When AI trading, decisions are free—the power to act is scarce
Market prices update every second, making decisions nearly free; what's truly expensive is the loss and responsibility each trade carries. AI trading's value isn't in predicting price moves accurately; it's in building a verifiable, auditable channel from decisions to actions.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Decisions are nearly free; real cost lies in responsible action
Wu Haozhe frames this year's most certain insight: in software development and similar fields, AI makes ‘doing’cheap but ‘deciding what to do’stays hard. Finance reverses this. Market prices update every second—each price represents someone's decision—and a research agent can issue hundreds of judgments daily across thousands of stocks, making decisions nearly free. What's truly scarce and expensive is action. To be responsible for one's trades, you must explain to yourself clearly *why* you're making them. That's why he repurposed the project from a quant research tool into a live trading dashboard: quant signals (IC, Sharpe ratio) are gibberish to retail investors—without a clear story, no one will spend money or take responsibility for them.
— Wu Hao ZheQuantitative trading locks responsibility into rule-writing, once
Traditional quant has humans lock decisions into rules upfront, then let machines execute them cheaply at scale. Win rate and strategy capacity are fixed at development time—so is responsibility. AI trading is different: the model continuously ingests unstructured material (news, filings, earnings calls, social media), so decisions surface continuously, not just at the start. You can't gate it once; you need constant auditing. In practice, good research must first separate facts in the material from the agent's assumptions, then test those assumptions (Is cash flow improvement from receivables collection or deferred payables? How are competitors moving?). What survives should be evidence-based with clear boundary conditions—if evidence falls short, stop at ‘unable to judge’, not at equivocation.
— Wu Hao ZheEarnings-surprise drift is being eliminated by AI
Before 2000, post-earnings drift was a classic anomaly: good earnings wouldn't move stock prices to their new level on day one; instead prices drifted upward over days or months because reading financials, tweaking models, and executing trades each took human time. Wu cites a paper titled ‘Rest in peace, earnings announcement drift’: this drift had already vanished from US indices by around 2006—surprise information is fully priced in on release day. He predicts AI will shrink this lag further, to seconds. That means being early and literate adds no alpha; excess returns migrate from single signals to combinations that require data fusion across sources.
— Wu Hao ZheIn the end, it all comes down to data alignment
Wu breaks an AI trading system into four layers: data alignment, a harness (auditing gate), the model (smarts and bias), and humans. A good system maximizes cheap decisions while minimizing actions—each action must have justification, clear bounds, and someone accountable. Returns and win rate aren't *causes*; they're the system's natural output. Poor systems either execute every cheap decision (overtrading) or become too cautious to act without clear reasons. He sees declining marginal value in model sophistication and proprietary data—anyone can query factors, anyone can run them. What's appreciating is data alignment and harness design. He views harness as the critical piece.
— Wu Hao ZheBuying data buys the right to test, not answers
Wu makes a blunt claim: more data doesn't equal alpha. Patterns discoverable on widely-available daily data and alternative data are already priced in; buying more data rarely uncovers new alpha. Buying data's real value is *making a thesis testable*. Often when you think an idea doesn't work, it's actually that data is incomplete or misaligned—not that the idea is wrong. He also cautions retail and small teams to know their limits: you can't outbid institutions on data breadth, but you can compete on the *quality* of data alignment and on ideas for combining disparate sources—that's how you find prices the market hasn't yet digested.
— Wu Hao ZheForecasting models deliver benefits roughly equivalent to zero
On the recent crop of forecasting-optimized models, Wu judges their help for trading at roughly zero. Three reasons: they optimize what's already cheap in finance (decisions), not what's actually scarce (converting decisions to actions credibly); finance's core facts—numbers, date comparisons, noisy or deliberately misleading text—are precisely where these models have public weaknesses; they output probabilities without reasoning, and once you're losing money, you may have lost a lot before the story breaks. He thinks these models work better as information gatekeepers (flagging rewrites, relevance) than for trade signals.
— Wu Hao ZheTrading is betting on probabilities, not specific outcomes
Trading's essence is betting on *probabilities*, not outcomes: a stock price is the market's weighted average of all possible futures, and profit comes from the gap between your probability estimate and the market's implicit one—position sizing follows from that gap, after costs. But Wu flags an overlooked trap: if a model says two claims each have 90% confidence (say, revenue beats *and* the beat drives price up), you can't simply multiply them to 81% joint confidence, because that silently assumes independence. In reality, many judgments chain on the same underlying logic, so if that chain breaks, all your 98% convictions fail together.
— Wu Hao ZhePerhaps AI doesn't need humans; humans need AI
Wu flips the question: everyone asks whether AI can replace humans and how many hours it saves—essentially pricing AI by the labor it displaces. But if models get strong enough to do many things better than humans, perhaps we should reverse the question: What can I still do that the world still needs? He predicts that even if AI could autonomously run trading end-to-end, we'd likely stay with humans in the loop. Not only because of oversight and auditing, but because humans derive direction and meaning through creating, deciding, and being responsible. We're not replaceable by design.
— Wu Hao ZheIn their own words · checked verbatim
Action is expensive—being responsible for your trades means you must explain to yourself why you're making them.
行动很贵 就是一个人要为自己的行动负责 他就必须要向自己去解释这个行动
Wu Hao Zhe3:02
AI trading isn't about letting the model make more decisions; it's about deciding which decisions deserve to become actions.
AI trading 更多的不是让模型去做出更多的决策 而决定哪些决策 更有资格去变成行动
Wu Hao Zhe6:04
My definition: AI trading is building a verifiable, auditable pathway between nearly-free decisions and expensive actions.
我的定义就是 AI trading是在几乎免费的决策 和代价昂贵的行动之间 去建立一条可被验证 可被审计的通道
Wu Hao Zhe6:04
A well-designed AI trading system should maximize cheap decisions while minimizing actions—each action needs justification, clear boundaries, and someone accountable.
一个做了好的AI trading的系统 应该是决策尽可能多 而且还便宜 但是行动尽量的少 每一个行动都要有它的来由 有它的边界 有人负责
Wu Hao Zhe20:15
More data really doesn't equal alpha, because patterns discoverable on widely-available daily and alternative data are already priced in.
数据多其实真的不等于有alpha 因为在大家都能拿到的日线和另类数据上 能够被找到的规律 其实已经被价格吃掉了
Wu Hao Zhe24:17
Forecasting models' help for trading? You can't say there's nothing, but it probably works out to roughly zero.
我觉得Jeff 对trading的帮助 就是也不能说没有吧 可能就是约等于零
Wu Hao Zhe27:17
The essence of trading is betting on probabilities, not outcomes.
交易的本质上 其实是对概率去下注 而不是去对答案下注
Wu Hao Zhe30:17
Figures
| GitHub stars (under six months open) | over 30,000 | 1:00 |
| Countries/regions with contributors | over 20 | 1:00 |
| Factors in early version | over 400 | 2:02 |
| Live trades following AI suggestions | 80–90% | 13:12 |
| Post-earnings drift disappearance in US indices | 2006 | 16:14 |
Glossary
- Harness
- The rules and processes that determine which AI-generated decisions qualify to execute as actual trades.
- Points in Time (PIT)
- Historical backtesting must use only data that was actually available at the time, never forward-looking corrections.
- Post-Earnings Announcement Drift
- The tendency for stock prices to drift upward over days or months following a positive earnings surprise, rather than repricing immediately.
- Vibe Trading
- An open-source AI trading tool that generates cross-disciplinary research teams from natural-language descriptions of investment theses.
How to listen
Practitioners in quantitative and AI-driven investing; individual investors seeking to grasp how AI trading differs from traditional quant approaches; and developers building AI agents who want to understand financial domain nuances.
Around 9:59, discussion of social media sentiment analysis covers familiar ground; skip if time is limited.