Research benchmark only. Not financial advice. Crypto trading can lose all capital.

Using an LLM for trading: what the live record shows

An LLM for trading is a large language model handed market data and asked to decide what to buy, sell or hold. Since January 2026, 56 LLM competitors have each managed $10,000 of simulated capital on live crypto and US stocks under one prompt and one rulebook, across 8 completed seasons and 2,724 trades. Only 39% of 77 model-seasons, one model running one season, finished profitable. Measured across 5 season-to-season transitions, a model's rank in one season does not detectably carry into the next. Pooled across Seasons 2 to 6, the field opened 32% of its positions long, in the one rising-market season the archives cover and in the falling ones alike. An LLM can read a market, state a thesis and place a coherent order. The record so far says it does not reliably make money doing so.

Data refreshed September 3, 2026. The settled-season figures are stamped July 18, 2026.

What an LLM can do in trading

People who search for LLMs for trading usually mean one of four jobs. Reading: turning news, filings and social feeds into a sentiment or a thesis. Building: writing and backtesting strategy code from a plain-English description. Analyzing: playing the analyst roles inside multi-agent frameworks such as TradingAgents and FinMem, where several model instances debate before a separate trader agent acts. Deciding: taking the market data and placing the order itself.

Academic benchmarks such as StockBench test that last job on historical windows, scored once. TradeRank tests the same job forward, on live prices, every day, with the reasoning published.

What the live record shows

The all-time counts in the summary above include every completed season, settled or not. Three of the four findings below use the 7 settled seasons, Seasons 0 through 6, whose archives are final; the consensus finding runs on its own snapshot, which reaches into the season that was in progress when it was taken. Later seasons join when their archives settle. Method and every figure: the settled-seasons study and the consensus study.

Rank does not carry between seasons

Across the 5 consecutive-season pairs with 7 to 10 models in common, the pooled rank correlation between one season's finishing order and the next is -0.001, with an interval of -0.35 to +0.35. A slot that finished in the top half repeated top-half 9 times in 20 tries; chance predicts 9.34. The interval is wide enough to hide a real but modest effect, so the finding is that nothing is detectable at this sample size, not that nothing exists.

The field has leaned short since Season 3

Share of opening orders that were longs, season by season: Season 2, 94%; Season 3, 5%; Season 4, 14%; Season 5, 25%; Season 6, 28%. Pooled across Seasons 2 to 6, that is 746 opening orders, 239 long and 507 short. The Seasons 2, 3 and 4 shares rest on partial decision archives (8%, 24%, 37% of cycles for the worst-covered model), so read them as the direction of the lean rather than a measured share; Seasons 5 and 6 (97% and 90%) clear the bar. The short lean began in Season 3 and held through every settled season since, in the one rising-market season the archives cover and in the falling ones alike. It does not move with the market. It is a default.

Win rate does not rank models

Across 51 model-seasons, the raw correlation between win rate and season return is -0.125: winning more often reads as slightly worse. That number is an artifact of pooling seasons with different average win rates and returns. Subtract each season's mean from both sides and the correlation is +0.31. Within a season, winning more often does go with finishing higher. It is a weak relationship, and the pooled number reports its opposite.

Agreement is not measurable signal

Three or more models backed the same direction on the same asset in the same daily cycle 59 times across Seasons 3 to 7, as of the August 2, 2026 snapshot; 2 never resolved. Over the next cycle the crowd was right 43.9% of the time, 25 of 57. A dummy that always says down scored 47.4% on the same events, and the gap is inside chance (p = 0.691069). The 42 bearish calls score identically to the dummy by construction, so the whole difference comes from 17 bullish events, and the consensus study found the sign of that gap flips with which events are included. These are not 57 independent bets: events cluster within cycles. Agreement carries no measurable information beyond drift here, which is not proof that it carries none.

How to use an LLM for trading

The rulebook around the model does more work than the prompt inside it. Each model gets one decision per day at a fixed time, so it cannot churn. It sees a one-row-per-asset summary of the whole tradeable universe plus raw multi-timeframe candles for the assets it holds or that the daily screen surfaces, so it reads prices rather than someone's indicator summary. It must answer in a fixed schema: an action, a size as a share of equity, a confidence, and an invalidation price on every open or add. A validator rejects anything outside the rules and allows one repair attempt. A monitor then checks every 15 minutes and closes the position once price has touched its invalidation level, whether or not the model would have. Its thesis is carried into the next day's prompt, so it argues with its own past self instead of starting blank.

The prompt template and the full rule set: how the LLM trading competition works. The same prompt shape, ready to copy: the copy-paste trading prompts. The arena runs on simulated capital only.

Where it goes wrong

Which LLM should you use?

The ranking changes season to season, so this page does not name one. The current standings and every settled season winner are on the best LLM for trading page; the all-time dataset and how it is built are on the live LLM trading benchmark. This season's field, each with its own record:

Frequently asked questions

Can an LLM be used for trading?

Yes, and it is being done live. Each model in the TradeRank arena receives the same market data and the same prompt once a day, returns a structured decision, and a validator executes or rejects it on a $10,000 simulated account. Placing the trade is the easy part. Whether it makes money is the question, and 8 completed seasons of results are below.

Are LLMs profitable at trading?

Mostly not, so far, and no model has shown it can repeat. Measured across 5 season-to-season transitions, a model's rank in one season does not detectably predict its rank in the next, and only 39% of 77 model-seasons across 8 completed seasons finished above their starting capital. Individual seasons produce winners. Research benchmark only. Not financial advice.

How do you use an LLM for trading?

TradeRank gives each model a fixed job, a fixed data feed and a rulebook it cannot leave: one decision per day from a one-row-per-asset summary of the whole tradeable universe plus raw candles for the assets it holds or screens; a structured order with a required invalidation price on every open or add; a confidence floor; and a validator that rejects anything outside the rules and allows one repair attempt. The prompt template and the full rule set are published on the how-it-works page.

Why do LLMs lose money trading?

The live record points at three habits. The field has leaned short since Season 3: 32% of 746 opening orders across Seasons 2 to 6 were long. Win rate misleads: across 51 model-seasons the raw correlation between win rate and return is -0.125, and only after subtracting season means does it turn positive, at +0.31. And agreement is not measurable signal: when three or more models backed the same direction, the crowd was right 43.9% of the time over the next daily cycle against 47.4% for a dummy that always says down, a gap inside chance.

Where the numbers come from

Four generated files. The projection this page reads is served at llm-for-trading-facts.json; it is built from research-facts.json and pillar-consensus.json, and the all-time counts come from traderank-facts.json.