AI能帮股民回本吗？一个抓住十倍股的人正在训练预测模型

Can AI Help Investors Break Even? A Person Who Caught a Tenfold Stock Is Training a Prediction Model

AI prediction is between

quantitative and subjective investment research, a new species.

Writer | Liu Sijie

Editor | Zhang Wei

Source | Future Human Lab (ID: LabforAI)

Cover source | Pixabay

In 2025, 25‑year‑old Henry went all‑in on a ten‑fold stock, Nebius. Thanks to that investment, he didn’t need to work for a long time.

Nebius is a company that does AI neocloud, i.e., the new cloud. It was not a company that was widely regarded as promising at the time; some researchers who took a wait‑and‑see attitude looked at its quarterly profits and judged its value based on the “computing power landlord” business model, but Henry thinks, “These judgments are short‑sighted because it’s hard to accurately predict every number.”

Henry’s perspective is different. He is bullish on Nebius, focusing on the company’s long‑term, fundamental driving forces: the people and the organization.

Beyond his investment track record, we want to highlight a more core, more interesting viewpoint presented by this young man—

“Stock selection and forecasting are essentially the same thing,” he says. “Understanding and handling them from different dimensions can help us forecast better.”

This perspective is also the foundation of his entrepreneurship. He earned his bachelor’s degree in computer science at Shanghai Jiao Tong University, pursued a master’s in Fintech at Shanghai Advanced Institute of Finance, worked in quantitative finance and discretionary research, and in May this year founded the AI forecasting company ProphetLab, combining AI predictions with human expert experience.

“Future Human Lab” exclusively learned that the company has now secured a multi‑million‑yuan angel round, including investment from Ivy Capital.

Before officially launching the startup, Henry spent more than half a year in a “hibernation period.” He lived in a rented apartment in Shanghai, independently researching the intersection of AI and investing, and during that time he seized Nebius, which later multiplied more than tenfold.

During those six months, Henry gradually formed a judgment: the combination of AI and investing is not merely automating existing quantitative methods, but could become a “new species” that sits between quantitative investing and subjective research.

In his view, quantitative investing excels at handling structured data and building models, while subjective investing relies on human experience, intuition, industry understanding, and a large amount of unstructured information that is hard to encode. Historically, there has been a huge gap between quantitative and subjective research, making talent and methods difficult to cross‑communicate. The emergence of large language models offers the first possibility of “building a bridge”: they can comprehend human analytical processes, abstract the judgments of subjective investors about industries, companies, and trends, as well as those originally unstructured, non‑programmatic pieces of information, and then combine them with quantitative data‑processing capabilities.

AI prediction companies are exactly this new species; past experiences repeatedly confirm that they can make accurate predictions with traceable evidence.

He wants to teach this ability to AI.

The information gathered must be abundant, and the ability to process it must be strong; AI surpasses humans in both aspects. After all, large models are increasingly adept at searching, analyzing, using tools, and completing tasks, and combined with human experts’ experience, “it may predict things that others cannot.” Henry expressed it this way.

For example, predicting Tesla sales: traditional analysts might only look at a small portion of online comments, whereas AI can examine many more reviews, as well as maps and other information, so it is likely to be more accurate.

ProphetLab’s first product is World Cup predictions.

Initially, they only ranked the predictions of major models. After funding arrived, the team began adding a self‑developed prediction framework: using API calls to large models, combined with historical matches, player injuries, referee styles, weather data, etc., to simulate sequences of events in a match, then running many simulations to obtain a probability distribution of scores and produce the final prediction.

According to Henry, this system achieved a prediction win rate of over 70%, correctly forecasting the tournament’s semifinalists, champion, and a 1∶0 final score.

Currently, ProphetLab is advancing collaborations with fund institutions to develop a benchmark for financial forecasting. They aim to test whether models can make evidence‑based judgments about events that have not yet occurred by tackling multi‑step analysis problems such as company sales and revenue. Financial experts participate in problem design and model evaluation, and the data and feedback gathered from the assessments will be used for subsequent model training.

They also did a fun little project, analyzing AI industry figure predictions. They collected their past public statements, checking each one to see if it was correct, seeing who tends to exaggerate and who is more precise. That way, next time they make a judgment, they’ll know whose words are more worth referencing. Henry calls this “analyzing human experts’ analyses to predict human experts’ predictions.”

Currently, the highest prediction accuracy belongs to former OpenAI chief scientist Ilya Sutskever, and the lowest is Elon Musk.

He believes in the allure and societal value of forecasting. If we can predict the future more accurately, people can make better decisions and reduce waste of resources.

However, regarding his investment track record, we asked him why he’d start a company when his investing skills already make money.

Henry mentioned his favorite game, Zelda. “Starting a company is like playing ‘Zelda’; there’s a boss you must defeat, you have to find ways to gather resources, carve a path, and get as close to that goal as possible.”

Exploration is the driving force behind his continuous investment. “Prediction is a long-term, ten-year challenge; I hope to create a prophet.”

When faced with something that hasn’t happened yet, how can AI identify truly important signals amid massive information and noise, forming judgments more accurate than humans? And when people begin to believe a prediction about the future, they may not take action to change that future.

Below is a conversation between the “Future Human Lab” (hereafter “Future Humans”) and Henry.

Henry

How does AI predict something that hasn’t happened yet?

Future Person: Why start with the World Cup for predictions? Do you personally like soccer?

Henry: I don’t play soccer, and I don’t watch it much. When people started talking about buying tickets for the World Cup, I thought, could a prediction company use this event hype to promote its brand? We also had hiring needs at the time.

At first it was mainly me and two AI PhD students working on it, and I spent some money myself. Later, posting the prediction results on social platforms every day did attract some people. Now the team includes a colleague with a Cambridge mathematics background who found the predictions interesting, traveled from Hangzhou to Shanghai several times to discuss these things with me, and eventually joined.

Future Person: I currently see on the official website that the World Cup predictions seem to mainly showcase results from multiple models. Which part are you actually doing?

Henry: When we first prepared for the World Cup, the first match was very close, we hadn’t secured funding yet, and resources were scarce. At that time we just wanted to make a simple leaderboard, let each model submit its predictions, and rank them after each match.

Later, with resources and friends and classmates helping, we started developing our own set of prediction models. The “model” here refers to the algorithm; we built a Harness—a framework for organizing model work and calling APIs. The end‑to‑end simulation you saw is done by calling APIs; the key is the framework and the raw data.

Future Person: How does this framework actually predict a match?

Henry: First we gather a lot of information related to football matches, and also purchase data, including player injuries, home/away status, referee style, weather and humidity, etc. This raw information goes into our framework and is mapped into the simulator.

The simulator makes very fine‑grained predictions, such as an event every minute or every few minutes: a wing cross, a ball cleared by a defender, a turn and a shot. The events combined together constitute a match.

We run the match many times, some ending 1‑0, some 2‑1, and finally obtain a probability distribution. For example, if we simulate 30 games and 5 end 1‑0, we consider the probability of that score to be one‑sixth. Summing the probabilities of all winning scores gives the overall win probability.

Future Person: Are these actions simulated, or did they actually occur in past matches?

Henry: It’s all simulated. But we use information from past matches. For example, in the last ten years of games, we know what events were triggered each minute. After the agent reads this historical data, it then decides its behavior in the simulation.

We also brought in a human expert. Because I don’t watch soccer much and noticed many AI errors, we consulted a referee from the Chinese Football Association to help us analyze the multi-step reasoning mistakes in football. He points out where it went wrong and how it should be corrected, and we store that in our memory.

At the end of each match, when a new game starts, the system reads these records, reminds itself not to repeat the same mistakes, and then makes subsequent predictions.

Future Person: A real referee? What exactly does he correct?

Henry: For example, when simulating a match between Japan and the Netherlands, the model often produces scenes where Japanese players leap high and score with a header. The referee thinks this shouldn’t happen so frequently; his judgment is that the players’ heights and aerial duels weren’t reasonably considered.

We record this information and feed it into the next prediction, so the model won’t make the same judgment so easily.

Future Person: How do you allocate weights to different pieces of information during prediction?

Henry: That’s decided by AI. For instance, in investment research, financial statements and company announcements have the highest weight and confidence, followed by analyst reports from reputable institutions, then mass-market reports and news, and the lowest are rumors on social platforms.

We also check whether the confidence design and AI judgments are good—that’s where human experts’ experience comes in. We don’t directly change specific judgments, but we evaluate how well the Agent performs.

Future Person: Have you ever seen a World Cup prediction that left a strong impression because it was wrong?

Henry: Of course. There were a few days in the World Cup when the predictions were way off, and the comment section said none of them were right, just random guesses. There were also three consecutive days when every prediction was correct, and the comment section called it amazing, calling us “gods.”

Football itself is random and can reverse. Even if the process is very accurate, the prediction can still be wrong. When we say a prediction is accurate, we mean it’s a bit more accurate than the market.

Future Person: So you only care about errors in the process, not about whether the final result is right?

Henry: The result is also stored in memory; the process gets a score, and the result gets a score. But we don’t punish a model just because it guessed a single match wrong. In the Japan‑Netherlands example earlier, there was a problem in the process, so we’ll correct it.

Future Person: If new information appears after a prediction, how do you update it?

Henry: By continuously updating through memory. For example, a prediction is made at time T, and at T+1 new raw data and information become available; the memory system uses the updated system to make a new prediction.

Future Person: Do you use this same simulation‑match approach for other events?

Henry: Football has a dedicated architecture; other events use a general framework with a Deep Research approach. It conducts analysis and ultimately produces a report that includes information from human analysts as well as the AI’s own analysis, all compiled together.

Can AI “predict the future” like an excellent researcher?

Future Person: Between World Cup predictions and business/financial forecasts, which abilities are universal?

Henry: The methods of collecting raw data, as well as Deep Research and multi-step reasoning, are universal. What isn’t universal is the data itself and the experts required for each task. Football needs football referees, finance needs investment researchers, fund managers, and other specialists.

Our current focus is on the commercial and financial sectors; politics and sports are lower priority.

Future Person: Regarding the financial benchmark you’re preparing with fund companies, what exactly will you be testing?

Henry: In financial evaluation, there are two categories: knowledge questions, such as whether the model has memorized financial formulas; and the second category assesses the model’s ability to work, such as the interrelationships among the three financial statements and the ability to use Excel and other tools.

If we compare to human analysts, these are more like the work of junior researchers. Our focus is on forecasting: as a fund manager or senior researcher, one must be able to judge a company’s future development. What we aim to do is this long‑term, complex task.

The model needs foundational knowledge, research, and tool capabilities; on top of that, it must tackle forecasting and complex financial problems. We want to outline the current capability boundaries of the model and then fill the gaps through reinforcement learning and other methods.

Future Person: Could you give an example of a business forecast?

Henry: For instance, what will Tesla’s sales be next week. We first need to freeze time, go back to a specific point in the past, and forecast the sales of a new model in China.

One possible approach is to first analyze the intensity of public discussion, then use satellite imagery to assess how crowded a factory’s parking lot is. Before a new model launch, inventory needs to be prepared, and the cars will be parked next to the factory—this serves as a signal. If enough cars aren’t ready, even strong demand may not translate into high sales.

There are also company financial reports and earnings calls. Revenue among companies in the same industry can be correlated, which provides relatively generic information.

Future Person: Through this task, what capabilities of the model are you trying to differentiate?

Henry: Multi-step reasoning ability. To give some initial observations, a good model might adjust the company’s earnings guidance with its own corrections; a typical model might simply take last year’s revenue and perform a linear extrapolation. This is how benchmarks differentiate model capabilities.

Future Person: What do team members with financial experience do in this process?

Henry: I believe that the experience of some human experts in the investment research process is still ahead of models. Constructing the evaluation questions themselves requires help from colleagues with financial experience.

In addition to the internal team, there are external expert consultants. They have their own research experience and will judge our models like football referees.

Future Person: Do you think AI could predict more accurately than researchers? Why?

Henry: The task is the same, it’s financial forecasting, but the approach may differ. Traditional analysts study a gaming company, aggregating game revenue, cash flow, and other data. AI can gather massive amounts of information, including details for each game and player reviews, and combine them, potentially yielding a better forecasting path than humans.

Human perspective is limited; AI can access knowledge from different domains and conduct web-wide searches, possibly predicting things that others cannot.

Future Person: Is more information always better? Or do we only need precise information?

Henry: I think the more information, the better, provided we have strong processing capabilities. AI may surpass humans on both counts. For example, when forecasting Tesla sales, a traditional analyst might only look at a small portion of online comments, whereas AI can examine many more reviews, locate maps, and search for other data, making its prediction likely more accurate.

Future Person: Do you ever have AI and researchers work on the same problem internally to compare who is more accurate?

Henry: That’s the next step, but we haven’t been developing long enough yet, so it’s not happening now.

Future Person: Do you have any concrete examples to show right now?

Henry: It’s still under development and being fine‑tuned. At the moment we can only give a rough sketch, and it will take more time to build it out.

How do you train “predictive ability”?

Future Person: You’re also hiring people for model training—are you planning to train a base model yourself, or do post‑training?

Henry: Our core team members come from Shanghai Jiao Tong University, Cambridge University, and Shanghai Advanced Institute of Finance, all holding AI PhDs and FinTech Masters. They have participated in training trillion‑parameter base models at Alibaba Qianwen, Tencent Hunyuan, Xiaohongshu Dots, and other Frontier Labs. Currently, we are doing post‑training. We start with benchmarks; once we have evaluations, we can generate data, then use that data to train smaller models and assess performance. This is a relatively standard post‑training pipeline for language models.

We are now doing post‑training on a 27B model, i.e., a model with roughly 27 billion parameters, and it is still in training.

Future Person: Then what is the new predictive architecture you mentioned?

Henry: Researchers are also exploring whether we can create new prediction or decision architectures, not necessarily Transformers. For example, on a Transformer, have the decoder output probabilities instead of text—something along those lines.

But we won’t demand that the research team produce results immediately, or that they surpass agents or fine‑tuned language models right away. It may remain a dream, perhaps one day it will show something interesting.

Future Person: In the end, will we use our own trained model or an external model plus an agent?

Henry: Use whichever works best. If a fund company needs it, or our own forecasts need it, we’ll go with the more effective approach. We’ll pursue several options; we’re quite pragmatic.

Future Person: Who will we sell the prediction results to? What’s the commercialization path?

Henry: Divide into short‑term and long‑term. In the short term, we have benchmarks and evaluation mechanisms, and the datasets behind the evaluations can serve as fuel to improve model capabilities. It’s like we act as a vendor, providing data and reinforcement‑learning environments.

In the long term, if we can predict things that others cannot, it inherently has economic value. Like weather forecasting, it can alert people and also assist decision‑making. Whether to build it as an enterprise‑focused or consumer‑focused product can be discussed further.

Future Person: Model companies and fund companies—do both have demand? Which side are we focusing on now?

Henry: Yes, the focus is probably still on model companies; they need to know the boundaries of model capabilities before they can fill the gaps. The key is still the reinforcement‑learning environments and data—they are the fuel for model training, and customers need them.

Future Person: Has this commercialization path been validated? Is the fund partnership confirmed?

Henry: It’s still running, and the benchmark is also under development. The fund partnership hasn’t been fully confirmed yet; some parts have been internally discussed, but the counterpart company still needs to go through approval.

Future Person: How will we determine whether the company is making progress moving forward?

Henry: In the next four months, first we hope to get the benchmark done well, second our self‑trained model can fill the capability gaps of the existing models, and third the predictions need to be more accurate.

Running our own benchmark may not carry much credibility, so we will also participate in other evaluations. There’s also the market standard, including the stock market. However, actual trading won’t be overly urgent, because human capability is still quite strong. It’s like we wouldn’t demand the research team to immediately produce a new architecture that surpasses all the strongest closed‑source models.

Spent a Decade Building a “Seer”

Future Person: Why did you choose to do AI forecasting?

Henry: Based on my own growth and work experience, I feel this direction suits me best.

Before starting my own company, I spent half a year researching how AI could be combined with investing. This stemmed from my background in AI research and subjective investment analysis, so I wondered how to give AI human-like thinking and decision‑making abilities. I gradually realized that the fusion of AI and investing could create a “new species” that sits between quantitative investing and subjective research.

AI can understand the human analytical process, abstract the subjective investor’s judgments about industries, companies, and trends, as well as the otherwise hard‑to‑structure unplanned information, and then combine them with quantitative data‑processing capabilities.

I did my undergraduate in Computer Science at Jiaotong University, then earned a FinTech master’s at Gaoxin. After graduating, I worked for a while in quantitative finance and then in discretionary investing, so I’m well suited to bridge the gap between quantitative and discretionary research.

Thus I started building forecasting models. This area requires combining AI research with human experts’ forecasting experience.

Future Person: Why switch from computer research to finance?

Henry: I wanted to start a business as early as my freshman year. I knew my ultimate goal was to run a tech company, and I pursued research precisely because I wanted to launch a company.

Later I realized that after two years of research, spending another four or five years would yield diminishing returns for me. Learning across disciplines and understanding how the business world evolves would be far more inspiring.

My parents might prefer that I pursue a PhD, and not many of my computer‑science peers switch to finance or business. But looking at it now, I think this path suits me better. I haven’t done any forecasting; even with strong computing skills, it may not be right for what I’m doing now.

Future Person: You heavily recommended and weighted Nebius last year. Why were you so bullish on it back then?

Henry: The company builds AI neocloud, essentially a new cloud.

I’m bullish on it because I focus on the company’s long‑term, fundamental drivers—people and organization. At the time, some analysts were on the fence or negative, judging value based on quarterly profits and the so‑called “compute landlord” business model. From another perspective, I found those judgments short‑sighted, since it’s hard to accurately predict every number.

Future Person: Which of these stock‑picking and research skills can be transferred to the AI forecasting you’re doing now?

Henry: Stock picking and forecasting are essentially one thing, but there are many ways to forecast. For example, using Excel for valuation modeling to see the gap between a company’s intrinsic value and its current stock price; some also look at changes in marginal cost and marginal revenue.

I focus more on a company’s long‑term, fundamental drivers, which are people and organization. By assessing these capabilities, I trust in a company’s future development. If you limit yourself to quarterly profits, the judgment can be short‑sighted because it’s hard to predict every number accurately. Understanding and handling different dimensions helps us forecast better.

Future Person: What attracts you most about forecasting? Why does it resonate with you so much?

Henry: In my previous investment work, forecasting was a necessity. Now in interviews, through one or two in‑depth conversations, I learn about a person’s experience and growth, forecast the value they’ll bring after joining, and assess whether they’ll work harmoniously with the team.

So I can naturally feel the value of prediction. When a prediction is correct and matches previous judgments, it provides positive feedback and reinforces my cognition. I increasingly realize that cognition is crucial for a person’s growth.

Future Person: Are there any other products related to prediction that we’re working on now?

Henry: We also did a small project analyzing AI‑industry figures’ predictions, collecting their past public statements and checking each one to see if they were right, identifying who tends to exaggerate and who is more rigorous. That way, next time we make a judgment, we know whose words are more worth referencing.

The original idea was that these people see things during model training that AI might not perceive. How can we leverage their insights? We gathered everything they’ve said publicly, later verified it, and used their judgments to predict AI’s future development. I call this “analyzing human experts’ analyses to predict human experts’ predictions.”

Currently, on our leaderboard, the highest prediction accuracy belongs to former OpenAI chief scientist Ilya Sutskever, and the lowest is Elon Musk.

Ranking of prediction accuracy for industry leaders

Future Person: What do you think distinguishes what you’re doing from Polymarket?

Henry: Our positioning is different. It provides a trading market, we are participants. I won’t evaluate its product, but what we do is socially meaningful: it can reduce a lot of unnecessary waste, improve the quality of decisions, and help people plan ahead. This is a very idealistic endeavor.

Future Person: But Polymarket’s prediction outcomes can also be used by commercial entities. It forms probabilities through mass betting, while you derive probabilities through model inference?

Henry: Yes. In that sense, it also has an altruistic side, achieved through the aggregation of human judgments. We may be more about guiding or leading the market. The appeal of prediction lies in the fact that the prediction itself changes the market.

Future Person: But that influence isn’t necessarily positive. For example, in financial markets, could a collective belief in a certain cycle actually bring about that cycle?

Henry：Yes, that can happen. You might, by predicting, prevent something you predicted from happening, or you might help it happen.

Future Person: So, what impact will prediction actually have on society? It may be hard to judge directly. You say it lets people act early and avoid some things, but also maybe because everyone believes something will happen, it ends up causing it?

Henry：Yes, it’s a game‑theoretic process. But simply, predictions can help companies or individuals make decisions, just like weather forecasts.

Future Person: If you can already make money through investing, why start a business?

Henry：I think money is just a byproduct. Investment and trading are said to allocate social resources better, but it’s still relatively indirect. Building tech products is more fun, you can directly feel that you contributed.

My ongoing motivation is still exploration. Forecasting is a long‑term, ten‑year challenge, and I hope to build a prophet, like the divine beast in “Zelda” that must be conquered—finding ways to acquire resources, carve a path, and get as close as possible to that goal.