The volatility in cryptocurrency markets, along with the continuous trading environment requires the deployment of machine learning trading strategies capable of discerning complex patterns to avoid high losses. This study evaluates systematically, the performance of three classifiers: Random Forest, Decision Tree and Extra Trees, across four different timeframes (30, 60, 120, and 240 min) using Bitcoin/USDT trading pair data downloaded from Binance and spanning from August 2017 to December 2024. By integrating the triple barrier method for label generation -which combines predefined price movement thresholds with time horizon constraints- this approach captures rapid market dynamics with enhanced nuance. In addition to that, a selected subset of trend technical indicators have been generated for each time frame. Mutual information was used to conduct a feature selection of 10 features from all the available features. Time series cross-validation is employed to assess the classifiers’ performance using different metrics such as accuracy, area under the curve (AUC), and Sharpe ratio. Our obtained results indicate that ensemble-based methods, particularly Random Forest and Extra Trees, outperform the Decision Tree model, especially at shorter intervals, while a general decrease in accuracy accompanies longer timeframes despite improvements in risk-adjusted returns. Our findings suggest that timeframe and model selection have influence in constructing robust automated trading strategies.