When AI learns to make predictions, it finds that Elon Musk's predictions are unreliable.

AI prediction is a new species that sits between quantitative investment and subjective investment research.

Text | Liu Sijie

Editor | Zhang Wei

Cover Source | Pixabay

In 2025, 25-year-old Henry took a heavy position in Nebius, a 10-bagger stock. Relying on the returns from this investment, he did not need to work for a very long period of time.

Nebius is a company that develops AI neocloud. It was not a company widely favored by the market at that time. Some researchers who took a wait-and-see attitude toward the company judged its value solely by its quarterly profits and defined its business model as a "computing power sublessor", but Henry thought, "These judgments are rather short-sighted, because it is extremely difficult to accurately predict every single figure."

Henry has a different perspective. His core logic for being bullish on Nebius focuses on the company's long-term, fundamental driving forces, namely its personnel and organizational structure.

Compared with his impressive investment track record, what we want to highlight more is a more core and interesting perspective demonstrated by this young man in the whole process —

"Stock selection and prediction are essentially the same thing," he said. "Understanding and processing information from different dimensions can help us make better predictions."

This perspective is exactly the cornerstone of his entrepreneurship. After completing his undergraduate study in Computer Science at Shanghai Jiao Tong University, he pursued a postgraduate degree in FinTech at Shanghai Advanced Institute of Finance. Having worked in quantitative finance and subjective investment research successively, Henry founded the AI prediction company ProphetLab in May this year, integrating AI prediction with the experience of human experts.

"Exclusive news from Future Human Lab shows that the company has obtained tens of millions of yuan in angel round financing from multiple investors including Ivy Capital."

Before officially starting his business, Henry had a "dormant period" of more than half a year. He lived in a rented apartment in Shanghai, studying the integration of AI and investment all by himself. It was during this period that he spotted Nebius, whose stock price later surged more than 10 times.

During that six months, Henry gradually formed a judgment: the integration of AI and investment is not just about automating existing quantitative methods, but may become a "new species" between quantitative investment and subjective investment research.

In his view, quantitative investment excels at processing structured data and building models, while subjective investment relies on human experience, intuition, industrial insight and a large amount of unstructured information that is hard to encode. In the past, there was a huge gap between quantitative and subjective investment research, and it was difficult for talents and methodologies to circulate between the two systems. The emergence of large language models for the first time makes it possible to "build a bridge": it can understand human analysis processes, abstract the subjective investors' judgments on industries, companies and trends, as well as the originally unstructured unplanned information that is hard to formalize, and then combine it with the data processing capabilities of quantitative investment.

AI prediction companies are exactly this new species. His past experiences have repeatedly verified that making accurate predictions follows traceable rules.

He wants to endow AI with this capability.

AI outperforms humans in both aspects of collecting sufficient information and having strong enough information processing capabilities. After all, large models are getting increasingly proficient at searching, analyzing, using tools and completing tasks. Coupled with the experience of human experts, "it is possible to predict things that others cannot predict", Henry explained.

For example, when predicting Tesla's sales volume, traditional researchers may only review a small part of the comments online, while AI can access far more reviews, search for map data and other relevant information, so it is very likely to deliver more accurate predictions.

The first product of ProphetLab is World Cup prediction.

At the initial stage, they only ranked the prediction results from various mainstream models. After the financing was in place, the team began to add self-developed prediction frameworks: calling large models through APIs, combining data such as historical matches, player injuries, referee styles and weather, simulating consecutive events that happen in the match, then counting the probability distribution of scores through multiple simulations, and outputting the final prediction result.

According to Henry, this system has achieved a prediction accuracy rate of over 70%, and successfully predicted the top 4 teams, the champion of this World Cup, as well as the 1:0 score of the final.

At present, ProphetLab is promoting cooperation with fund institutions to develop a benchmark for financial prediction. They try to test whether the model can make well-founded judgments on events that have not yet happened through questions requiring multi-step analysis such as corporate sales and revenue. Financial experts participate in question design and model evaluation, and the accumulated data and feedback from evaluation will be used for subsequent model training.

They also worked on an interesting small project to analyze the prediction performance of figures in the AI industry. They collected all their past public statements, verified each one to see if the prediction came true, and identified who tends to make empty promises and who speaks more rigorously. In this way, when making judgments next time, people will know whose opinions are more worthy of reference. Henry calls this "analyzing the analysis of human experts, and predicting the predictions of human experts".

At present, Ilya Sutskever, former Chief Scientist of OpenAI, ranks first in prediction accuracy, while Elon Musk ranks last.

He believes in the charm and social value of prediction. If people can predict the future more accurately, they can make correct decisions and reduce resource waste.

However, going back to his outstanding investment track record, we asked him: you have already been able to make money with your investment capabilities, why do you still choose to start a business?

Henry mentioned his favorite game, *The Legend of Zelda*. "Starting a business is like playing *The Legend of Zelda*. There is a divine beast that you must defeat. You need to find ways to acquire resources, carve out a path, and get as close to that goal as possible."

Exploration is the driving force that keeps him investing continuously. "Prediction is a 10-year long-term project, and I hope to build a prophet."

Facing an event that has not yet happened, how can AI identify truly important signals from massive information and noises, and form more accurate judgments than humans? And when people start to believe in a certain judgment about the future, will they take actions that change the future itself?

The following is the conversation between "Future Human Lab" (hereinafter referred to as "Future Human") and Henry.

Henry

 How does AI predict something that has not happened yet?

Future Human: Why did you start prediction with the World Cup? Do you love football very much yourself?

Henry: I don't play football, and I seldom watch football matches. At that time, some people started talking about buying World Cup tickets, and I thought, as a prediction company, can we take advantage of this event hotspot to promote our brand? We also had recruitment needs at that time.

At the very beginning, it was mainly me and two AI PhD students working on this project, and I invested some money myself. Later, we posted prediction results on social platforms every day, which did attract some people. Now there is a team member with a mathematics background from the University of Cambridge. He thought the prediction results were very interesting after seeing them, traveled from Hangzhou to Shanghai several times to talk to me about these ideas, and later joined our team.

Future Human: From the official website I visited, it seems that the World Cup prediction mainly displays the prediction results of multiple models. Which part did you develop on your own?

Henry: When we first started preparing for the World Cup prediction, the first match was very close, and we had not received financing yet, so resources were very limited. At that time, we just wanted to make a simple ranking list, let each model output its prediction results, and rank them after each match.

Later, with more resources and help from friends and classmates, we began to develop our own set of prediction models. The "model" we mentioned here refers to algorithms. We built a set of Harness, which is the framework that organizes the work of models and calls APIs. The whole match simulation process you saw is implemented by calling APIs, and the core lies in the framework and raw data.

Future Human: How does this framework predict a specific match in detail?

Henry: First of all, we collect a lot of information related to football matches, and purchase data including player injury status, home and away status of the match, referee style, weather and humidity, etc. These raw information will be imported into our framework and mapped to the simulator.

The simulator will make very fine-grained predictions, for example, an event occurs every minute or every few minutes: cross from the wing, headed out by the defender, turn and shoot. Combining all these events together forms a complete match.

We will repeat the simulation of the match many times, some end with 1:0, some end with 2:1, and finally we get the probability distribution. For example, if we simulate 30 matches and 5 of them end with 1:0, we will consider that the probability of this score is 1/6. Adding up the probabilities of all winning scores gives the probability of winning the match.

Future Human: Are these actions simulated, or did they happen in past matches?

Henry: They are all simulated. But we will use information from past matches. For example, in football matches over the past decade, what event was triggered every minute. After the Agent reads this historical information, it will generate corresponding behaviors in the simulation.

We also introduced a human expert. Because I seldom watch football, I found that AI made a lot of mistakes, so I invited a referee from the Chinese Football Association to help us analyze the fallacies in the multi-step reasoning of football prediction. He marks out where the mistakes are and what the correct logic should be, and we input these contents into our memory module.

When each match ends and a new match starts, the system will read these records, remind itself not to make the same mistakes again, and then make subsequent predictions.

Future Human: A real-life referee? What exactly does he correct?

Henry: For example, when simulating the match between Japan and the Netherlands, the model often generated plots where Japanese players jump high and score with a header. The referee thought this situation should not be so common, and his judgment was that the height of players from both sides and the header confrontation situation were not reasonably considered.

We recorded this information and imported it into the next prediction, so the model will no longer easily make the same judgment.

Future Human: How do you assign weights to different information when making predictions?

Henry: These are judged by AI. For example, in investment research, financial reports and corporate announcements have the highest weight and confidence, followed by research reports from analysts of high-quality institutions, then public research reports and news, and the lowest are rumors on social platforms.

We will check whether the confidence level is designed properly and whether AI makes good judgments, which is the role of human experts' experience. We will not directly modify specific judgments, but we will evaluate whether the Agent performs well.

Future Human: Have there been any impressive prediction errors in the World Cup prediction?

Henry: Of course there are. There were a few days during the World Cup when we made many mistakes, and people left comments saying that none of the predictions were right and we were just guessing randomly. There were also times when we got all predictions right for three consecutive days, and the comment area was full of praise calling us "godlike".

Football itself has randomness, and there will be come-from-behind victories. Even if the process is very accurate, the prediction may still be wrong. When we say our predictions are accurate, we just hope to be more accurate than the market average.

Future Human: So you only care about the errors in the reasoning process, not the correctness of the final result?

Henry: The result will also be stored in the memory module. The process will get a score, and the result will also get a score. But we will not punish the model directly just because it guessed wrong in one match. Like the example of Japan vs Netherlands I mentioned earlier, if there is a problem in the reasoning process, we will correct that part.

Future Human: If new information emerges after the prediction is made, how do you update the model?

Henry: We update it continuously through the memory module. For example, we make a prediction at time T, and at time T+1 there is new raw data and new information. We will use the updated system to make a new prediction through the memory system.

Future Human: Do you use this match simulation method for other events as well?

Henry: We developed a dedicated architecture specifically for football, and for other events we use a general framework that adopts the Deep Research approach. The model will conduct analysis and finally generate a report that includes information from human analysts and the content analyzed by AI itself, and then summarize all the content.

Can AI "judge the future" as well as excellent researchers?

Future Human: What capabilities are universal between World Cup prediction and commercial/financial prediction?

Henry: The raw data collection method, as well as the work of Deep Research and multi-step reasoning are universal. What is not universal is the data itself and the experts that need to be matched for each task. For football, the expert is a football referee, and for finance, the experts are investment researchers, fund managers and other relevant professionals.

Our company is currently focusing on the commercial and financial fields, and we give lower priority to politics and sports.

Future Human: What exactly does the financial benchmark you plan to develop in cooperation with fund companies test?

Henry: In financial evaluation, one type of questions is knowledge-based, such as whether the model has memorized financial formulas; the second type of questions tests the model's practical ability, such as the interlocking logical relationship between the three financial statements, and the ability to call Excel and other tools.