AI Trading — Decisions Are Cheap, Actions Are Expensive | A Conversation with Haozhe, Author of 30k+ Star Vibe-Trading

Summary

• Haozhe’s core judgment is the opposite of what holds in most industries: in finance, “decisions are almost free; what is truly expensive is acting.” Markets quote every second, and an agent can issue 100 calls a day on each of thousands of stocks; AI trading is neither an oracle predicting up or down nor a hands-off money machine, but “a verifiable, auditable channel between nearly free decisions and costly actions,” with the harness and the human deciding which decisions qualify for execution.
• The open-source project Web Trading went public on April 1 this year and surpassed 30k GitHub stars in under six months, with its success driven by a shift from a professional quant tool to an ordinary investor’s trading desk. Its predecessor, Quant Agent, had 400-plus factors and 9 market backtesting engines, but “quant is a foreign language to ordinary people”—if people are responsible for acting, they must be able to explain the action to themselves. The product therefore shifted to clear rationales around cash flow, catalysts, and risk, while catching a March-April window when there was no all-in-one financial desk agent on the market.
• His most market-relevant call is that AI adoption will compress information processing to seconds, potentially taking the edge out of “reading” and shifting alpha from single signals to cross-desk combinations. He recalls—perhaps from a 2022 or 2023 paper titled “RIP Post-Earnings Announcement Drift”—that post-earnings drift in some US indexes had apparently disappeared from 2006 onward; once everyone can run factors with one click, single-factor alpha may quickly die. Errors will also spread faster: someone apparently used AI to generate content claiming that an explosion had occurred near the Pentagon, briefly sending US stocks lower; research on multiple agents suggests they may independently produce collusion-like or insider-like behavior, while accountability and regulation “have no answer right now.”
• His competitive ranking is clear: the marginal value of model intelligence and proprietary data is falling, while data alignment and the harness are gaining value, with alignment the most important. “Having more data really does not equal having alpha”; the real purpose of buying data is simply to make a hypothesis testable. Alignment pitfalls include adjusted versus unadjusted prices, thousands of dollars versus dollars, whether “1M” means one month or one minute—a real bug fixed the day before yesterday—and fallback data sources whose amounts can differ by 6 or 7 orders of magnitude, while backtests must enforce PIT and never look ahead.
• He poured cold water on the recent JARVIS trading hype: the benefit may not be zero, but it is probably close to zero. JARVIS optimizes the decision side, which is already nearly free in finance; its supposed weaknesses in the official documentation—mathematical calculation and date comparison—are “exactly what finance looks like,” and it returns probabilities without reasons. A 50% output may simply be a coin toss and, at bottom, a hallucination; 90% × 90% = 81% quietly assumes independence even though financial judgments often share the same potentially wrong news. Its proper role is “the sorter at the door, not the courier.”
• In practice, he follows AI’s recommendations 80% to 90% of the time: storage stocks ran hot in July, then sold off across the sector for nearly half a month; as the market debated whether Chinese producers’ capacity expansion would depress storage prices, the agent distinguished the impact on memory, flash, and hard disk drives and bought Seagate before earnings. He still rejects full automation because of “accountability”: someone must own every entry and every loss. He closes on a philosophical question—“Could it be that people need AI to demonstrate their own value?”—and expects human-in-the-loop systems to endure.
• His first lesson for beginners is to isolate the risk and open a separate account: “Don’t treat it as a get-rich code.” Any method, including sentiment as a contrarian signal, will be countered once it becomes mainstream—“it’s like playing a game while your opponents and the version environment keep changing.” A better use is to have AI help explore markets one has never tried but that fit one’s own strengths, whether commodities, fundamental stock selection, or technical charts.

Deep dive

1. The Web Trading pivot: Quant is a foreign language to ordinary investors

• Haozhe’s career is an AI-trading timeline in miniature: a mathematics graduate of Wuhan University and a PhD student in HKU DS, he has personally worked through “several generations,” from factor mining and deep learning to having large models read text and building agents. The project went open source on April 1, surpassed 30k stars in under six months, and drew contributors from more than 20 countries and regions.
• Its predecessor, Quant Agent, was built for quant researchers, with 400-plus factors, 9 market backtesting engines, and a range of statistical tests. The team later found that demand was moving in the opposite direction and rebuilt it as a secondary-market trading desk spanning macro, fundamental, quant, and risk review. “Quant is a foreign language to ordinary people”; IC and the Sharpe ratio cannot explain why someone should buy a particular stock.
• The underlying principle is simple: “Actions are expensive. If someone is responsible for an action, they have to explain it to themselves. They won’t pay for or take responsibility for a quant signal they can’t understand.”
• He attributes the breakout to timing: in March and April this year, there was no all-in-one financial desk agent on the market, while professionals could use the platform to turn fundamental experience directly into automatically executable quant strategies.

2. Defining AI trading: in finance, decisions are free; actions are expensive

• The dynamic is the opposite of other industries. In software development, AI makes doing cheaper while deciding what to do remains difficult; in finance, “decisions have never been scarce.” Markets quote every second, and an agent can issue 100 calls a day on each of thousands of stocks. “What is truly expensive is acting.”
• First, what it is not: it is not “AI predicting whether the market will be up or down tomorrow, like a god issuing an oracle,” nor an automated order-entry system that lets people make money while doing nothing. His definition is a verifiable, auditable channel between nearly free decisions and costly actions.
• The roles are divided three ways: the model generates decisions, cheaply and at a cost that will only fall; the harness determines which decisions qualify to become actions; and the human sets the rules and makes the call on anything outside them. That last layer is the most expensive.
• This is where AI trading diverges from traditional quant. Quant starts with humans fixing decisions into rules, after which machines execute them cheaply; responsibility is settled at the moment the rules are written. AI trading reads unstructured news, announcements, earnings calls, and social media, with decisions constantly emerging, so it needs “something like a gate” that keeps screening them.

3. Research methodology: separate facts from assumptions and stop at “cannot determine”

• Take a cash-flow question. AI first organizes the financial-report disclosures, lists the source and date for every number, separates facts in the materials from the agent’s assumptions, and then tests the thesis: did the improvement come from collecting receivables or delaying payments? What happened to peers? Is there contrary evidence?
• The key discipline is that when the evidence is insufficient, the system “must stop at cannot determine, rather than give you an ambiguous result.” Whether to place the order remains a human decision.
• Asked where human know-how fits into cross-validation, Haozhe put it bluntly: AI can judge all industry know-how because that knowledge has already been distilled through model iterations. What humans must manage is risk—how much each person can tolerate—which AI does not know.
• Why not simply ask GPT? Codex and Claude can be black boxes, can hallucinate, and may produce uncertain results when rerun. A purpose-built tool can “fully control what the model is allowed to generate and what must be done through correct calculations,” making the process auditable, verifiable, and repeatable.

4. Sentiment contrarianism works, but every edge gets arbitraged away

• Scraping retail sentiment from Twitter, Reddit, and Xiaohongshu “can reasonably be called effective,” because forum sentiment is often a contrarian indicator. But once everyone does it, institutions can deploy bots to manufacture false information and counter the signal. “At its core, this is a zero-sum game: what you earn is what someone else loses.”
• His broader worldview is definitive: “There is never a permanently unchanged way to make money.” It is like playing a game while the opponents and the version environment keep changing; the method, risk appetite, and thesis all have to keep changing too.

5. Live trading: following AI 80% to 90% of the time, the Seagate case, and why he refuses full automation

• He trades with his own tool and “follows AI’s recommendations 80% to 90% of the time,” limiting his role to checking whether the reasoning chain is sound.
• The case came this July. The storage sector ran hot and then fell broadly for nearly half a month. As the market debated whether Chinese producers’ capacity expansion would depress prices across the storage industry, the agent asked whether memory, flash, and mechanical hard drives would face the same shock. They clearly would not, but the market was pricing the sector as a whole, creating a potential underpricing. It selected Seagate, whose business is concentrated in mechanical hard drives, compared its overall post-earnings move with historical samples, and reached a cautious conclusion. The human-machine process led to a purchase before earnings.
• Why not let AI operate autonomously? “More than anything, it is a question of accountability.” Even in highly automated quant trading, “you still need to take responsibility for every position you open and every loss you incur.” He does not expect AI trading to become a purely scalable system in which humans are completely uninvolved.

6. When everyone has an agent: the death of post-earnings drift and the rapid spread of errors

• The first impact is faster price discovery. Before 2000, the famous “post-earnings announcement drift” meant that a stock would continue rising for days to months after a strong report because reading the filing, updating the model, holding meetings, and placing the trade all required human work. He recalls a 2022 or 2023 paper titled “RIP Post-Earnings Announcement Drift” arguing that the drift in some US indexes had apparently disappeared from 2006 onward. His view is that once AI is widespread, any information could be digested within seconds, and “being the first to read or understand something will no longer be worth as much.”
• The second impact is that excess returns move from single signals to combinations of conditions. Once everyone can run factors with one click, single-factor alpha may disappear quickly. The real value will lie in obtaining, aligning, and combining data across desks at the same time.
• The third impact is faster error propagation. Someone apparently used AI to generate content claiming that an explosion had occurred near the Pentagon, briefly sending US stocks lower. Haozhe and his team have studied multiple agents trading in the same environment, and some results suggest that they may independently produce similar collusion-like or insider-like trades. Whether responsibility lies with the model provider, the user, or the exchange “is a question that has no answer right now.”

7. Ranking the four elements: alignment and the harness gain value, while intelligence and proprietary data lose it

• A good system generates “as many decisions as possible, as cheaply as possible, and as few actions as possible,” with every action grounded in a rationale, bounded by rules, and owned by someone. Returns and win rates “are not the reason; they are more like the result.” Bad systems take two forms: turning every cheap decision into an action and overtrading, or becoming so cautious that they place no trades while being unable to explain why.
• The division of labor is clear: model intelligence determines what the system can think of; proprietary data determines whether others have thought of it; data alignment determines whether the idea is actually valid; and the harness determines whether a wrong idea can be stopped. The marginal value of the first 2 is falling, while that of the latter 2 is rising, with data alignment the most important.
• The real alignment traps include whether prices are adjusted and whether the unit is thousands of dollars or dollars. “Does 1M mean one month or one minute? This is not a joke; it is a real bug we fixed the day before yesterday.” When a small exchange’s data source went down and the system fell back to a crypto data source denominated in US dollars, the amounts differed by 6 or 7 orders of magnitude.
• There is also PIT, or point-in-time integrity. If a 2020 stock-picking exercise uses a number the company did not correct until later, “it is equivalent to looking at the answer in advance, so the backtest will definitely look good.” When data is missing, the agent should proactively pause its report rather than silently continue running the backtest.

8. Demystifying proprietary data and positioning JARVIS as a “sorter”

• Qu Kai pushed back: if institutions can legally gain early access to investor-relations teams and earnings calls, can exclusive information really have no major impact? Haozhe was unequivocal: “Having more data really does not equal having alpha.” In daily and alternative data available to everyone, the patterns that can be found have generally already been absorbed into prices. The real purpose of buying data is “to make your hypothesis testable”; many ideas fail simply because the data is insufficient or the definitions are not aligned. Retail investors cannot match institutions on data breadth, but they can compete on alignment quality and combinations of ideas.
• On using the recently hyped JARVIS for trading, his answer was blunt: “You can’t say it has no benefit, but it may be approximately zero.” JARVIS makes repeated calls to large models cheap and fast, but it optimizes the side that is already cheap in finance—“it makes something that was not very valuable even cheaper.” Its supposed weaknesses in the official documentation are mathematical calculation and date comparison, while “the things it is not good at happen to be what finance is actually like.”
• Returning probabilities without reasons is a fatal weakness. Without an analytical chain, “by the time you realize what is happening, it may already have lost a lot of money.” Its proper role is initial screening: whether a news item concerns the company, whether a document is an original filing or a repost, and whether the material is relevant. These tasks can run tens or hundreds of thousands of times a day with limited consequences if wrong. It is “better suited to being the sorter at the door, not the courier entrusted with delivering the entire chain.”
• The deeper issue in forecasting is that trading is “betting on probabilities, not betting on answers.” The payoff is the difference between your probability and the market-implied probability, less costs. A 50% output may simply be a coin toss: “Isn’t that also essentially a hallucination? Then you should not act.” Multiplying two 90% probabilities to get 81% quietly assumes independence, even though multiple judgments often rely on the same news; if the news is wrong, they are all wrong.

9. AI for finance ≠ AI for trading: layered gains and a philosophical close

• Haozhe uses Codex and Claude daily, and says that for capabilities related to discretionary investing, “they already do a very good job.” Discretionary investing is fundamentally Deep Research, a question that already had a clear answer last year, and AI can indeed do it better than people. AI for trading asks whether the market can be beaten consistently. AI for finance—such as SLP’s financial template open-sourced in May for valuation checks and reconciliation, or the similar product Kimi launched in September—is about connecting to institutional data, understanding how a team works, and producing work that colleagues can take over. “Model progress is no longer that important.”
• The gains are layered. People scoring 20-30 points gain basic risk control and stop believing every anonymous stock tip or Douyin recommendation. Those at 50-60 points can conduct deeper fundamental and macro research. Those at 80-90 points can validate ideas across multiple dimensions: macro researchers may not know how to code a backtest, while quant researchers with technical backgrounds may not understand the economics behind prices. As for belief disputes such as “baijiu will never disappear,” he classifies them as “emotional companionship”: people talk to themselves, and AI responds with logic.
• The philosophical closing is: “Could it be that people need AI to demonstrate their own value?” Even if AI eventually no longer needs humans to review and supervise it, people may still need to participate to retain a sense of direction and meaning. “Human decision rights cannot disappear as machine capabilities rise.”
• His 0-to-1 advice is to define risk tolerance first and isolate funds in a separate account. “Don’t treat it as a get-rich code,” because countermeasures will inevitably emerge. Then use AI to explore markets one has never tried: someone may have a natural feel for commodity logic, someone else for the fundamentals of thousands of stocks, and someone else for technical charts.