Scholarlink

<p>Cryptocurrency markets focus on digital currency trading mechanisms, price prediction, and market dynamics through financial analysis, risk management, and regulatory frameworks. Investment interest in cryptocurrencies has increased due to their anonymous and decentralized nature, but price prediction through artificial intelligence (AI) in volatile markets remains challenging, with scarce research in high-frequency trading (HFT). This is caused by the swiftness required for data retrieval and prediction to execute a successful trade. Thus, achieving successful HFT necessitates leveraging the full potential of AI to execute timely decisions. This study proposes a hybrid convolutional neural network (CNN) model based on long short-term memory for multi-class crypto prediction. The main contribution is the statistical analysis of the impact of AI prediction on investors’ financial decisions and profitability improvements achievable in multi-granular models. The feature engineering of Fibonacci retracement levels captures support and resistance zones, providing additional information for generating trade signals based on market trends. The methodology applied for profitability analysis research is grounded in a simulated investment environment, which involves data collection, preprocessing, sampling, training, prediction, and finally investment simulation for evaluation. This study demonstrates that incorporating Fibonacci retracement levels significantly enhances model performance and profitability across various architectures and temporal domains. It has been observed that C-LSTM consistently outperformed other models, particularly on 1-minute data for both BTC and ETH, while MLP and LSTM showed mixed results depending on prediction type and time granularity. Trend and trend-strength predictions benefited from Fibonacci Retracement Levels, yielding ROI improvements in up to 70% of configurations. The highest long ROI gain reached 45%, while trend-strength classification using MLP showed a maximum increase of 31.71% and a potential loss of -56.73%. For long-short trading, C-LSTM provides stable and consistent gains, especially on 1-minute data. These findings highlight the importance of fine-grained temporal data, model selection, and further study of reward-risk management in high-frequency trading scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI crypto trading: multi-class multi-granular analysis for boosting high-frequency trade predictions with fibonacci and hybrid convolutional neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月3日

查看全文
  • Bilal Hassan Ahmed Khattak,
  • Chaudhry Hamza Rashid,
  • Imran Shafi,
  • Sultan Alfarhood,
  • Mejdl Safran,
  • Imran Ashraf

摘要

Cryptocurrency markets focus on digital currency trading mechanisms, price prediction, and market dynamics through financial analysis, risk management, and regulatory frameworks. Investment interest in cryptocurrencies has increased due to their anonymous and decentralized nature, but price prediction through artificial intelligence (AI) in volatile markets remains challenging, with scarce research in high-frequency trading (HFT). This is caused by the swiftness required for data retrieval and prediction to execute a successful trade. Thus, achieving successful HFT necessitates leveraging the full potential of AI to execute timely decisions. This study proposes a hybrid convolutional neural network (CNN) model based on long short-term memory for multi-class crypto prediction. The main contribution is the statistical analysis of the impact of AI prediction on investors’ financial decisions and profitability improvements achievable in multi-granular models. The feature engineering of Fibonacci retracement levels captures support and resistance zones, providing additional information for generating trade signals based on market trends. The methodology applied for profitability analysis research is grounded in a simulated investment environment, which involves data collection, preprocessing, sampling, training, prediction, and finally investment simulation for evaluation. This study demonstrates that incorporating Fibonacci retracement levels significantly enhances model performance and profitability across various architectures and temporal domains. It has been observed that C-LSTM consistently outperformed other models, particularly on 1-minute data for both BTC and ETH, while MLP and LSTM showed mixed results depending on prediction type and time granularity. Trend and trend-strength predictions benefited from Fibonacci Retracement Levels, yielding ROI improvements in up to 70% of configurations. The highest long ROI gain reached 45%, while trend-strength classification using MLP showed a maximum increase of 31.71% and a potential loss of -56.73%. For long-short trading, C-LSTM provides stable and consistent gains, especially on 1-minute data. These findings highlight the importance of fine-grained temporal data, model selection, and further study of reward-risk management in high-frequency trading scenarios.