. At the end of 2025, major models like DeepSeek, GPT, and Qianwen participated in the competition, leading Zichen to discover that general large models had evolved a certain level of independent judgment ability. Perhaps in the future, his proud investment decision-making ability could also be replicated by large models.
Rather than panic, it’s better to dive into the waves. He found Fang Zheng, who was engaged in model research at Meta, and the two hit it off, beginning an AI entrepreneurial journey named FreeRide. “Dark Waves” has learned exclusively that FreeRide AI recently completed a multi-million dollar seed round of financing, with L2F Light Source Entrepreneur Fund as the investor.
Securing funding even before having a Demo is due to two key points: one is the chosen direction, and the other is the background and potential of the founding team. After graduating from the mathematics department at Oxford University, Zichen worked at institutions like Morgan Stanley, JPMorgan, and Huatai International in London and Hong Kong, engaging in complex derivative trading, eventually serving as a senior trading manager.
Fang Zheng’s background is even more diversified. After graduating from the physics department at Oxford University, he worked at UBS, Evercore, and Macquarie Capital, managing investment banking business in Europe and the USA. Recognizing the immense potential of AI in the financial sector, he returned to Cambridge University to pursue a PhD in computer science and participated in AI model training and AI system research at major companies like Meta and Oracle.
The two began to explore: large models already exhibit decision-making intelligence, but how many more mountains must be climbed to enable AI to possess the capabilities of a “Wall Street trader”?
The first unavoidable challenge is “data.” Text and code can be massively collected, but truly high-quality experiences that impact financial trading often only exist in the minds of frontline practitioners. Even if many types of databases can be purchased, the challenge of judging data value and weight and converting it into real-world decision-making intelligence remains vast.
Moreover, financial trading data is inherently high-noise, and the long process models adapted around coding agents are challenging to migrate directly.
From an architectural perspective, financial trading tasks often rely on extremely high-dimensional logical analysis. Many tasks are not point-based or unidirectional solution paths; relying solely on conventional model training methods is insufficient, and specialized system learning architecture is needed.
Even if a finance-specific model possesses the necessary expertise and ability, financial markets are ever-changing. Only by allowing AI to learn autonomously and self-evolve can it capture real market signals.
These obstacles have clarified their technical route: to create an AI financial trading brain, one must self-train a specialized vertical model foundation while also developing a self-evolution Agent. Based on experiences gathered in the financial industry, they adjust the parameter weights of generalized large models to evolve a more practical LLM foundation, first allowing AI to become a junior employee with 2-4 years of experience; then through the self-evolution Agent, enabling the model to adapt to the ever-changing real-time financial trading.
Self-developed models are undoubtedly expensive and challenging. However, Zichen and Fang Zheng believe that this is the essential path to reach the “pinnacle of finance.” Once established, vertical models may reach the pinnacle of Wall Street investment banks and frontline funds, participating in derivatives pricing, complex investment research, and decision-making; in the future, they could radiate into a broader financial market and even serve as a trading “co-pilot” for ordinary people.
In their view, while model giants and internet platforms possess more computing power and resources, they find it difficult to deviate from the current mainline of AI development and focus on this segmented financial scene. Future AI systems might resemble a symphony orchestra: general large models collaborating with specialized models to accomplish complex tasks.
Aside from training finance-specific models, they plan to publish a paper in 2026 to preliminarily verify the feasibility of self-evolving Agents in high-noise, high-variation financial domains. This “Alpha-seeking researcher” mechanism was tested in the CSI 300 market, where after 1,600 evolution steps, it raised the median IR of generating factors from 1.6 to 3.9. This performance is quite commendable; in traditional quantitative fields, an IR greater than 1.0 is considered an excellent factor.
With the initial technological verification complete, FreeRide will next delve into specialized tasks like derivatives pricing, producing models usable by financial professionals, before expanding to broader financial tasks.
The following is a conversation between “Dark Waves” and FreeRide co-founders Zhang Zichen and Shen Fangzheng (edited):
Part01
AI Symphony Orchestra:
General models leading a group of specialized models
Dark Waves: If general models can accommodate Agents and Harness, why still train underlying models yourself?
Zichen: We need to apply first principles to solve intelligence issues. Currently, the architectures of Agents and Harness are becoming increasingly complex, but if you do not grasp the underlying model weights, even the most exquisite top-level design will face risks, such as the general model not meeting the quality of task for specific scenarios.
On the other hand, institutions will eventually have their own Agents, but which model will drive this Agent? The current general large models may not be suitable; we hope to create a model more suited for the financial scene, giving these Agents a chance to utilize the FreeRideAI model.
We have compared GPT, Claude, DeepSeek, Qianwen, MiniMax, and Zhipu. Regarding derivatives pricing issues, they indeed understand many foundational concepts, but they are more like good interns. If you are interning at an investment bank and provide such answers, I think it is OK; but as a formal employee, you would not meet the standards, as your answers are too superficial and lack practical market experience.
Dark Waves: How do investors view financial vertical models? Are they concerned about being swallowed by general models?
Zichen: A few months ago, everyone was worried that general models could consume everything, but recently, there has been a shift.
General large models have several limitations, including performance, cost, and speed in specialized fields. For specific professions, model size isn’t necessarily better; the final form may be a collaboration between the general large model and a series of specialized models to complete tasks.
The current general models have excelled in programming-related areas but still require specialized models in fields like finance and healthcare. This is because certain hands-on experiences in finance cannot be found online and won't be in books either. A tech worker finds it challenging to replicate top abilities in vertical fields.
Fang Zheng: Rather than calling them vertical models, calling them specialized models may be more accurate. Similar to liberal education, modern AI has built a broad knowledge base during early training, but to excel independently in a specific area like a professional, highly specialized knowledge and professional skills must be cultivated.
Dark Waves: How do you determine what financial models should learn?
Fang Zheng: If we compare model training to “alchemy,” we first need to know who this “elixir” is for and what problem it solves. Because our team has rich professional experience in international financial centers like London, New York, and Hong Kong, we know what financial institutions and professionals need AI models to do, allowing us to deduce the necessary capabilities and types of data required.
Dark Waves: Is it a must to buy large amounts of commercial financial databases to train specialized financial models?
Zichen: We do need to acquire databases. But the true experience includes two points: first, finding this data; second, judging their value and weight. When faced with 100 data points, which is more important? How to turn this data into something useful for decision-making? What is the chain of decision-making thought? Such experiences previously existed only in a small financial circle, and people wouldn’t record and share them.
For instance, for a specific parameter in option pricing, there are very few online resources or documentation on how to choose under specific conditions. But a team with frontline trading experience knows what is right and wrong, which existing data weighs more and which weighs less.
If you just throw the data to a model expert saying, “I have this data, please train it,” the resulting intelligence level will be very low. It’s like a manager asking a junior analyst to analyze a problem, and if they just throw data back, the manager will definitely ask: “What do I want you to do?”
Dark Waves: Is it sufficient to rely on the experiences of a small circle, given the data volume?
Zichen: Firstly, our team is capable of amplifying the relevant training data volume based on real scenarios targeting the problems that need solving.
Additionally, apart from experiences within the financial circle, we are also doing something more interesting, which is creating a mechanism for generating “simulated experiences.” It’s somewhat like an AI Stanford town, simulating human behaviors where various roles interact in the AI world. We hope to use existing experiences and AI to scale up the data volume and simulate more valuable experiences.
Dark Waves: Do you have specific examples?
Fang Zheng: Key technical details are not convenient to disclose too much at this stage.
But as an example, take the STAR mechanism disclosed in the paper; its original data scale is quite small, consisting entirely of limited, discrete, and highly concentrated information and noise.
Even so, we still see valuable observational results can be deduced, guiding state judgments and decision-making behavior in later stages. Although applied to self-evolution and traditional post-training differs, it confirms that a lot of information in the financial market can be effectively extracted and can serve as signals to drive the overall iteration and optimization of AI systems and models.
Dark Waves: How do existing benchmarks in the financial field look, and how do your specialized models perform?
Fang Zheng: The current benchmarks in the financial field aren’t very ideal. For instance, finding some numbers in financial reports is essentially information retrieval; or simple calculations which essentially place general tasks in a financial context.
We see that in recent times some benchmarks seek professional annotations, for example, evaluating what angles to consider in a particular question. However, many are still quite rough and do not reflect real industry concerns.
Zichen: We won’t blindly chase rankings. The financial sector is vast, and a benchmark set might only encompass a few hundred questions. Tax, accounting, and retail financial issues grouped with institutional financial issues have relatively narrow coverage. We will selectively evaluate which benchmarks are more suitable for us.
Part02
Self-evolving Agent:
Instantly perceiving market changes
Dark Waves: Why train a model while also allowing the Agent to self-evolve?
Fang Zheng: They are complementary. Due to the limitations of general models in financial tasks, solely focusing on Agent evolution still has efficiency and effectiveness bottlenecks. Meanwhile, if only training models, the financial market changes very quickly, making it impossible for models to adjust in real-time based on market dynamics.
Sending a competent but basic ability “high school student” to Wall Street for trial and error on financial tasks means their progress to becoming an internationally top-tier professional will far lag behind introducing a well-prepared graduate with solid expertise and high adaptability from a notable university. Training finance-specific models is about producing “AI financial graduates.”
We hope the Agent can continuously perceive changes in the market, evolving slowly up the career ladder like humans.
Dark Waves: How does the STAR self-evolution mechanism dig for Alpha, and how does it differ from common Alpha digging? How will this mechanism be used in products?
Zichen: It’s quite different. In general, people directly search for Alpha with predetermined directions. However, our design is for the Agent to evolve into a researcher proficient in finding Alpha.
STAR has incorporated a meta-cognition module, enabling adjustments to the researcher’s prompts, memory, tools, and research processes. During searches, it doesn’t just choose the currently highest-scoring individuals; it prioritizes expanding lineages that can generate stronger offspring. The reduced number of evolution steps and costs in the paper has nearly tripled the number of operators used by the generated factors.
These are insights it arrived at through self-evolution, which we did not inform it of.
Reflecting on the product, we can allow users to define the direction they want to evolve. For example, if I currently have a certain idea or strategy, can you help me evolve through the STAR mechanism to see how it improves for future generations? This is one of the capabilities we have a chance to develop into an “AI researcher.”
Dark Waves: However, the experimental environment of the paper is very simple. How far are we from genuinely assisting institutions in making investment decisions?
Zichen: The core goal of this STAR paper is not to optimize a single Alpha factor, but to evolve a “Alpha researcher” system with autonomous refinement capabilities. As a researcher, it can integrate market conditions, distill思维框架, historical experience memory, and advanced mathematical operator tools, collaboratively outputting quantifiable factors suitable for strategies.
There is no constant wealth code in this world.
Nurturing research-oriented Agents capable of integrating strategy frameworks, experiences, and mathematical tools is more meaningful than simply optimizing factors. It also aligns better with the working logic of human researchers. The factors proposed by the model can indeed be evaluated through various samples and indicators like IC, IR, and Sharpe ratios in backtesting, but transforming these factors into final investment decisions and returns still requires going through investment portfolio construction, risk control, and trading execution stages — this is precisely the standard operational process for professional financial institutions.
Most institutions already have matured mechanisms. Our initial aim is to integrate with them. For example, I provide you with a good Factor, and you have many more Factors; you could blend them to create a composite.
Dark Waves: How do you assess if the analysis provided by the Agent is professional enough?
Fang Zheng: There can be many approaches here, one of which is to convert “high-difficulty validation” into “low-difficulty validation.” Depending on different scenarios, there can be many specific implementation details, but the core goal remains consistent, that is to simplify the overall validation processes through the conversion into correctness determination standards.
However, various validation methods cannot entirely replace the importance of the industry’s top experts. We have a very large team of experts to support us, comprising senior practitioners from the top financial institutions on Wall Street, with generally over ten to twenty years of experience.
Zichen: Let me give you an example. For instance, if you ask it how to price an option, a typical answer might involve looking at the stock price, volatility, borrowing rate, and dividend yield of the stock, which are all things seen in textbooks.
However, we hope it considers more. For example, whether this pricing method encompasses analyses about the industry and similar sectors, whether predictions for future events, analyses on future capital flows for stocks and industries, etc., are taken into account. If an AI model tells me how an option should be priced and covers all these points, then I would deem it professional in the financial scene.
Part03
The Path Ahead
Dark Waves: Who does FreerideAI first aim to serve? What are their biggest pain points?
Zichen: First and foremost, we hope that financial institutions and financial professionals will be satisfied. If we can achieve this, it indicates it has reached a high level of intelligence. Later, we can extend it to more ordinary people, which is something we also aspire to in the future.
The current issue is that people fundamentally lack confidence in allowing AI to assist in investment decisions. For example, if you ask DeepSeek whether to buy the dip tomorrow, it will give you many professional-sounding answers, but some contain AI hallucinations, and others sound like textbooks providing only concepts. From a professional standpoint, they are basically unusable.
Dark Waves: At this stage, what level do you hope the model will reach? When can we expect to see initial products?
Fang Zheng: We aim to achieve the capabilities of a junior employee in foreign investment banks or top hedge funds soon, essentially the level of employees with 2 to 4 years of work experience.
As the model training advances and the adaptation mechanisms improve, we can gradually enhance the experience level. Obviously, initially, this will focus on scenes and professional fields we are training, and later we will progressively broaden the scope.
Zichen: We plan to launch an initial Demo in Q4, at which point specific scene previews will be available. Derivatives pricing is one area we will focus on, and in the future, it will gradually generalize to secondary market trading, primary market research, and more.
Dark Waves: Are you currently in close contact with investors? When do you plan to start the next round of financing?
Zichen: Recently, we’ve been heavily focused on product development and model training. Once the Demo comes out next month, we will begin meeting with some investors to prepare for the next round of financing.
Typeset by｜Guo Xukun
Image source｜Unsplash
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
