Machine Learning-Based Bitcoin Trading Under Transaction Costs: Evidence From Walk-Forward Forecasting

Battle Count: 82

Highly relevant to quantitative trading. The paper directly addresses the prediction-to-trading gap, demonstrating that transaction costs can completely reverse the conclusion from gross equity curves. The cost-aware execution filter is a practical and implementable solution for reducing turnover in short-horizon strategies. The walk-forward protocol, bootstrap inference, and multi-architecture comparison provide a rigorous framework for evaluating trading strategies. The finding that execution discipline matters more than model complexity is a key insight for practitioners.

Moderate-to-high complexity. Requires implementing three distinct model architectures (XGBoost, LSTM, iTransformer), a 27-fold walk-forward protocol with Optuna hyperparameter tuning, feature engineering (technical indicators, EGARCH), cost-aware execution logic, and paired circular block-bootstrap inference. The cost-aware filter itself is simple (threshold-based), but the full experimental pipeline with 1,350+ XGBoost trials and 1,080+ neural network trials per configuration across multiple feature tiers and loss functions is computationally intensive. The iTransformer implementation adds complexity due to variable-as-token attention mechanism.

The paper provides detailed hyperparameter search spaces, walk-forward protocol specifications, feature engineering procedures, and statistical inference methods. Data is sourced from Binance public REST API. Code is available from the corresponding author upon request but no public repository is linked. The 27-fold walk-forward design, Optuna hyperparameter tuning, and bootstrap procedures are well-documented. However, exact random seeds and full pipeline code are not publicly available.