Twenty stocks. Five ratings. One running test.
TradingAgents
Field Test
Every US trading day a team of LLM analysts rates 20 stocks Buy to Sell. The ratings are traded on an MT5 demo account and compared with other ways of reading the same ratings, and with simply holding all 20.
Latest complete run
Waiting for first run
No market cutoff recorded
01 / The traded portfolio
Stateful five-tier
Long only. What the MT5 demo EA executes, modelled from daily closes: no spreads, commissions or fill slippage.
Model equity
$10,000.00
0.00% since start
Current target book
Each of the 20 bars is one $500 sleeve. Fill shows the target position the EA holds after the latest signal; the rest is cash earning 0%.
02 / Same ratings, different portfolios
Interpretation race
TradingAgents says Buy or Underweight, but never how much. Each line reads the same daily ratings a different way. Toggle lines below.
| # | Portfolio | Equity | Return | vs B&H | Max DD | Invested |
|---|
03 / Research only — not traded
Long / short
Treats Underweight and Sell as real short bets. Do the negative ratings carry profitable information, or only "own less"?
Scores are not positions. In the fixed-sleeve portfolio Sell is a full −$500 short, whatever the other 19 ratings are. In the two score portfolios, Sell (−2) only means twice the short weight of Underweight (−1); its final size depends on every other rating and the normalization rule.
04 / Passive benchmark
Twenty-stock buy-and-hold
Each selected range starts at $10,000 with $500, or 5%, invested in each symbol. Fractional shares are bought once and then held without rebalancing. Adjusted closes include split and distribution adjustments.
Contribution over the selected range
05 / The universe
Constituents
Deliberately mixed: sectors, sizes, volatility and retail popularity. Click a stock for its price history.
Symbol tape
—
06 / Research resources
Run cost and scale
Measured time, calls and tokens. API cost is an estimate, never a bill.
07 / Rules of the game
How every rating becomes a position
Four different kinds of rule live in this table and they are not interchangeable: exposure is a fixed share of a $500 slot, stateful depends on yesterday's position, score is relative weight normalized across all 20, and control ignores the ratings.
| Portfolio | Buy | Overweight | Hold | Underweight | Sell |
|---|
Timing. A rating generated from data up to session t only earns returns from close t to close t+1. No portfolio ever profits from a move that happened before its signal existed.
Prices. Daily adjusted closes for all 20 stocks, refreshed after every run. A missing price counts as a 0% move, never an invented one.
What is not modelled. Spreads, commissions, overnight financing, borrow fees for shorts, the broker's 0.1-share step, and the gap between the close and the 09:31 open fill. Cash earns 0%.
Not advice. A public research experiment on a demo account. Nothing here is a recommendation to buy or sell anything.