Scopevest — Wang Qicai

What it is

Scopevest (观衡) is an AI research workspace for China’s A-share market. It gathers daily prices, fund flows and limit-up data, has a language model write a research judgment for each stock, and records the evidence behind it and the conditions that would prove it wrong. An AI paper-trading account follows those judgments with fixed rules. It never connects to a broker.

The Scopevest AI workflow: market switch, after-close screening, per-stock AI analysis and AI ranking, then the next-day trade plan, open buy, close sell and per-trade result — The AI workflow: after the close it checks market conditions and screens candidates, the AI analyzes and ranks each one, and the next day the paper account trades them under fixed rules. Stock names are blurred; the interface is in Chinese.

The question

Earlier versions came with impressive numbers: one limit-up rule set claimed +54% a year from a three-month backtest of 140 trades. A strict reproduction came in lower, and 27.7% of 1,000 random portfolios did better. Before putting real money behind anything, I wanted to know which of those claims survive a proper test.

How it was tested

Every hypothesis and parameter was written down and committed before the run (pre-registration), so results couldn’t be picked after the fact.
Long samples: 116,563 limit-up events from 2016 to 2026, monthly portfolios back to 2010, with separate training and test periods.
Real trading constraints: buy at the next open, no fill when it opens limit-up, T+1, and about 0.3 percentage points in round-trip costs.
Multiple-testing corrections, with every run counted as a trial.

Pre-registered results table for Study 3: point estimates for three periods, pooled mean with 95% interval, Holm-adjusted p-value and verdict for five hypotheses — Pre-registered results for Study 3: of five hypotheses, two were supported, one was falsified and two were inconclusive.

What we found

Long-only, daily data, public information: none of the speculative strategies we proposed made money after costs out of sample. Buying a limit-up stock at the next open loses about 1% to 1.4% per trade. The “2–3 consecutive limit-ups” rule that looked best over three months did 0.90% per trade worse than other limit-up stocks on the same day across ten years. Technical scores and multi-timeframe signals that worked in training reversed in testing, and news used as a direction signal did no better than a placebo.

What held up was a list of things not to do:

Don’t chase limit-up stocks at the next day’s open.
Don’t buy lottery-like stocks; they lag by about 1% a month.
Don’t buy stocks that hit limit-up in the past five days; they lag by about 1.7% a month.
Don’t chase prominent good news such as contract wins.

What I take from it

Limit-ups, streaks and headline announcements are where the market’s attention is most crowded. Faster, better-funded traders sit on the other side, and whoever buys at the next open pays for it.
Complexity didn’t create an edge. Better prompts and more indicators can’t get past the limit of public information.
The most valuable output is the method: pre-registration, out-of-sample tests, real costs and fixed decision deadlines. It kept me from putting money behind a strategy that only looked good for three months.

The AI paper-trading account keeps running as a forward test, with a verdict at 250 trades or April 15, 2027, under rules written down in advance.

The AI paper-trading account’s hypothesis and verdict rules: at trade 250 or April 15, 2027, whichever comes first, the 95% interval of per-trade returns decides whether it works — Verdict rules for the AI paper-trading account, fixed before any trade closed and unchanged until the deadline.

This is a research record, not investment advice.