Comparison of short-term and long-term forecasts using machine learning across S&P 500 industries with financial and market indicators
摘要
Stock-price prediction is essential for investment strategies, risk management, and financial market efficiency. While fundamental analysis relies on macroeconomic indicators and financial statements, and technical analysis focuses on historical price trends, the relative effectiveness of these assessment strategies across different investment horizons, such as short-term (daily) and long-term (quarterly) timelines, remains relatively underexplored. This study compared fundamental and technical information sets across time horizons by forecasting one-day-ahead and one-quarter-ahead prices for twenty-two sector-balanced S&P 500 constituents using eight model families spanning kernel methods, tree ensembles, and deep neural networks. Company financial statements and macroeconomic indicators were used for quarterly prediction, while financial and technical indicators were used for daily prediction. To reflect the dynamic nature of markets, this study implemented a rolling, walk-forward re-estimation design that repeatedly retrained models at each forecast origin and reported out-of-sample performance using MAPE and R2, with sector-level summaries to characterize cross-sectional error structure. Crucially, the rolling protocol combined with model-specific hyperparameter grids yielded an exhaustive computational workload exceeding 150,000 training runs. Therefore, the study carried out GPU-accelerated training and testing on a high-performance computing (HPC) cluster that enabled transparent benchmarking at scale in the context of Explainable AI (XAI) in financial market forecasting. Therefore, comparing the results of long- and short-term predictions under various machine learning models provided insights into the strengths and limitations of the models employed across varying time horizons.