FinAgent

Inspiration

The goal of this project is to observe if LLM agents framed as traders can take rational trading decisions and end up maximising their profits in a simple simulation.

What it does

LLM Agents to compete between themselves in a basic market simulation. We define a playground for several AI agents which trade between each others and we analyse the results associated.

How we built it

Mistral Large LLM with Mistral API using in-context learning with few-shot examples to make agents mimic an expected output for each step (streaming price + trading)

Challenges we ran into

Creating the trading simulation, enforce expected outputs for the LLM, generating suitable reference examples for the LLM and finding suitable rules, ...

Accomplishments that we're proud of

We have created the simulation and obtained financial behaviour with only prompt/text and an LLM.

What we learned

LLM seems to behave rationally in this simulation between. We have found interesting results in terms of streaming prices: we observe convergences in terms of prices. Many more results are to be explored...

What's next for FinAgent

Pursue analysis of results with more assets, more agents and round. We can also add all the history in the prompt. Adding specific rules. An extension can be made easily in which one person could play a trading game vs LLM Agents. Add RL in the game.

Built With

• api
• mistral
• python