The financial market has long been an important domain for the application of machine learning (ML). Since the 1980s, researchers have leveraged ML techniques to uncover patterns and trends in financial data. Yet, stock market prediction remains inherently challenging due to its volatility and non-linearity. This paper presents a comprehensive multi-model forecasting framework integrating classical statistical models and modern deep learning methods. We employ three baseline models—Linear Regression (LR), Autoregressive Integrated Moving Average (ARIMA), and Long Short-Term Memory (LSTM)—as well as two extended architectures: an ARIMA-LSTM ensemble and a Transformer-based model. Our experiments span five representative U.S. stocks (AAPL, KO, NVDA, PFE, TSLA), reflecting diverse volatility profiles. A unified preprocessing pipeline—including sliding window segmentation, normalization, and volatility-aware sampling—is applied across all models to ensure fair evaluation. Results show that while ARIMA performs best on volatile assets, LSTM benefits significantly from input normalization, and ensemble approaches offer enhanced robustness. This work contributes an extensible, model-agnostic framework for financial time series prediction, providing actionable insights for both academic research and practical trading applications.