Advances in Time Series Forecasting Across Domains

Abstract

Time series forecasting plays a critical role in predictive analytics, enabling accurate modeling of temporal data across various domains such as finance, energy, healthcare, and climate science. This study explores the advancements in forecasting techniques by systematically comparing traditional statistical models (ARIMA, SARIMA), machine learning algorithms (Random Forest, Gradient Boosting, SVR), and deep learning architectures (LSTM, GRU, Transformer). Using domain-specific datasets spanning 5–10 years, the research employs rigorous preprocessing, feature engineering, and model optimization to evaluate performance based on metrics such as MAE, RMSE, MAPE, and R². The results reveal that Transformer and GRU models consistently outperform conventional methods, achieving higher accuracy and robustness in capturing nonlinear and long-term dependencies. Cross-domain transfer learning analysis demonstrates strong generalization capabilities, with the Transformer model retaining over 95% of its predictive efficiency when adapted across domains. Feature importance analysis using SHAP and LIME confirms the dominance of temporal and cyclical variables in driving model accuracy, while ANOVA and Diebold-Mariano tests validate the statistical significance of performance differences. The radar chart, cluster dendrogram, and correlation heatmap provide visual insights into model clustering, adaptability, and metric interrelationships. Overall, the study highlights that deep learning architectures, particularly Transformer networks, mark a paradigm shift in time series forecasting, offering scalable, interpretable, and domain-independent solutions. These findings underscore the transformative potential of AI-driven forecasting frameworks in advancing decision intelligence and predictive analytics across industries.