LLM trading

LLM trading with Gemini, Grok, DeepSeek and Llama

LLM trading means using a large language model such as Gemini, Grok, DeepSeek or Llama to help you trade, and it works best when the model writes a rule you can test instead of guessing where prices go. Any of them can turn an idea into exact entry and exit rules, and to backtest those rules on real market history you connect the model to a tool such as Keel over MCP, where your app allows it. In Season 1 of Nof1's Alpha Arena, where models traded real money on their own, four of the six lost money.

Start free Prompts that end in a backtest
By Keel Research Team
Updated October 18, 2026
Two ways to use an LLM

The model trades

It sees prices and decides each trade. Nothing to test beforehand, and results shift with the prompt.

The model writes a rule

It turns your idea into rules. The rules are backtested with costs, then paper traded, before any money moves.

What LLM trading means

Two ways to use a language model for trading

The first way is to let the model trade: give it prices and an account and let it decide, trade by trade. That is what Alpha Arena tested, and it is the version most people picture. The model has no edge it can explain, its answers shift with small changes in wording, and fees punish the many quick trades it tends to make.

The second way is to let the model write. You describe an idea, the model turns it into exact rules, and those rules are backtested and paper traded before any money moves. The model does what it is good at, which is language and structure, and the testing is done by software that charges real costs. An AI trading assistant, or trading copilot, in this sense is a model that helps you build and check rules, not one that picks trades.

Everything below assumes the second way. It works with any capable model, so use the one you already like.

Model by model

How to use each model for trading

Trading with Gemini

Gemini here means Google's AI model, not the Gemini crypto exchange. Give it your notes, a research paper or a strategy you found, and ask it to turn the idea into rules with every setting stated.

On MCP: Google documents MCP servers for Gemini CLI, its command-line agent, and experimental MCP support in its Gen AI SDK. Google's help centre also describes adding custom MCP apps to the Gemini app; check that page for what it requires. Where your Gemini setup can add a remote MCP server, it can connect to Keel.

In Alpha Arena: Gemini 2.5 Pro took part in Alpha Arena's first season and finished down more than half. See the results

Trading with Grok

Grok comes from xAI. Use it to work through a trading idea: ask it for the rule, then ask it to argue against its own rule and list how it could fail.

On MCP: xAI documents remote MCP tools in its API, and its Grok connectors page describes adding a custom connector on grok.com by entering an MCP server's address. Where your account offers connectors, that is the route to connecting Grok to Keel.

In Alpha Arena: Grok 4 finished Alpha Arena's first season down a little under half. In Season 1.5, on US stocks, Nof1 named Grok 4.20 the winner. See the results

Trading with DeepSeek

Use DeepSeek's models to turn a vague idea into a precise rule and to check that rule for look-ahead mistakes.

On MCP: DeepSeek's API supports tool calls and is compatible with the OpenAI and Anthropic formats, and DeepSeek's own docs show its models running inside Claude Code. The MCP connection then comes from the app you run the model in, so a DeepSeek model in an app that supports MCP can use Keel's tools.

In Alpha Arena: DeepSeek v3.1 finished second in Alpha Arena's first season, one of the two models that made money. See the results

Llama and other open models

Meta's Llama models are open weights: you download them under a licence for each version and run them on your own hardware or a provider of your choice. That suits traders who want to keep their ideas on their own machine.

Whether an open model can use MCP depends on the app that runs it, not on the model. Check whether the app you run it in supports MCP. Whatever the model, read every line of the rule it writes before you test it.

Connecting a model to Keel

What any model needs to backtest with Keel

Keel's MCP server lives at https://mcp.usekeel.io/mcp. Any app that can add a remote MCP server connects to it and signs you in to Keel; nothing to install and no key to paste. Keel has step-by-step guides for Claude, ChatGPT, Claude Code, Cursor, Codex, Windsurf and VS Code. For other apps, use the app's own add-a-server flow with the same address.

Setup for each assistant Other MCP servers for trading
What to ask any model

Ask for the rule, then for the holes in it

Whichever model you use, the useful requests are the same. Ask it to write your idea as rules a computer can follow, with the entry, exit, size and coins on separate lines. Ask it what market would break the rule. Ask it to list what could make a good backtest misleading. Leave out questions about what will go up next: no model can see the future, and on its own it has no live prices to work from.

Prompts that end in a backtest
Testing what it builds

Test the rule before it trades

A rule only means something once it has run on history. With a model connected to Keel, ask it to build the rule from Keel's blocks and backtest it. Keel runs the backtest on Hyperliquid history with fees, slippage and funding charged, and gives you a link to the result in the Keel app, where you can read the strategy as blocks and see every trade.

If the result holds up, paper trade it in the Keel app on live Hyperliquid prices with simulated money and no wallet. If it doesn't, you have lost nothing but a few minutes, which is the point of testing first.

Using ChatGPT instead Using Claude instead Open-source trading agents
FAQ

Questions about LLM trading

Can Gemini trade for me?

Not well on its own. Gemini has no account access or live prices by default, and in Alpha Arena's first season Gemini 2.5 Pro, trading on its own, lost more than half its money. Use it to write a rule, then backtest the rule before anything trades.

Is DeepSeek good at trading?

DeepSeek v3.1 finished second in Alpha Arena's first season, one of only two models to make money, but one short contest says little. Turning ideas into precise rules you then test is the safer way to use it.

Can Grok build a trading bot?

It can write the rules for one. To backtest and run them you need a tool that executes strategies; xAI documents custom MCP connectors for Grok, which is how it can reach one. With Keel, the rule is built from Keel's blocks and backtested on Hyperliquid history.

Which LLM is best for trading?

There is no reliable answer yet. In Alpha Arena's first season Qwen3-Max and DeepSeek made money and four models lost, and in Season 1.5 Grok 4.20 won, but Nof1 itself warns that early results may be luck. Pick the model you like and judge its rules by a backtest.

What is an AI trading assistant?

A language model that helps you build and check trading strategies: it turns ideas into rules, explains backtest results and points out what could go wrong. Connected to a tool like Keel, it can also run the backtest. ChatGPT and Claude have their own guides.

Sources

Facts about each model's MCP support come from its maker's own documentation, and Alpha Arena results from Nof1's own pages, read on the date shown. These change often, so check the source before you rely on one.

Use the model you like, test what it writes

Connect your assistant to Keel, have it build a rule, and backtest it on Hyperliquid history before anything trades.

Start free