COMPARATIVE EVALUATION OF STATISTICAL AND MACHINE LEARNING MODELS FOR STOCK MARKET FORECASTING: EVIDENCE FROM GLOBAL EXCHANGES

Abstract

This study compares statistical and machine learning models for stock market forecasting using daily closing prices from the Korea, Shanghai, Tokyo, Pakistan, and New York stock exchanges. Five models, ARIMA, GARCH, Naïve, Artificial Neural Networks (ANN), and Long Short-Term Memory (LSTM) were evaluated using MAE, RMSE, MPE, and MAPE. The results show that forecasting accuracy depends on market dynamics. Specifically, ARIMA performed best for Korea and Tokyo, LSTM was most effective for Shanghai, GARCH provided the most accurate forecasts for Pakistan, and the Naïve model outperformed others in New York. In general, LSTM excelled in capturing complex nonlinear behavior, while ARIMA and GARCH proved more reliable in volatile or stable environments. Overall, no single model dominates across all exchanges, but the evidence highlights LSTM’s strength in emerging markets and the continued relevance of traditional statistical methods in mature exchanges. These findings provide valuable insights for investors and policymakers in selecting context-appropriate forecasting tools.