Abstract

This study presents a comparative analysis of six forecasting techniques ETS, ANN, ETS-ANN, SNR, LSTM, and K-NN using the well-known Air Passengers dataset, which consists of monthly measurements of U.S. air travel passengers across a twelve-year span (1949–1960). The analysis examines both the theoretical foundations and empirical performance of statistical, machine learning, and hybrid forecasting models. ETS demonstrates strong interpretability and effective modeling of trend and seasonality, while ANN and LSTM provide enhanced capabilities for capturing nonlinear patterns and long-term dependencies. The hybrid ETS–ANN model is designed to combine these strengths by modeling residual structures left unexplained by ETS, although its empirical performance depends on the presence of exploitable nonlinear patterns. K-NN offers a simple instance-based forecasting approach, whereas the SNR-based method incorporates noise-sensitive estimation. To identify the best-performing model, five evaluation criteria, ME, MAE, RMSE, MAPE, and MASE, were applied to forecast values for the year 1960. The results indicate that the ANN model consistently achieved the lowest error across all metrics, establishing it as the most accurate method in this study. In contrast, the SNR model exhibited the highest error values and performed poorly relative to the other approaches. Overall, the findings highlight the superiority of ANN for forecasting monthly airline passenger data and support the effectiveness of nonlinear machine learning techniques in time-series prediction.