An autonomous crypto-trading agent that leverages Large Language Models for decision-making, featuring a real-time Dash control panel.

DeepTrade: An LLM-Powered Trading Agent

Core Thesis

This project explores a fundamental question: Can a Large Language Model, when provided with a rich context of real-time market data and governed by a strict set of rules, function as a viable decision-making engine for short-term algorithmic trading?

DeepTrade is a framework built to test this hypothesis on the Binance Futures Testnet, focusing on a robust prompt-engineering strategy to guide the LLM's analytical process.

Architecture Overview

The system operates in a continuous loop, executing the following steps:

1. Data Ingestion & Feature Engineering: The prompt_builder.py script connects to the Binance API to fetch the latest OHLCV data for multiple assets (BTC, ETH, etc.) across different timeframes. It then enriches this raw data with a suite of technical indicators using pandas-ta.

2. Context-Rich Prompt Construction: All market data, technical indicators, open interest, funding rates, qualitative news sentiment, and current account status (positions, P&L) are dynamically assembled into a single, comprehensive prompt. This gives the LLM a full situational snapshot.

3. LLM Reasoning & Decision: The prompt is sent to the deepseek-reasoner model. A carefully crafted system_prompt.md acts as the model's constitution, defining its rules of engagement, risk parameters, and the mandatory JSON output format. This is where the core trading logic resides.

4. Action & Execution: The bot parses the JSON response from the LLM. Based on the signal (e.g., buy_to_enter, close_position, hold), it e