12 times backtesting curve 94% net profit comes from 5 trades, Uweb eight students online dismantle the verification path of AI investment research
TECHUB NEWS Reporter Alma Li in Hong Kong reports｜August 30, 2026
Key points: On the evening of August 29, Hong Kong's Web3 educational institution Uweb (University of Web3) held an online sharing session for students, with eight students taking turns to present on AI investment research. A backtesting report with a cumulative yield of about 12 times was dismantled on site, with 94% of the net profit coming from 5 trades; the sharers on options, precious metals, and on-chain arbitrage unanimously focused on "verification" rather than "prediction." The consensus of the evening was: The role of AI is not to provide answers but to execute discipline and audit evidence.
Bitcoin is experiencing a surge. Public market data shows that since mid-August, Bitcoin has risen from about $62,800 to over $81,000, with an increase of around 21% in the past month, fluctuating around the $80,000 mark on August 28. With the market heating up, discussions about tools have also intensified. On the evening of August 29 (Saturday), Uweb held an internal student sharing session, where the host opened with a question: During this week of intense market fluctuations, has anyone conducted in-depth research using their AI tools? For nearly three hours, eight students took turns speaking, offering varying answers but surprisingly consistent in direction: how to use AI for investment research and how not to be deceived by AI (and oneself).
Five Gates: On-site Dismantling of a 12 Times Curve
Ryan Yue, a sharer from the quantitative direction, was the one with the highest data density that evening. He first showcased a backtesting report run on TradingView: 52 trades, with a cumulative yield of about 12 times. The curve looked good, but his next move was to dismantle it on site.
According to his dismantling, the first 5 trades contributed to 94% of the total net profit; excluding these 5 "big winners," the yield shrank from 12 times to 65%. Out of the 52 trades, 31 were profitable, with a win rate of 59% and a profit-loss ratio of 2.02; the average single trade return was 6.7%, with the median only 4.3%, a figure he stated he trusts more. After accounting for fees and slippage, the discrepancy reached hundreds of thousands of dollars. He also analyzed the correlation between holding period and final returns, with a correlation coefficient of only 0.44. The intuition that "the longer you hold, the more you earn" does not stand firm in this data.
"Profit concentration is a warning signal to be wary of." Ryan Yue mentioned that concentrated profits may simply coincide with a specific sector's market movements rather than being attributed to the strategy itself. He set five gates for backtesting: data integrity, strategy logic verifiability, execution authenticity, profit distribution, and robustness verification. For execution authenticity, he provided an example: signals triggered by daily closing prices may not necessarily transact at assumed prices the next day; the confirmation time for signals and execution rules must be pre-agreed rather than retroactively recognized. The strategy must also be tested separately in rising, fluctuating, and declining market conditions before conducting out-of-sample verification; otherwise, in his view, it can only be regarded as a "hypothetical research."
As for the role of AI within this, his definition contrasts with most people's imaginations: it is not about generating more attractive curves but about acting as an auditor, producing "downgraded conclusions" and "evidence gaps," while humans are responsible for defining rules and judging boundaries. "One cannot judge the quality of a strategy solely by its returns." He also reminded that many precise entry and exit points seen on social media hide future functions.
Equipping Discipline with Exoskeleton
The second sharer, Lin Jiarui, diagnosed the common investor: losing money is not due to stupidity but a triple misalignment. In his words, most people ask, "Can I buy at this price?" whereas a complete investment has at least four dimensions (selecting targets, judging structure, determining entry and exit points, and managing positions); "using one dimension to make four-dimensional decisions inevitably leads to losses." The other two misalignments are emotion and energy: relying on willpower to combat greed, anger, and ignorance is unfeasible, and limited energy cannot keep track of a 7×24 market.
The solution is to systematize feelings: expanding the analysis cycle to weekly and monthly charts, entering and exiting in batches, allocating dedicated funds for accounts, and then allowing AI to execute discipline. "True AI does not give you answers; it executes the system for you." Lin Jiarui emphasized that AI lacks greed, anger, and ignorance. When the market crashes at 3 AM and people are scared and shaking, it simply reports "reached the entry range" per preset rules; its greatest value is "not being smarter than humans but being steadier." He offered a three-step path for ordinary people: first, learn cognition before using tools (AI amplifies existing cognition; if the rules are wrong, it will help people efficiently worsen the situation); use a sum of money that does not affect one's life to run the entire process; let AI be the co-pilot, gradually granting authority, keeping decision-making and ordering in human hands. He revealed that he has been running this system for two to three years, with an annualized return of about 30%.
The first student to take the stage that evening, Jingming, turned the same topic into a product. His developed "Investment Workbench" has over ten long and short positions, enabling multiple rounds of debate on the same asset, providing bullish or bearish conclusions, as well as investment or non-investment verdicts; then an investment radar operating weekly filters targets based on signal strength across four categories (clear buy, candidate, observation, risk alert), fed back to the analytical terminal for secondary verification; finally, layering in a set of trading rules he has formulated. If the first two layers are bullish but the rules determine that buying is not advisable, the system directly gives a risk warning: "The trading of the asset has violated discipline." From filtering, in-depth analysis to final decision-making, these three layers form a closed loop, which he calls using tools to constrain emotional trading. The development path was also publicly shared: first allowing AI to form a "virtual team" with professional skills to produce requirement documents, then handing it to Claude Code for development, and finally running a subscription membership embedded version directly in a production environment.
Options are a Price Quotation for Expectations
Tracy Cui shifted to a market dimension: options. Her question also did not concern predictions but rather "What risks has the market paid for in advance before major events like earnings reports and FDA approvals?"
She understands options as "a price quotation for future volatility": if call options become more expensive, it shows the market values upward potential; if put options become more expensive, it indicates that downward risk is being considered; an increase in implied volatility represents a growing expectation of volatility without pointing to any specific direction. She broke the signals down into four dimensions: direction, magnitude, time, and crowding. Taking crowding as an example, if the options positions are concentrated around a certain strike price, that area is likely to be a key support or resistance level. Ryan Yue immediately posed a question about this, and she provided a simplified judgment: look at where the trading volume is concentrated.
When the signals in four dimensions converge, they produce what she calls the "pre-event expectation radar," outputting four fixed fields: direction, magnitude, price, boundary. "It will not forcefully compress research into a fixed conclusion of buying or selling." After the event, analyzing whether the actual volatility falls within or outside the expected range is the key focus of research. She provided an example: if the option price implies a volatility range of ±10% after earnings, and the actual movement is only 4%, it means the result did not exceed expectations; a volatility of 12% would signify that expectations have been broken. "The focus of research is not the outcome itself, but the difference between the result and the expectation."
Methodologically, she insists on first showing AI the evidence, allowing it to list evidence before making conclusions; otherwise, "it can easily package some scattered information into seemingly definite answers." She assigns four roles to AI: information organizer, data analyst, case retriever, and questioner; the last role requires it to actively seek counterexamples. The list of pitfalls also comes from her practical experience: a single large order does not equal a direction; the execution of a call option may be a sell; high volatility does not equal bearish; any price point must be viewed in historical percentile context. "The most important capability in AI investment research is not finding a magical indicator but building a checklist that does not easily misinterpret signals."
Noise Reduction: From Earnings Reports to WeChat Groups
Guandu presented a financial report scoring system that integrates real market data, scoring from several dimensions: valuation, profitability, growth, and risk, outputting four conclusions: increase, hold, observe, and decrease. He also demonstrated with Tesla, Apple, and Micron stocks. More remarkable than the tools themselves are his three judgments: financial reports are standardized information that AI can handle 90% of; financial reports are merely "arrangement tools" used to validate investment logic and do not represent future cash flows, "a good financial report does not mean it can be bought immediately," because financial report cycles may contradict investment cycles; and AI manages information collection, data processing, and logical analysis, while the final decision should still belong to humans, as investment requires independent thinking and critical thinking.
Xiaowanzi's topic was the most "down-to-earth" and relatable: opinion verification radar, addressing the question of "Big names all make sense, but whom to believe." She used precious metals investment as a case: it is well-known that gold is bullish in the long term, but during the market from 1974 to 1980, those who initiated positions in 1973 based on correct direction had to endure a price halving for two years, and then waited 40 months for an 8-fold increase, only for it to drop back to half in the subsequent two months. Her radar reduces noise across five layers: cyclical nesting (over monetary cycles, there are monetary system cycles, "what we want are the interest payments on US bonds, what they want is our principal"), information source funnel (ranking information sources by impact and credibility), data calibration (self-purchasing Wind data interface, with a focus on London and New York markets), technical levels (cross-market metrics like gold-silver ratio, gold-copper ratio), and emotional thermometer. Lastly, the most relatable aspect: monitoring the density of chats in WeChat groups, keyword rankings, and red envelope activity; when the market is good, red envelopes flood the chat; when the market is bad, the group falls silent. According to her calculations, the daily trading volume of gold far exceeds annual production, and the main perspective on gold prices is the trading volume in financial markets, not the supply and demand news. The tool's output is not simply a buy or sell recommendation but the impact and credibility of each piece of information along with action recommendations for different phases, coupled with ideas for switching carriers: leaning towards options and futures in good market conditions, reverting to stocks and spot during poor conditions.
Lulu, a student from the blockchain industry, shared a lightweight multi-source cross-checking method: it’s not that there is insufficient information, but rather too much fragmented information. Breaking down big questions into six to eight smaller questions, matching each question with the most suitable information source (search breadth for general information, semantic search management for understanding, news timeliness for up-to-date information, social media for sentiment, web scraping for depth), merging and deduplicating multi-source results before outputting conclusions. She is currently testing a search aggregation tool charged per session, which aggregates mainstream search engines and crawlers, with failed sources automatically switching routes.
On-Chain Arbitrage: The Visible Price Difference is Fake
Programmer Xiaobai took the stage last, with the clearest stance: extremely averse to uncertainty, only seeking certain opportunities. His topic is to audit the mismatch between resources, information, and rules using arbitrage thinking.
The on-chain examples come with reversals: a visible price difference of about $80 between two decentralized exchanges, turning over 1,000 DAI yields a gross profit of over $80, seemingly pocketing it in just 12 seconds; however, deducting gas fees and "bribery fees" paid to on-chain block builders results in a net loss, causing the system to skip this transaction. The visible price difference is fake, which is the lesson he wanted to convey most that evening. He also showcased an on-chain arbitrage trade that allegedly captured about $450,000 at a cost of around $17,000, involving hundreds of consecutive operations. From this, he established a three-step verification method: mathematically valid, mechanically executable (completing low buy and high sell within the same block), and a real-world net profit that is positive. He wrote the formula for opportunities: transferable differences, minus all friction costs, minus unbearable uncertainty.
He also promoted arbitrage thinking to daily life: working remotely with a dollar salary in low-cost areas, living in Shenzhen while working in Hong Kong, the information gap regarding visas and tax systems. "The barrier is twofold: if I can’t cross it, it’s my barrier; if I have crossed it, it’s a barrier for others." In his view, information disclosure does not equate to "simultaneous, uniform, and objective"; the value of AI lies in organizing non-structured data like announcements and news into verifiable structured information, while humans must engage in adversarial review. He reminded that AI's attention mechanism naturally retrieves along the user's conversational lead, and the model is trained to be more pleasing, "it’s naturally pleasing you," thus one must proactively push it to find counterexamples.
His on-site demonstration of the "liquidation following" strategy perfectly matched this week's market: monitoring on-chain liquidation maps, after dense liquidations are triggered, buying induced by forced liquidations of short positions will continue to push prices higher, creating a rolling effect. This mechanism is not just theoretical; on-chain data analysis firm Glassnode attributed part of this round of Bitcoin's rise to short liquidations on August 19 in public analysis. The host further reminded that one market maker lecturer he invited earlier had spoken of various "tricks" on the order book, including hidden orders and manipulating buying and selling, which are routine operations; establishing mechanisms does not mean making money is simple. On the tool level, Xiaobai recommended using Grok's automated tasks to monitor on-chain information on Twitter (Grok is a native Twitter product with natural advantages in crawling), and also through Coinglass's API to monitor total liquidations across the network and funding rates.
Beyond the 150 Point Advancement Line
This sharing session was also part of a selection mechanism. The host emphasized the initial advancement threshold three times that evening: 150 points; those who do not meet this will be eliminated, and students can make up points through check-ins, social media posts, and signing up to share. He also announced that an in-depth research report on Bitcoin's rise would be released the next day, and students' access to on-chain data interfaces is expected to be established within two weeks; the first student project roadshow on Sunday will feature six students presenting; from September 5 to 6, Uweb will also hold an AI investment research practical training course offline in Shenzhen.
Uweb, short for University of Web3, was founded in 2022 by the original team of Huobi University (established in 2018) under the leadership of Dr. Yu Jianing, based in Hong Kong, offering Web3 and digital asset courses for entrepreneurs and investors. The host recalled during the opening that investment discipline and separate accounts were topics already discussed "when it was still called Huobi University." Eight years have passed, the podium has changed to students, the topics have shifted from discipline to AI, but the core has not changed.
For nearly three hours, none of the eight taught "how to let AI predict tomorrow's rise and fall," but rather they all did the same thing: delineating boundaries for AI, allowing it to list evidence, find counterexamples, and execute rules. The more urgent the market, the more valuable this work becomes. To conclude the session, Lin Jiarui stated: "You are responsible for setting the rules, and it is responsible for not letting your emotions break the rules."
(Compiled by: Alma Li. This article is a compilation based on the verbatim transcript of the Uweb student online sharing session on August 29, 2026, with edits and not reviewed by the speakers; student names are based on their on-site aliases and transcript details.)
Editor’s Note: This article uses the verbatim transcript of the sharing session as the sole source of content; ambiguous proprietary terms, inconsistent numbers, and verbal errors (such as price points from verbal market reports) are handled under the "better to omit than guess" principle; all case data narrated by speakers is retained in its attributed language. The background of the organizing body and market data is verified through the aforementioned public sources.
Disclaimer: This article is event reporting; all viewpoints and data in it belong to the individual speakers and do not constitute any investment advice.
Related public information:
3. Bitcoin Historical Price Chart (BTC123)
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
