An Adaptive Multi-Agent Bitcoin Trading System

Battle Count: 210

Highly relevant to quantitative trading. The paper directly addresses Bitcoin portfolio management, market regime classification, alpha generation, and risk-adjusted returns. It introduces a novel verbal feedback mechanism as an alternative to traditional fine-tuning for LLM-based trading systems. The evaluation uses standard quantitative finance metrics (Sharpe ratio, cumulative returns, regret). The multi-agent architecture with specialized roles mirrors real-world trading desk structures. The focus on profitability over accuracy and the regime-based analysis are directly applicable to quantitative trading practice. However, it is limited to a single asset (Bitcoin) and daily frequency, which constrains broader applicability.

Moderate complexity. The system requires: (1) API integrations with multiple data sources (Binance, GNews, SentiCrypt, Alternative.me, Dune Analytics), (2) technical indicator computation using Python TA module, (3) structured prompt engineering for 5+ agents, (4) JSON output parsing and validation, (5) feedback injection logic for daily and weekly cycles, (6) deterministic rule-based evaluator for weekly feedback, and (7) backtesting infrastructure. The core logic relies on LLM API calls (DeepSeek-r1) with carefully crafted prompts. No model training or fine-tuning is required, reducing computational complexity. The main challenge lies in prompt engineering, ensuring consistent JSON outputs, and managing the feedback loop without information leakage between agents.

The paper provides detailed system prompts in the appendix, specifies data sources (Binance API, GNews API, SentiCrypt API, Alternative.me Fear & Greed Index, Dune Analytics), and describes the agent architecture clearly. However, no GitHub repository or code is mentioned. The exact prompt templates, feedback injection logic, and backtesting code are not fully provided. The use of DeepSeek-r1 API is specified but exact API parameters and versioning are unclear. The weekly feedback rule set is provided in the appendix, aiding reproducibility.