Analysis of the inferred causal structures showed that the returns improved when trades were executed only during rolling-window trials in which specific directed edges to the target variable were detected, and proposed a causal filter strategy that restricts trading to periods in which specific directed edges to the target variable are detected.
Abstract
In the highly volatile cryptocurrency market, trading decision support based on price prediction remains a challenging task. Although machine learning and deep learning techniques have been widely applied to cryptocurrency price prediction, many existing approaches rely on correlation-based black-box models, which limits interpretability and robustness. In this study, we employed a NOTEARS-Linear-based Prediction Model (NLBPM) that directly incorporated causal structures inferred through a causal discovery method as structural constraints within the prediction model. Unlike conventional approaches that focus primarily on minimizing prediction error, the NLBPM emphasized return maximization as its objective function, thereby prioritizing practical economic value. Using Bitcoin as a case study, we constructed a model to predict the direction of price movement four hours ahead and evaluated its performance using a rolling-window scheme with a one-month sliding window. Analysis of the inferred causal structures showed that the returns improved when trades were executed only during rolling-window trials in which specific directed edges to the target variable were detected. Based on this finding, we proposed a causal filter strategy that restricts trading to periods in which specific directed edges to the target variable are detected. In the data period analyzed in this study, the selected edge was the one from the opening price (Open) to the target variable. Backtesting experiments incorporating a transaction fee of 0.1% demonstrated that, while the benchmark LSTM model achieved a negative monthly average return of −3.20% and the NLBPM without filtering yielded −0.72%, the NLBPM with the Open filter attained a higher monthly average return of 10.35%. This study supports the usefulness of using inferred causal structure for cryptocurrency trading decision support.
Accurately forecasting cryptocurrency price movements following market overreactions is crucial for traders, investors, and risk managers operating in highly volatile environments. This study presents a novel multi-source, multi-output deep learning framework designed to predict the direction of price changes in four major cryptocurrencies — Bitcoin (BTC), Ethereum (ETH), Binance Coin (BNB), and Ripple (XRP) — immediately after overreaction events. By jointly modeling these assets, our approach captures their interconnected market dynamics, enhancing predictive accuracy. We compile an extensive dataset with over 656 features from diverse sources, including historical trading data, on-chain metrics, technical indicators, and social sentiment data from Google Trends, collected at both daily and intraday frequencies. To improve model interpretability and performance, we introduce two engineered features — price change magnitude and price variation speed — that effectively represent intraday volatility. Feature selection using a Random Forest approach reduces the feature set to 30 key variables, ensuring robustness and avoiding overfitting. Using three advanced deep learning architectures — LSTM, RNN, and CNN — we train models to classify the next-day price movement as upward or downward. Empirical results demonstrate that the multi-output LSTM achieves an F1-score of 73.42%, outperforming both single-asset models (62.95–68.25%) and alternative architectures. These findings highlight the benefits of joint modeling, leading to more reliable forecasts during turbulent market conditions. Our framework offers a practical tool for algorithmic trading, portfolio management, and risk mitigation in the dynamic cryptocurrency landscape.
Unknown authors· International Journal of Fin...· 0 citations
Current machine learning-based quantitative high-frequency trading suffers from a significant Prediction-to-Trading Gap. Academic and industrial research overly focuses on the optimization of model prediction errors while ignoring the devastating impact of real-market trading frictions on strategy returns. Taking the hourly high-frequency trading of BTC/USDT perpetual contracts as the research scenario, this paper constructs a quantitative trading system covering three heterogeneous time-series models (XGBoost, LSTM, and iTransformer) under the constraint of 10 bps full-dimensional transaction costs. A Cost-Aware Execution Filter (CA) is introduced, combined with the 27-fold non-anchored Walk-Forward Optimization (WFO) framework and a three-layer nested feature engineering system for empirical research. The results show that all three machine learning models can achieve excess returns in the frictionless scenario, with the annualized return of iTransformer reaching 181.76%. However, when 10 bps real trading frictions are introduced, all naive unfiltered strategies incur comprehensive losses, with annualized returns collapsing ranging from -64.00% to -98.00%. The CA filtering mechanism substantially reduces the strategy turnover rate by two orders of magnitude and effectively repairs strategy returns. The optimal strategy achieves an annualized return of 65.40% with a Sharpe ratio of 1.09. The empirical results verify that the core bottleneck of high-frequency quantitative trading is not model prediction accuracy but the cost adaptation mechanism of signal transformation, and a simple and efficient transaction cost filtering strategy is far more valuable than blindly iterating complex time-series models. This study provides an important reference for the research and development, real-market implementation, and standardized backtesting system construction of cryptocurrency high-frequency quantitative strategies.
Weijian Huang, Zhanwei Wang, Xianpeng Jiang· Journal of the European Acad...· 0 citations
In the field of high-frequency quantitative trading for cryptocurrencies, the industry has long fallen into an algorithm arms race centered on deep learning models such as LSTM and Transformer. Researchers excessively pursue marginal improvements in predictive indicators including Mean Squared Error (MSE) and Directional Accuracy, which leads to a widespread dilemma: high prediction accuracy accompanied by poor live trading performance. Taking hourly high-frequency trading of Bitcoin as the research object, this paper constructs a comprehensive transaction cost system covering explicit handling fees, bid-ask spreads and slippage. Adopting the Walk-Forward dynamic backtesting framework, this study systematically compares the performance of traditional time series models, complex deep learning models and transaction cost-aware filtering strategies. The empirical results show that: first, there is a significant disconnect between model prediction accuracy and actual net returns; marginal improvements in prediction brought by complex models cannot offset profit losses caused by transaction frictions in high-frequency trading. Second, simple signal filtering rules designed based on transaction costs deliver far better profitability improvements than iterative optimization of deep learning architectures. Third, although the target strategies achieve impressive returns in single-path backtesting, their returns are extremely unevenly distributed across time intervals with weak statistical significance, indicating prominent stability risks in live trading. The conclusions of this research provide theoretical basis and practical references for the R&D of cryptocurrency quantitative strategies, the construction of standardized backtesting systems and live trading risk control.
Weijian Huang, Zhanwei Wang, Wenchang Jiang· Journal of the European Acad...· 0 citations
This work attempts to provide a thorough comparative analysis mapping the precise accuracy–efficiency trade-off between Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models under a standardized Grid Search hyperparameter optimization pipeline using a recent Bitcoin closing-price dataset.
Findings from the application of Deep Temporal Convolutional Networks in high-frequency cryptocurrency price forecasting are synthesized, highlighting TCNs' advantages in computational efficiency, robustness, and adaptability to rapidly shifting trading environments.
Xue Cheng· Applied and Computational En...· 0 citations
Cryptocurrency markets are highly volatile, nonlinear, and affected by several internal and external market factors, making price forecasting a challenging task. Accurate cryptocurrency price forecasting can support investors, traders, and financial analysts in making informed decisions. This research paper presents a comparative analysis of machine learning and deep learning models for cryptocurrency price forecasting using historical Aave (AAVE) cryptocurrency data. The dataset consists of 275 records and 10 features, including Date, High, Low, Open, Close, Volume, and Marketcap. The Close price is selected as the target variable, while High, Low, Open, Volume, and Marketcap are used as predictor variables. Five models are implemented and compared: Linear Regression, Support Vector Regression, Random Forest Regressor, XGBoost Regressor, and Long Short-Term Memory. The models are evaluated using Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, R-squared score, and directional accuracy. Experimental results show that the LSTM model achieved the best performance with the lowest RMSE of 2.74, MAE of 1.78, MAPE of 3.91%, and R-squared score of 0.965. The results indicate that deep learning models, especially LSTM, are more suitable for capturing temporal dependencies and nonlinear patterns in cryptocurrency price data.
Rukhsar Zaka, Faiza Irfan, Sidra Rehman et al.· International Journal of Adv...· 0 citations