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

Abstract

This study examines the viability of translating machine learning predictions of hourly BTC-USDT returns into profitable trading strategies once transaction costs are factored in. The analysis utilizes a dataset comprising roughly 70,000 hourly data points spanning from 2018 to 2026, applying a 27-fold walk-forward validation protocol to assess the performance of XGBoost, LSTM, and iTransformer models.

While all three algorithms demonstrate positive gross trading results in certain configurations, their viability diminishes significantly when a transaction cost of ten basis points is applied; under these conditions, simple sign-based strategies fail to generate profit. However, the implementation of a cost-aware execution filter—which restricts trading only when forecasted magnitude falls below a threshold derived from transaction costs—significantly curtails turnover and re-establishes profitability in specific setups. Notably, the most effective long-only XGBoost strategy achieved annualized returns exceeding 65% alongside a Sharpe ratio greater than one.

Further investigations reveal that while technical indicators can enhance performance in certain scenarios, features derived from EGARCH models do not consistently yield robust improvements. Additionally, while XGBoost appears descriptively superior to its neural network counterparts, bootstrap analysis fails to confirm formal statistical dominance. The impact of loss functions and model selection is found to be secondary and statistically unstable. Ultimately, the findings suggest that the primary challenge in hourly cryptocurrency trading lies not merely in limited predictability, but in the critical process of converting forecasts into executable trades.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC