Paper: arXiv 2510.08068

Authors: Aadi Singhi

Abstract

This paper presents a Multi Agent Bitcoin Trading system that utilizes Large Language Models (LLMs) for alpha generation and portfolio management in the cryptocurrencies market. Unlike equities, cryptocurrencies exhibit extreme volatility and are heavily influenced by rapidly shifting market sentiments and regulatory announcements, making them difficult to model using static regression models or neural networks trained solely on historical data. The proposed framework overcomes this by structuring LLMs into specialised agents for technical analysis, sentiment evaluation, decision-making, and performance reflection. The agents improve over time via a novel verbal feedback mechanism where a Reflect agent provides daily and weekly natural-language critiques of trading decisions. These textual evaluations are then injected into future prompts of the agents, allowing them to adjust allocation logic without weight updates or finetuning. Back-testing on Bitcoin price data from July 2024 to April 2025 shows consistent outperformance across market regimes: the Quantitative agent delivered over 30% higher returns in bullish phases and 15% overall gains versus buy-and-hold, while the sentiment-driven agent turned sideways markets from a small loss into a gain of over 100%. Adding weekly feedback further improved total performance by 31% and reduced bearish losses by 10%. The results demonstrate that verbal feedback represents a new, scalable, and low-cost approach of tuning LLMs for financial goals.

Complexity vs Empirical Score

  • Math Complexity: 2.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; no out-of-sample test; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper’s mathematical complexity is low, relying mostly on standard financial metrics like Sharpe ratio and volatility rather than dense formulas, but it shows high empirical rigor with specific backtesting results on real Bitcoin data, detailed performance metrics across regimes, and a proposed implementation-ready LLM agent system.

Research Flowchart

  flowchart TD
  A["Research Goal"] --> B["Develop LLM Multi-Agent System"]
  B --> C["Data & Inputs"]
  C --> D["Core Process"]
  D --> E["Key Outcomes"]
  
  A(["Research Question:<br>Can LLM agents improve<br>crypto trading via verbal<br>feedback without retraining?"])
  
  C(["Data & Inputs:<br>BTC Price Data (Jul 2024 - Apr 2025)<br>Market Sentiment & News"])
  
  D(["Core Process:<br>1. LLM Agents (Tech/Sentiment/Decision)<br>2. Daily/Weekly Verbal Feedback<br>3. Natural Language Prompt Updates"])
  
  E(["Key Outcomes:<br>- 15% Overall Gain vs Buy & Hold<br>- 100% Gain in Sideways Markets<br>- 31% Perf. Boost w/ Weekly Feedback<br>- Lowered Bearish Losses"])

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a multi‑agent Bitcoin trading system that uses large language models with a dual verbal feedback mechanism for alpha generation and portfolio management.

Source sentence
This paper presents a Multi Agent Bitcoin Trading system that utilizes Large Language Models (LLMs) for alpha generation and portfolio management in the cryptocurrencies market.

Assumes 4

  • Bitcoin price movements are driven primarily by non‑fundamental forces such as regulatory announcements, influencer statements, and speculation. market
    Source sentence
    Bitcoin’s price is largely driven by non-fundamental forces such as regulatory announcements, influencer statements, and speculative
  • The system trades once per day using daily closing prices, ignoring intraday dynamics. execution
    Source sentence
    The system was evaluated on daily Bitcoin data from July 2024 to April 2025.
  • LLMs can effectively process real‑time textual signals without parameter updates. model
    Source sentence
    LLMs offer a potential solution to this unpredictability, as they are particularly suited for large-scale textual analysis and pattern recognition.
  • Transaction costs, slippage and market impact are not incorporated into the reported headline results. execution
    Source sentence
    even highly accurate models may fail if execution costs, slippage, or regime shifts negate expected returns.

Data

Backtest sample: 2024-07 – 2025-04 · Bitcoin · crypto · daily

  • daily Bitcoin price data market-prices unknown
    July 2024 to April 2025 · Bitcoin
    Source sentence
    The system was evaluated on daily Bitcoin data from July 2024 to April 2025.
  • news data via GNews API · GNews text-news public
    Source sentence
    First, news data from trusted financial outlets (e.g., CNBC, Yahoo Finance, Forbes, Business Insider, Seeking Alpha, and Fortune) is aggregated via the GNews API
  • Fear and Greed Index · Alternative.me alternative public
    Source sentence
    Second, the Fear and Greed Index, obtained from Alternative.me, provides a composite daily score ranging from extreme fear to extreme greed
  • Senticrypt social media sentiment scores · Senticrypt alternative public
    Source sentence
    Finally, social media sentiment scores are collected through the Senticrypt API, which processes Twitter discussions to quantify retail investor mood in real time

Recognised sources: Twitter / X Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
even highly accurate models may fail if execution costs, slippage, or regime shifts negate expected returns.

Finds 3

  • The Quantitative agent delivered over 30% higher returns in bullish phases and 15% overall gains versus a buy‑and‑hold benchmark. performance
    Source sentence
    the Quantitative agent delivered over 30% higher returns in bullish phases and 15% overall gains versus buy-and-hold
  • The sentiment‑driven agent turned sideways markets from a small loss into a gain of over 100%. performance
    Source sentence
    the sentiment-driven agent turned sideways markets from a small loss into a gain of over 100%
  • Adding weekly verbal feedback improved total performance by 31% and reduced bearish losses by 10%. performance
    Source sentence
    Adding weekly feedback further improved total performance by 31% and reduced bearish losses by 10%

Checked by others

Contradicted by

  • ATLAS: Adaptive Trading with LLM AgentS Through Dynamic Prompt Optimization and Multi-Agent Coordination (2025)
    On the question: Does providing feedback to adaptive multi‑agent trading systems improve performance?
    Compare the findings

    This paper: Adding weekly verbal feedback improved total performance by 31% and reduced bearish losses by 10%.

    That paper: Reflection‑based feedback fails to provide systematic gains and can be inconsistent across models.

    A shows +31% total performance and -10% bearish losses with weekly verbal feedback; B finds reflection‑based feedback gives no systematic gains and is inconsistent across models.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.