Machine Learning in Economic Forecasting: Integrating Traditional Methods with a Tunable LSTM

The surge of digital data in tourism, finance and consumer markets demands predictive models capable of handling volatility, nonlinear dynamics, and long-term dependencies, where traditional econometric tools often fall short. This study makes two contributions. First, we benchmark four classical machine learning methods-K-Nearest Neighbors (KNN), Reinforcement Learning (RL), K-Means clustering, and Principal Component Analysis (PCA)-to establish their strengths and limitations in economic applications. KNN provides accurate cancelation predictions but lacks sequential awareness; RL adapts dynamically, yet suffers from long-horizon instability; K-Means reveals static consumer clusters but cannot capture temporal shifts; and PCA condenses macroeconomic indicators while discarding dynamic structure. Second, we extend this comparison with an in-depth ablation study of a Long Short-Term Memory (LSTM) framework, systematically varying activation functions, loss functions, and training regimes. This analysis reveals how architectural design governs the accuracy, robustness, and sensitivity of rare events forecasting, and shows that LSTM addresses the shortcomings of classical models by learning temporal dependencies while remaining tunable between tasks. Across baselines, KNN achieves AUC = 0.81 in cancelation classification, while the tuned LSTM (Huber loss + Sigmoid head) achieves MAE ≈ 11.3 and Directional Accuracy (DA) ≈ 0.68, outperforming static models in both magnitude error and trend capture. Collectively, our findings provide practical guidelines for choosing between interpretable classical baselines and adaptive deep learning architectures in dynamic economic and tourism environments.