This review synthesizes recent applications of large language models (LLMs) for stock price forecasting from a hedge-fund perspective, covering sentiment extraction, financial report analysis, price series tokenization, and multi-agent trading systems. It critically evaluates practical pitfalls often understated in academic literature, including fragility in sentiment analysis due to context dependence and regime shifts, short evaluation horizons that fail to span full market cycles, misalignment of evaluation metrics (e.g., MSE, accuracy) with real-world trading objectives (e.g., risk-adjusted returns), data leakage from cross-day event drift and literal label statements, and the impact of illiquidity premiums on strategy feasibility. The review also discusses the limits of stock price predictability, suggesting that markets alternate between random-walk and predictable regimes, and advocates for regime-switch detection. Practical guidelines are provided for deploying LLMs in production-grade trading pipelines, such as using long-horizon datasets, risk-adjusted metrics, rigorous leakage controls, and benchmarking against non-LLM baselines. The paper highlights the gap between academic prototypes and real-world trading systems, emphasizing the need for robustness under realistic market frictions.