| Abstract [eng] |
Reliable prediction of financial market movements remains a challenging task due to high volatility, complex interdependencies, and sensitivity to external shocks. This study assesses the performance of advanced machine learning models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer networks, Extreme Gradient Boosting (XGBoost), and Deep Multi-Layer Perceptron (DMLP), as well as proposes their ensemble combinations, in forecasting daily closing prices of five major stock indices (S&P 500, NASDAQ-100, Dow Jones Industrial Average, FTSE 100, and DAX). Results indicate that although all models achieved high predictive accuracy, profitability outcomes varied substantially across models and markets. Among single-model approaches, LSTM generally exhibited more stable positive returns in several indices, while other models showed pronounced variability depending on market conditions. Meanwhile ensemble strategies frequently ranked among the top-performing configurations, often matching or exceeding the performance of adaptive weighting schemes. Performance was strongly index-dependent, with S&P 500 and NASDAQ-100 exhibiting comparatively stronger profitability, whereas FTSE and Dow Jones showed weaker and less differentiated results. These findings emphasize that statistical accuracy (e.g., RMSE, R^2 metrics) alone is insufficient for profitable trading, underscoring the importance of financial performance metrics such as total return, drawdown, and risk-adjusted measures when evaluating predictive models. |
