Xiao, Sun, Luo & Wang — arXiv 2412.20138 (Tauric Research)
A committee of language-model agents reads each company — four analysts, a bull and a bear researcher who argue it out, a trader, a risk desk — and votes buy, sell or hold.
no picks
Booked, per closed position
0 closed · average excess vs the market
—
Closed positions ahead of the market
no picks
Open book, marked to market
0 names still held
—
Average sessions held
window is 20 sessions
The ledger below is empty because the record is only 15 sessions old and nothing has yet left the portfolio. The open book is real and marked to the last close.
The rule, as implemented here
every condition, in the order it is applied
Every afternoon the framework is given the twenty most liquid names in the universe, by median traded value over twenty sessions.
Four analyst agents read price, fundamentals, news and social data for each name, as the framework fetches them.
A bullish and a bearish researcher debate; a trader agent decides; a risk desk and a portfolio manager approve or reject.
The portfolio is every name the committee voted buy. Sells and holds are not acted on. A name it could not decide counts as hold.
Ranking: In the framework’s own order; no ranking of ours is applied. Portfolio is the top 12 names, equal-weighted.
Where our version differs: Run exactly as released, on DeepSeek, over the twenty most liquid names each afternoon. We record its buy list before the next session and score it like every other row here. Its written reasoning is generated prose about real companies and is not published anywhere on this site.
Tonight’s run: 15 names put to the committee for the close of 9 September 2026 (carried forward — no completed run for tonight’s session). It voted buy on 0, sell on 2, hold on 12, and could not decide 1 — on its own five-tier scale: hold 12, underweight 2; overweight counts as a buy here and underweight as a sell. Framework TradingAgents on deepseek / deepseek-chat, 1 debate round, analysts: market, social, news, fundamentals. Finished 2026-09-10 18:10 IST.
Today’s portfolio
as of 10 September 2026
nothing clears these rules today
Portfolio record by window
every published day, measured forward and averaged. “Typical day” is the median excess, which ignores one runaway session
Window
Portfolio
Universe
Excess
Typical day
Days ahead
Days scored
1d
no picks
no picks
no picks
no picks
no picks
0
5d
no picks
no picks
no picks
no picks
no picks
0
10d
no picks
no picks
no picks
no picks
no picks
0
20d
not yet
not yet
not yet
not yet
not yet
0
3m
not yet
not yet
not yet
not yet
not yet
0
Booked positions
closed, with the market over the identical days
No position has closed yet. A name closes when the rules drop it or when it has been held 20 sessions, whichever comes first — so the first entries appear here about 20 sessions after the book opened.
Open book
still held, marked to the last close
No position is open. This book publishes its picks in advance and holds each for up to 20 sessions, so the shelf is empty either between runs or because every name it took has already been closed and scored below.
How these numbers were made. Computed here from NSE and BSE end-of-day bhavcopy, adjusted for splits and bonuses, not supplied by anyone being measured. Portfolios are recorded on the day they are published and scored forward from that date, so no figure on this page could have been chosen after the outcome was known. The code that produced them is stamped in build.json.
This is our implementation, not TradingAgents’s portfolio. Run exactly as released, on DeepSeek, over the twenty most liquid names each afternoon. We record its buy list before the next session and score it like every other row here. Its written reasoning is generated prose about real companies and is not published anywhere on this site. None of the authors named on this site are associated with it, and nothing here is advice or a recommendation to buy or sell anything.