Paper: arXiv 2508.02356

Authors: Wěi Zhāng

Abstract

This paper explores neural network-based approaches for algorithmic trading in cryptocurrency markets. Our approach combines multi-timeframe trend analysis with high-frequency direction prediction networks, achieving positive risk-adjusted returns through statistical modeling and systematic market exploitation. The system integrates diverse data sources including market data, on-chain metrics, and orderbook dynamics, translating these into unified buy/sell pressure signals. We demonstrate how machine learning models can effectively capture cross-timeframe relationships, enabling sub-second trading decisions with statistical confidence.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced neural network architectures like multi-head CNNs with attention mechanisms and statistical modeling, showing high mathematical density. It details real-world implementation challenges, data processing, and claims consistent performance metrics, indicating strong empirical rigor and backtest readiness.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Neural Network-Based Trading<br>in Crypto Markets"] --> B["Data Collection & Integration<br>Market Data, On-Chain, Orderbook"]
  B --> C["Computational Process<br>Multi-Timeframe Analysis<br>High-Frequency Prediction"]
  C --> D["Statistical Modeling<br>Systematic Market Exploitation"]
  D --> E["Key Findings:<br>Positive Risk-Adjusted Returns<br>Sub-Second Decisions<br>Statistical Confidence"]
  
  classDef goal fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
  classDef data fill:#fff3e0,stroke:#f57c00,stroke-width:2px;
  classDef process fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px;
  classDef outcome fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px;
  
  class A goal;
  class B data;
  class C,D process;
  class E outcome;

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 neural‑network based algorithmic trading system that combines multi‑timeframe trend analysis with high‑frequency direction prediction networks.

Source sentence
Our approach combines multi-timeframe trend analysis with high-frequency direction prediction networks, achieving positive risk-adjusted returns through statistical modeling and systematic market exploitation.

Assumes 5

  • Market information flows hierarchically across time‑frames, with larger‑scale trends constraining shorter‑scale movements. market
    Source sentence
    Our core hypothesis rests on the hierarchical nature of market information flow across different time‑frames.
  • Transaction costs are 0.05% per trade and are included in back‑test results. execution
    Source sentence
    The testing framework simulates realistic trading conditions including 0.05% transaction costs and accounts for variance in slippage to ensure results reflect achievable performance.
  • Feature normalization uses predetermined scaling parameters rather than running statistics to avoid data snooping. data
    Source sentence
    ... with feature normalization performed using predetermined scaling parameters rather than dataset running statistics to prevent data snooping and ensure model stability.
  • Direction networks are periodically retrained to remain relevant as market regimes shift. model
    Source sentence
    Direction networks undergo periodic retraining as market regime shifts occur, ensuring model relevance and preventing performance degradation over time.
Show 1 more
  • Risk controls (position limits, correlation limits, drawdown monitoring) operate independently of signal generation. execution
    Source sentence
    These controls operate independently of the signal generation process, providing multiple layers of protection against model failures or market disruptions.

Data

Backtest sample: cryptocurrency markets · crypto · intraday

  • Historical cryptocurrency market data market-prices unknown
    2018‑present · single asset
    Source sentence
    The data pipeline integrates three primary information sources into a unified framework for trading signal generation, utilizing historical data spanning from 2018 to present with the largest timeframe being daily intervals.
  • On‑chain analytics crypto-onchain unknown
    2018‑present · single asset
    Source sentence
    For single asset, the system integrates multiple data sources including on‑chain analytics capturing network activity and transaction patterns.
  • GDELT sentiment data · GDELT text-news public
    15‑minute frequency · global news and social media
    Source sentence
    ... providing 15‑minute frequency sentiment data from global news and social media sources.
  • S&P 500 index movements macro unknown
    S&P 500
    Source sentence
    overall economic indicators such as S&P 500 index movements

Recognised sources: GDELT

Costs

Modeled Costs are part of the reported headline results.

0.05% transaction costs per trade

Source sentence
The testing framework simulates realistic trading conditions including 0.05% transaction costs and accounts for variance in slippage to ensure results reflect achievable performance.

Finds 4

  • The CNN with soft‑attention architecture achieved the highest profit factor of 1.15, outperforming LSTM and CNN‑only variants. performance
    Source sentence
    CNN + Soft Attention (Current) 1.15 Optimal performance
  • Inference latency per prediction is sub‑50 ms, with total trade execution (including network and exchange latency) typically 100‑300 ms. performance
    Source sentence
    The inference pipeline achieves sub‑50ms latency per prediction, with complete trade execution typically requiring 100‑300ms including network latency and exchange processing time.
  • Direction networks are compact (~520,000 parameters) and optimized for fast inference, enabling sub‑second trading decisions. performance
    Source sentence
    The direction networks are relatively compact with approximately 520,000 parameters, optimizing for both performance and inference speed.
  • The hierarchical multi‑timeframe approach enables consistent profitability across diverse market regimes. performance
    Source sentence
    This paper demonstrates that neural network‑based approaches can achieve consistent profitability in cryptocurrency markets through systematic exploitation of cross‑timeframe relationships and microstructural patterns.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 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 (34)

  1. High-frequency trading: a practical guide to algorithmic strategies and trading systems (2013)
  2. Empirical market microstructure: The institutions, economics, and econometrics of securities trading (2007)
  3. High-frequency trading (2011)
  4. High-frequency trading and price discovery (2014)
  5. High frequency trading and the new market makers (2013)
  6. Surveying stock market forecasting techniques–Part II: Soft computing methods (2009)
  7. Deep learning for finance: deep portfolios (2017)
  8. Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500 (2017)
and 26 more
  1. Conditional time series forecasting with convolutional neural networks (2017) — link
  2. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019 (2020)
  3. Forecasting stock prices from the limit order book using convolutional neural networks (2017)
  4. Attention is all you need (2017)
  5. Neural machine translation by jointly learning to align and translate (2014) — link
  6. A dual-stage attention-based recurrent neural network for time series prediction (2017) — link
  7. The inefficiency of Bitcoin (2016)
  8. A new look at cryptocurrencies (2018)
  9. Cryptocurrencies as a financial asset: A systematic analysis (2019)
  10. Cryptocurrency trading: a comprehensive survey (2020)
  11. Twitter mood predicts the stock market (2011)
  12. Giving content to investor sentiment: The role of media in the stock market (2007)
  13. The color of money: The role of sentiment in financial markets (2013)
  14. GDELT: Global data on events, location, and tone (2013)
  15. GDELT-based event embedding for stock movement prediction (2022)
  16. Multi-time scale analysis of a single time series (2007)
  17. A hybrid approach based on GARCH models and discrete wavelet transform for stock price prediction (2016)
  18. Analysis of cryptocurrency bubbles (2019)
  19. Tokenomics: Dynamic adoption and valuation (2021)
  20. The evaluation and optimization of trading strategies (2008)
  21. Mutual fund performance (1966)
  22. The Calmar ratio: A smoother tool (1991)
  23. Measuring the information content of stock trades (1991)
  24. The price impact of order book events (2014)
  25. Deep learning (2016)
  26. Deep learning (2015)

0 of 34 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.