Stock prediction by means of XGBoost, LSTM, and Transformer

Stock price prediction remains a long-standing and difficult problem in financial time series analysis as market data shows non - non-stationarity, noise, and high volatility. As a result of the fast development of machine learning and deep learning, data-driven models are essential for modeling complex time dependencies in the stock market. This paper uses historical data of NVIDIA (NVDA) from 1999 to 2025 to look at the effectiveness of three representative models - eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Transformer - for stock prediction. Using daily adjusted closing prices and trading volumes, a unified feature engineering and sliding window framework is constructed to make sure of fair comparisons between models. The findings from the experiments show that while XGBoost performs robustly as a traditional machine - learning reference point, deep - learning models, particularly the Transformer, show a better ability to capture long-term dependencies and changes in the market mechanism. These findings give empirical outlooks on the advantages and limitations of different modeling frameworks in financial time series prediction.