Raising lobsters and trading stocks: Is it "science" or "mysticism"?

Interface News Reporter | Liu Litong

Interface News Editor | Song Yejun

Recently, “Lobster Trading” (deploying, training, and using open-source AI agents OpenClaw) has gone viral across the internet, attracting many investors to join the trend.

Interface News has noticed that discussions about “using lobsters to trade stocks” on social media are becoming increasingly heated. Some praise the efficiency and convenience of “lobsters” being able to monitor stocks 24/7 with AI, while others lament that “using lobsters to trade stocks costs tokens that are more than ten times higher than trading fees.” Some are seeking information about OpenClaw everywhere, and others are questioning the safety and reliability of “lobster trading.”

Since DeepSeek became popular last year, more and more A-share investors have begun to explore AI in various ways, but their actual experiences vary widely.

Investor Chen Xue (pseudonym) has sought the “secret to wealth” across multiple AI large model platforms but suffered an overall loss of nearly 20% during the bull market. She said, “A sincere effort still ended up being a misinvestment.”

Qin Peng (pseudonym), head of a quantitative team in South China, views AI “partners” as “investment research tools,” significantly boosting work efficiency.

How is AI performing in stock trading scenarios?

Efficient, but not necessarily reliable

For Guangdong speculator He Feng (pseudonym), the first instinct when encountering problems is to “check Baobao.”

Whether it’s breaking news or new thematic concepts, he can usually get a preliminary answer within 1-2 minutes. If deeper research is needed, he adjusts keywords and questions, and within a few minutes, he can get a more satisfactory answer.

Before the advent of large AI models, He Feng would spend a lot of time browsing news sites, stock forums, and social media to gather information. After collecting enough data, he would need to analyze and synthesize it himself to reach a somewhat acceptable conclusion.

Qin Peng prefers to combine his quantitative stock selection model with large AI models. His model automatically screens stocks daily based on fund flow, market heat, price-volume trends, etc., then performs secondary filtering based on fundamentals and hot topics to identify final targets. With AI assistance, the time spent on manual filtering has shrunk from 3-5 hours to 30-50 minutes, greatly improving efficiency.

Additionally, Qin Peng occasionally delegates simple tasks like writing or modifying stock selection models to AI.

“Efficiency” is the first word many investors think of when discussing AI stock trading. With over 5,000 listed companies in China and continuous 24-hour financial information updates, extracting the needed data from this vast sea of information is beyond any individual investor’s capacity. For AI, this is a “small dish.”

However, many interviewees also agree that AI large models often produce unreliable answers.

For example, asking an AI model about the relationship between a stock and a hot topic usually yields a seemingly logical answer, but much of the content lacks factual basis.

Some investors give an example: asking AI to find the 10 stocks with the lowest PE ratios in the market. It might only analyze data from dozens of stocks and give an answer, some of which could be outdated or even incorrect.

AI “partners” also often display a “people-pleasing personality.”

For instance, if you ask, “Is A better than B?” it will list many supporting points. But if you reverse the question to “Is B better than A?” it will do the same. If you first ask it to analyze a certain industry, then inquire about sectors worth watching, the previously mentioned industries often appear again.

Almost all interviewees have experienced “AI hallucinations”—where AI’s answers seem reasonable and comprehensive but are actually fabricated, containing false facts, data, or events, and sometimes contradict basic common sense. This is a serious problem in investment, as any decision mistake can lead to real financial losses. These phenomena cause additional issues: although investors can get quick answers from AI, they often spend multiple times longer correcting or adjusting the AI’s responses, trying to get more reliable information.

Where do the problems lie?

Chen Xue was initially motivated to try AI stock trading after learning that DeepSeek’s underlying quantum algorithms are impressive.

Many top quantitative funds publicly state they are deploying AI, but few outsiders truly understand how AI influences their investment decisions or how much it impacts their returns—especially how much high-frequency trading contributes.

A senior person from a leading Shanghai quantitative fund believes that asking AI casually during stock trading is very different from fully applying AI in quantitative investment.

Quantitative investing generally uses mathematical models, statistical methods, and computer programs to replace subjective judgment, characterized by discipline, data-driven decisions, diversified holdings, and strict risk control.

For most ordinary investors using AI models, the ultimate decision still rests with humans, making it a form of subjective investing. Their holdings are usually limited, making it difficult to diversify enough to hedge risks caused by AI errors.

Furthermore, many investors are accustomed to using general large models like Baobao, Qianwen, DeepSeek, etc., which differ fundamentally from proprietary AI models developed by quantitative funds.

According to Interface News, quantitative funds mainly focus on three AI elements: data, computing power, and algorithms.

An industry insider told Interface News that high-quality, real-time, complete data is crucial for training AI in finance. General large models are mostly trained on text data, lacking sufficient high-quality financial data.

In terms of computing power, while general large models may have higher hardware investment overall, their broader scope and larger training volume mean more resource consumption.

On the algorithm front, leading quantitative funds generally adopt a “self-developed” approach. Their core algorithms are often aligned with general large models but are fine-tuned differently. These core algorithms are usually top secrets and rarely disclosed.

The industry insider also mentioned that some brokerages are actively developing vertical AI models for finance. Although these institutions focus more on finance and have access to the latest financial data, their AI research is limited by computing costs and regulatory constraints, making it difficult to fully meet investors’ expectations.

“Even if their AI models are far from those of quantitative funds, general large models incorporate a wealth of investment knowledge. Why can’t they provide more reasonable investment advice like seasoned investors?” many investors, including Chen Xue, have wondered.

In response, Chengdu speculator Ren Yu told Interface News, “Subjective investors may not require the same level of data precision as quantitative investors, but their decisions still need to be based on the latest, relatively accurate data. General large models often fetch data that isn’t timely and may include some contaminated information, making their analysis unreliable.”

“The key issue is that AI large models lack a complete investment system. Every strategy has its characteristics and suitable market environment. From different strategic perspectives, conclusions about buy or sell points can vary greatly. For example, a stock might be a good buy from a medium- to long-term perspective but a sell from a very short-term view. AI models have learned many strategies but lack practical data supporting these strategies, making it hard to understand their underlying logic and differences,” Ren Yu explained.

Would feeding AI with the investment frameworks and philosophies of top investors improve its recommendations? Qin Peng has tried this but opposed the idea. He believes that the “input” consists only of publicly available views and logic, which top investors may not fully or openly share. Moreover, their investment systems evolve with market conditions.

Even if AI could give more reasonable investment advice, would investors strictly follow its strategies? The answer is probably no.

Human-AI collaboration is the consensus

How should ordinary investors use AI “partners” that are efficient but potentially unreliable?

“Relying solely on AI for investment decisions is impossible; you still need to build your own investment system,” Chen Xue concluded after over a year of experience.

Recently, she paused real trading to focus on learning more investment knowledge. Once she feels more confident, she plans to restart her trading. During this process, she discovered a new advantage of AI models: “Their text analysis ability is really impressive—searching and summarizing investment knowledge is fantastic!”

Qin Peng, who is more satisfied with AI “partners,” shared his experience with Interface News. He said that during the information collection phase, AI’s efficiency far surpasses humans, so this part can be delegated more. In the analysis phase, AI is also more efficient but prone to errors, so adjusting questions and prompts can help improve its reasoning. The decision-making stage is more complex and critical, requiring human judgment.

Most interviewees believe that AI models will become increasingly useful and that many specialized AI tools for finance will emerge. However, AI will not completely replace human decision-making but will serve as an auxiliary tool. Human-AI collaboration will remain the main trend.

On one hand, AI models are trained by humans; the amount of computing power invested, the data fed into AI, and the algorithms used are all decided by humans. For the foreseeable future, AI cannot operate independently without human oversight.

AI generally finds patterns in historical data, but stock markets never repeat exactly. “Black swan” events can happen at any time. AI models lack the innate ability to handle such scenarios, making it unlikely to produce a universal “agent.”

On the other hand, from a technical perspective, AI trading might someday outperform humans. But a series of potential risks makes it unlikely that we will hand over all decision-making authority to AI.

For example, strategy convergence is a long-standing concern. As more institutions and investors use similar data and methods to train AI, strategies tend to become more similar, increasing the risk of herd behavior and systemic market fluctuations.

Additionally, AI models often operate as “black boxes,” making their decision processes difficult to trace. In case of anomalies, the models cannot be held responsible, and it’s hard to determine whether human factors influenced the outcomes. If decision rights are fully delegated to AI, some groups could manipulate or influence AI to control the market more covertly. From a regulatory standpoint, to mitigate such risks, the application of AI in finance will likely be limited within certain boundaries.