Paper: arXiv 2407.09546

Authors: Yuan Li, Bingqiao Luo, Qian Wang, Nuo Chen, Xu Liu, Bingsheng He

Abstract

The utilization of Large Language Models (LLMs) in financial trading has primarily been concentrated within the stock market, aiding in economic and financial decisions. Yet, the unique opportunities presented by the cryptocurrency market, noted for its on-chain data’s transparency and the critical influence of off-chain signals like news, remain largely untapped by LLMs. This work aims to bridge the gap by developing an LLM-based trading agent, CryptoTrade, which uniquely combines the analysis of on-chain and off-chain data. This approach leverages the transparency and immutability of on-chain data, as well as the timeliness and influence of off-chain signals, providing a comprehensive overview of the cryptocurrency market. CryptoTrade incorporates a reflective mechanism specifically engineered to refine its daily trading decisions by analyzing the outcomes of prior trading decisions. This research makes two significant contributions. Firstly, it broadens the applicability of LLMs to the domain of cryptocurrency trading. Secondly, it establishes a benchmark for cryptocurrency trading strategies. Through extensive experiments, CryptoTrade has demonstrated superior performance in maximizing returns compared to traditional trading strategies and time-series baselines across various cryptocurrencies and market conditions. Our code and data are available at \url{“https://anonymous.4open.science/r/CryptoTrade-Public-92FC/"}.

Complexity vs Empirical Score

  • Math Complexity: 1.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper relies on LLM prompting and traditional financial indicators (e.g., MACD, MA) rather than dense mathematical theory, but features heavy empirical backtesting with specific crypto assets, market conditions, and published code/data.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Bridge LLM gap in<br>cryptocurrency trading"] --> B["Data Sources<br>On-chain & Off-chain signals"]
  B --> C["LLM-based Agent<br>CryptoTrade Framework"]
  C --> D["Core Process<br>Reflective Decision Loop"]
  D --> E["Execution<br>Zero-shot Trading"]
  E --> F["Key Outcomes<br>Superior returns vs baselines<br>Benchmark established"]

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 CryptoTrade, an LLM‑driven trading agent for cryptocurrencies that combines on‑chain and off‑chain data with a self‑reflective mechanism.

Source sentence
We introduce CryptoTrade, an innovative trading agent in the cryptocurrency domain, driven by LLMs.

Assumes 5

  • Trading decisions are executed daily (day‑to‑day frequency). execution
    Source sentence
    Another limitation is the frequency of trading actions, which is currently set to day-to-day.
  • A proportional transaction fee is applied to each trade. execution
    Source sentence
    When a trading decision is made, a transaction fee is charged in proportion to the traded value.
  • The model operates in a zero‑shot manner without fine‑tuning on validation data. model
    Source sentence
    Notably, CryptoTrade operates in a zero-shot manner without fine-tuning based on validation sets
  • Test period is chosen after September 2021 to avoid data leakage from the LLM’s knowledge cutoff. data
    Source sentence
    We carefully select the test period after September 2021, the GPT-3.5’s knowledge cutoff date, to prevent data leakage.
Show 1 more
  • Initial capital is $1 million split equally between cash and the three cryptocurrencies. execution
    Source sentence
    We initialize the trading agent with 1 million US dollars, split equally between cash and BTC/ETH/SOL

Data

Backtest sample: BTC/ETH/SOL · crypto · daily

  • On‑chain data · Dune.com crypto-onchain public
    BTC, ETH, SOL
    Source sentence
    The foundation of our methodology relies on a…
  • Gnews API news articles · Google News (Gnews API) text-news public
    cryptocurrency‑specific outlets (Bloomberg, Yahoo Finance, crypto.news)
    Source sentence
    on filtering news from reputable financial and cryptocurrency‑specific outlets such as Bloomberg, Yahoo Finance, and crypto.news
  • Market price data for BTC, ETH, SOL market-prices unknown
    Bitcoin, Ethereum, Solana
    Source sentence
    CryptoTrade using GPT‑42, and GPT‑4o3, evaluating its proficiency in making daily trading decisions for Bitcoin (BTC), Ethereum (ETH), and Solana (SOL).

Recognised sources: Bloomberg

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not modeled

Paper text extraction: Modeled - A transaction fee proportional to the traded value is charged for each trade.

Source sentence (the authors' words)
When a trading decision is made, a transaction fee is charged in proportion to the traded value.

Finds 4

Reported performance (CryptoTrade (GPT-4o)): Sharpe 0.23 (benchmark 0.25)

  • CryptoTrade outperforms time‑series baselines and achieves comparable performance to MACD across bull, sideways, and bear markets. performance
    Source sentence
    and achieves comparable performance to trading signals like Moving Average Convergence Divergence (MACD)
  • During the ETH bullish test period, CryptoTrade exceeded the Buy‑and‑Hold return of 22.59% by a remarkable 3%. performance
    Source sentence
    For instance, during the ETH bullish test period, the Buy and Hold strategy secured a 22.59% return, while CryptoTrade exceeded this by a remarkable 3%.
  • CryptoTrade (GPT‑4o) improves total return and Sharpe ratio by about 3% in ETH’s bull market. performance
    Source sentence
    CryptoTrade (GPT‑4o) excels in all metrics under ETH’s bull market by 3% in total return and sharpe ratio.
  • Even without fine‑tuning, CryptoTrade surpasses more than half of the trading signals across different market conditions. performance
    Source sentence
    it consistently surpasses more than half of the trading signals across different market conditions, even without fine‑tuning.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 4 papers in the archive (4 by independent authors)

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 (55)

  1. Bloomberggpt: A large language model for finance (2023) — link
  2. Transforming Financial Decision-Making: The Interplay of AI, Cloud Computing and Advanced Data Management Technologies (2023)
  3. Stock Market 101: From Bull and Bear Markets to Dividends, Shares, and Margins—Your Essential Guide to the Stock Market (2024)
  4. Bitcoin volatility in bull vs. bear market-insights from analyzing on-chain metrics and Twitter posts (2023)
  5. Investigating the effectiveness of task-agnostic prefix prompt for instruction following (2024)
  6. Interoperability in blockchain: A survey (2023)
  7. The hidden shortcomings of (d) aos–an empirical study of on-chain governance (2023)
  8. Cryptocurrency uncertainty and volatility forecasting of precious metal futures markets (2023)
and 47 more
  1. What is mature and what is still emerging in the cryptocurrency market? — → 2305.05751
  2. Summarization is (almost) dead (2023) — link
  3. Holistic evaluation of language models (2022) — link
  4. Harnessing the power of llms in practice: A survey on chatgpt and beyond (2024)
  5. ChatGPT informed graph neural network for stock movement prediction — → 2306.03763
  6. Pixiu: A large language model, instruction data and evaluation benchmark for finance (2023) — link
  7. Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models — → 2402.03659
  8. A survey of large language models (2023) — link
  9. Multimodal chain-of-thought reasoning in language models (2023) — link
  10. Chain-of-thought prompting elicits reasoning in large language models (2022)
  11. Transformers in time series: A survey (2022) — link
  12. Timesnet: Temporal 2d-variation modeling for general time series analysis (2022)
  13. Frequency-domain MLPs are more effective learners in time series forecasting (2024)
  14. A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
  15. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
  16. Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
  17. Non-linear prediction of security returns with moving average rules (1996)
  18. Predicting stock price trend using MACD optimized by historical volatility (2018)
  19. The profitability of Bollinger Bands trading bitcoin futures (2023)
  20. Llama 2: Open foundation and fine-tuned chat models (2023) — link
  21. Temporal Data Meets LLM–Explainable Financial Time Series Forecasting — → 2306.11025
  22. Fingpt: Democratizing internet-scale data for financial large language models (2023) — link
  23. The rise and potential of large language model based agents: A survey (2023) — link
  24. SocioDojo: Building Lifelong Analytical Agents with Real-world Text and Time Series (2024) — link
  25. Offline reinforcement learning: Tutorial, review, and perspectives on open problems (2020) — link
  26. Reflexion: an autonomous agent with dynamic memory and self-reflection (2023) — link
  27. Cryptocurrency price prediction using traditional statistical and machine-learning techniques: A survey (2021)
  28. The promises and pitfalls of machine learning for predicting stock returns (2021)
  29. Stock price prediction using reinforcement learning (2001)
  30. Stock return prediction under GARCH—An empirical assessment (2017)
  31. Stock price prediction using artificial neural network (2014)
  32. Deep learning-based stock price prediction using LSTM and bi-directional LSTM model (2020)
  33. Cryptocurrency price prediction using tweet volumes and sentiment analysis (2018)
  34. Cryptocurrency price prediction using time series and social sentiment data (2019)
  35. A lstm-method for bitcoin price prediction: A case study yahoo finance stock market (2019)
  36. Reflexion: Language agents with verbal reinforcement learning (2024)
  37. Self-refine: Iterative refinement with self-feedback (2024)
  38. Automatic prompt optimization with" gradient descent" and beam search (2023) — link
  39. Refiner: Reasoning feedback on intermediate representations (2023) — link
  40. Bitcoin and cryptocurrency technologies: a comprehensive introduction (2016)
  41. The Theory of Parsing, Translation and Compiling (1972)
  42. Publications Manual (1983)
  43. Alternation (1981) — link
  44. Scalable training of $L_1$ -regularized log-linear models (2007) — link
  45. Algorithms on Strings, Trees and Sequences (1997) — link
  46. Yara Parser: A Fast and Accurate Dependency Parser (2015) — link
  47. An algorithm for the machine calculation of complex F ourier series (1965) — link

4 of 55 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.