This study investigates the “Accuracy Paradox” in algorithmic equity trading by examining whether a machine learning model with near-random directional prediction accuracy can still generate profitable trading outcomes when combined with effective risk- management and market-regime rules. An XGBoost classifier was developed and evaluated using approximately six years of daily market data from five major technology stocks—NVIDIA, Apple, Microsoft, Alphabet, and Meta—covering the period from January 2020 to February 2026. The model incorporated twelve predictors based on technical indicators, including moving averages, RSI, MACD, rate of change, Bollinger Band position, volatility measures, and VIX-based sentiment proxies. A regime-aware trading framework was implemented using a 200-day simple moving average filter, a probability threshold for trade execution, a 2% stop-loss, a 4% take-profit target, and transaction costs. Walk-forward optimization with rolling training and testing windows was employed to reduce look-ahead bias and assess the robustness of the strategy. The results show that XGBoost achieved an average directional accuracy of approximately 49.8%, yet the resulting trading strategy generated positive returns in four of the five stocks. Meta produced the highest return of 32%, followed by Microsoft at 18%, Alphabet at 17%, and Apple at 5%, while NVIDIA recorded a loss of 13%. The findings indicate that predictive accuracy alone may not determine trading profitability and that regime filtering, disciplined trade execution, and asymmetric risk management can play a more important role in controlling losses and improving portfolio performance.
