Effective climate change mitigation and adaptation strategies require robust, interpretable, and scalable forecasting frameworks. Northwestern Iran is highly vulnerable to increasing aridity, yet existing assessments often lack sufficient temporal resolution and predictive rigor. To address this gap, this study evaluates historical and future aridity dynamics using a modified UNEP aridity index (MUi), in which two additional classes were introduced to better capture extreme dry conditions. Monthly, seasonal, and annual MUi datasets were constructed for 16 synoptic stations covering 1967-2024. Five advanced forecasting approaches representing two fundamentally different modeling paradigms were developed and compared: three optimized machine learning models, Support Vector Machine with Pelican Optimization Algorithm (SVM-POA), Artificial Neural Network with POA (ANN-POA), and Bidirectional Long Short-Term Memory (Bi-LSTM), and two probabilistic state-based models, namely First-and Second-Order Time-Inhomogeneous Markov chains (FOI-Markov and SOI-Markov). Model performance was evaluated across temporal scales using multiple accuracy metrics, and the best-performing model at each scale was subsequently employed to forecast aridity conditions for the coming decade (2025-2035). Results demonstrate a clear scale-dependent model superiority. SVM-POA consistently outperformed other models at monthly and seasonal scales, while SOI-Markov and ANN-POA showed superior performance at the annual scale, highlighting the complementary strengths of machine learning and state-transition approaches. Trend analysis based on observational records (1967-2024) reveals a statistically significant drying tendency across most stations, while model-based projections (2025-2035) suggest a continuation of this tendency under the adopted forecasting framework. Overall, the findings provide a robust, multi-scale assessment framework and offer actionable insights for climate adaptation and water resource management in arid and semi-arid regions.