Cryptocurrency markets exhibit high volatility and structural inefficiencies, creating both opportunities and risks for traders. Traditional trading strategies rely on fixed heuristics, which often fail to adapt to dynamic price movements. This study explores the integration of machine learning (ML) classifiers—Random Forest, Decision Trees, K-Nearest Neighbors (KNN), and Extra Trees—with optimized raw return labeling and hyperparameters using Tree-structured Parzen Estimator (TPE) to enhance BTCUSDT trading performance. Using Binance historical data from August 17, 2017, to November 30, 2024, models were trained using trend technical indicators and OHLC (Open, High, Low, close) data. To mitigate class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) was applied. A trading simulation compared ML-driven strategies against a passive buy-and-hold benchmark. Results demonstrated that ML classifiers significantly outperformed buy-and-hold. The Random Forest model achieved the highest return of 716.52%, representing a 142.7% improvement over buy-and-hold, with an 86.21% win rate across 29 trades. Decision Trees followed with a 704.93% return, outperforming buy-and-hold by 138.7%, executing 44 trades with a 72.73% win rate. Extra Trees delivered a 453.90% return, a 53.8% increase over buy-and-hold, with a 76.00% win rate across 25 trades. The findings suggest that machine learning-based trading models can exploit inefficiencies in cryptocurrency markets, achieving superior risk-adjusted returns. However, considerations such as transaction costs and model generalizability warrant further investigation.