Comparative analysis of classical, machine learning, deep learning, and adaptive neuro-fuzzy models for forecasting the S&P 500 index using financial, macroeconomic, and technical indicators

Abstract

Forecasting the stock market, particularly the S&P 500 index, is an essential yet challenging task due to inherent volatility, nonlinearity, and structural instability in financial data. This study investigates the effectiveness of different predictive models. It examines classical statistical approaches such as ARIMA, VECM, ETS, and Holt-Winters. It also evaluates machine learning algorithms including Random Forests, XGBoost, LightGBM, SVR, k-NN, and MLP. Additionally, it analyzes deep learning architectures like LSTM, GRU, and ARIMA-LSTM. The study further includes adaptive neuro-fuzzy inference systems. All models use financial, macroeconomic, and technical indicators. These indicators were collected from reputable sources between May 2020 and May 2025. The results indicate a significant enhancement in predictive accuracy when incorporating macroeconomic and technical indicators, validating the primary hypothesis. Among evaluated models, ANFIS demonstrated superior forecasting capability, effectively capturing market complexities and uncertainties, thus supporting the second hypothesis. Furthermore, feature selection and dimensionality reduction techniques considerably improved model accuracy and robustness, confirming the third hypothesis. However, limitations such as data non-stationarity, structural breaks, and inherent market noise constrained the forecasting precision. This research underscores the potential of advanced and hybrid predictive methodologies, providing valuable insights for stakeholders navigating the dynamic landscape of financial markets.