Paper: arXiv 2510.08268

Authors: Kairan Hong, Jinling Gan, Qiushi Tian, Yanglinxuan Guo, Rui Guo, Runnan Li

Abstract

Cryptocurrency markets present unique prediction challenges due to their extreme volatility, 24/7 operation, and hypersensitivity to news events, with existing approaches suffering from key information extraction and poor sideways market detection critical for risk management. We introduce a theoretically-grounded multi-agent cryptocurrency trend prediction framework that advances the state-of-the-art through three key innovations: (1) an information-preserving news analysis system with formal theoretical guarantees that systematically quantifies market impact, regulatory implications, volume dynamics, risk assessment, technical correlation, and temporal effects using large language models; (2) an adaptive volatility-conditional fusion mechanism with proven optimal properties that dynamically combines news sentiment and technical indicators based on market regime detection; (3) a distributed multi-agent coordination architecture with low communication complexity enabling real-time processing of heterogeneous data streams. Comprehensive experimental evaluation on Bitcoin across three prediction horizons demonstrates statistically significant improvements over state-of-the-art natural language processing baseline, establishing a new paradigm for financial machine learning with broad implications for quantitative trading and risk management systems.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper is mathematically dense with formal theoretical guarantees and optimal property proofs for its fusion mechanism and multi-agent system, but it focuses more on architectural design and statistical significance from experiments rather than providing exhaustive backtesting details like code, datasets, or detailed performance metrics, placing it in a high-math, moderate-rigor category.

Research Flowchart

  flowchart TD
  A["Research Goal: Predict crypto trends<br>using off-exchange public info"] --> B["Input: Heterogeneous Data Streams<br>(News, Social, Technical, On-chain)"]
  B --> C["Step 1: Multi-LLM News Analysis<br>Extract 6-Impact Features<br>(Sentiment, Regulatory, Volume, Risk, Technical, Temporal)"]
  C --> D["Step 2: Multi-Agent Coordination<br>Distributed Processing with<br>Low Communication Overhead"]
  D --> E["Step 3: Adaptive Volatility-Conditional Fusion<br>Dynamically combine signals based on<br>Market Regime Detection (Bull/Bear/Sideways)"]
  E --> F["Key Findings & Outcomes<br>Significant accuracy improvements<br>over SOTA NLP baselines<br>Enhanced sideways market detection<br>for risk management"]

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 cryptocurrency trend prediction framework with three innovations: an information‑preserving multi‑dimensional news analysis system, an adaptive volatility‑conditional fusion mechanism, and a distributed low‑communication multi‑agent architecture.

Source sentence
We introduce a theoretically‑grounded multi‑agent cryptocurrency trend prediction framework that advances the state‑of‑the‑art through three key innovations: (1) an information‑preserving news analysis system

Assumes 5

  • Bitcoin is used as a representative asset for cryptocurrency markets. market
    Source sentence
    7.9% market share, BTC serves as a representative asset.
  • Prediction horizons are fixed at 1, 7, and 15 days with asymmetric tolerance thresholds. model
    Source sentence
    Thresholds, adapted from stable triplet labeling [28], apply asymmetric tolerances: ±0.30% (1‑day) for short‑term sensitivity, ±0.60% (7‑day) to reduce noise, and ±0.40% (15‑day) for persistence–consolidation balance
  • Financial news articles are quantified across seven dimensions using large language models. model
    Source sentence
    We propose a seven‑dimensional quantitative system capturing market impact, regulation, volume, risk, technical correlation, and timing, with provable information content preservation and analytical guarantees.
  • Volatility‑conditional fusion weights reflect sentiment‑led price discovery and temporal precedence. model
    Source sentence
    Our approach dynamically fuses news sentiment and technical indicators via volatility‑driven weights, reflecting sentiment‑led price discovery and temporal precedence in trading markets.
Show 1 more
  • The multi‑agent system operates with low communication complexity suitable for real‑time processing. execution
    Source sentence
    distributed multi‑agent coordination architecture with low communication complexity enabling real‑time processing of heterogeneous data streams.

Data

  • Binance BTC price data · Binance market-prices commercial
    2025-07-21 to 2025-09-06 · BTC
    Source sentence
    Historical Transaction Data Timestamp, Open, High, Low, Close
  • Financial news articles text-news unknown
    Source sentence
    News Size 9528 financial news articles

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • The multi‑dimensional news analysis framework improves accuracy by 10.4 to 20.4 percentage points over keyword‑extraction baselines across 1‑, 7‑, and 15‑day horizons. performance
    Source sentence
    The multi‑dimensional approach achieves accuracy improvements ranging from 10.4 percentage points (1‑day horizon) to 20.4 percentage points (15‑day horizon) compared to keyword extraction baselines.
  • For the 1‑day horizon, the proposed system attains 0.4375 accuracy, 0.3596 macro‑F1, and 0.4722 balanced accuracy versus baseline 0.3333, 0.2518, 0.3500. predictive-accuracy
    Source sentence
    1d Baseline 0.3333 0.2518 0.3500 Ours 0.4375 0.3596 0.4722
  • Balanced accuracy improves markedly, reaching 0.6071 for 15‑day predictions compared to baseline 0.0333. predictive-accuracy
    Source sentence
    15d Baseline 0.0625 0.0417 0.0333 Ours 0.2667 0.2509 0.6071

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.

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.

Cited References (40)

  1. Breaking news headlines: Impact on trading activity in the cryptocurrency market (2023) — link
  2. Deep Learning and NLP in Cryptocurrency Forecasting: Integrating Financial, Blockchain, and Social Media Data — → 2311.14759
  3. Multilevel Analysis of Cryptocurrency News using RAG Approach with Fine-Tuned Mistral Large Language Model (2025) — link
  4. Predicting the price of bitcoin using machine learning (2018)
  5. Bitcoin price prediction using machine learning (2020)
  6. Text mining for market prediction: A systematic review (2014)
  7. FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link
  8. Agent-based computational finance (2006)
and 32 more
  1. Market-Derived Financial Sentiment Analysis: Context-Aware Language Models for Crypto Forecasting — → 2502.14897
  2. Controllable and Reliable Knowledge-Intensive Task-Oriented Conversational Agents with Declarative Genie Worksheets (2025) — link
  3. Stock market prediction via multi-class classification (2017)
  4. A multi-class classification approach for stock trend prediction (2020)
  5. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation (2020)
  6. Not all words are equal: Sentiment and jumps in the cryptocurrency market (2024) — link
  7. Using Sentiment and Technical Analysis to Predict Bitcoin with Machine Learning (2024) — link
  8. The impact of news media on Bitcoin prices: modelling categorical discourse (2022) — link
  9. Optimizing Forecast Accuracy in Cryptocurrency Markets: Evaluating Feature Selection Techniques for Technical Indicators (2025) — link
  10. Nonlinear nexus between cryptocurrency returns and COVID-19 news sentiment (2022) — link
  11. The impact of regulation on cryptocurrency market volatility in the context of the COVID-19 pandemic: Evidence from China (2023) — link
  12. Long and short-term impacts of regulation in the cryptocurrency market (2021) — link
  13. Cryptocurrency Market Reactions to Regulatory News (2020) — link
  14. Forecasting seasonals and trends by exponentially weighted moving averages (1957)
  15. The moving average convergence/divergence (MACD) indicator (1979)
  16. New concepts in technical trading systems (1978)
  17. The stochastic oscillator (1950s)
  18. When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks (2011)
  19. Bollinger on Bollinger Bands (1980s)
  20. A comparison of transformer-based models and traditional TF-IDF for cryptocurrency sentiment analysis (2021)
  21. The predictive power of traditional keyword extraction for Bitcoin price movement (2020)
  22. Performance comparison of traditional and contextual representations for cryptocurrency sentiment analysis on Twitter (2023) — link
  23. Lexicon-based sentiment analysis versus transformer models in cryptocurrency prediction: A comparative study (2022)
  24. Keyword frequency analysis and sentiment scoring for cryptocurrency price movement prediction: Limitations and improvements (2021)
  25. Attention-based CNN–LSTM for high-frequency multiple cryptocurrency trend prediction (2024) — link
  26. LangGraph: A framework for building stateful, multi-actor applications with LLMs (2023)
  27. The Balanced Accuracy and Its Posterior Distribution (2010)
  28. Cryptocurrencies as a financial asset: A systematic analysis (2019) — link
  29. Volatility estimation for Bitcoin: A comparison of GARCH models (2017)
  30. Deep learning and NLP in cryptocurrency forecasting (2025) — link
  31. Multilevel analysis of cryptocurrency news using RAG approach with fine-tuned Mistral Large Language Model (2025) — link
  32. FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link

2 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.