The evolution towards sixth-generation (6G) and vehicle-to-everything (V2X) networks requires precise quality-of-service (QoS) levels, making network behavior prediction crucial for applications like autonomous driving. While machine learning (ML) is a promising solution, its performance is highly dependent on hyperparameter tuning, a complex and computationally expensive challenge. This study proposes a rigorous methodology for tuning Random Forest (RF) hyperparameters for downlink throughput prediction in vehicular communication scenarios using the Berlin vehicle-to-everything (V2X) dataset, collected in West Berlin (Germany). The core of the methodology is a novel statistical approach named Analysis of Variance and Scott-Knott (ANOVASK), which utilizes Analysis of Variance (ANOVA) and the Scott-Knott clustering algorithm to systematically rank combinations of eight RF hyperparameters. Two distinct ANOVA models: a one-way and a three-way factorial design were employed to assess the influence of hyperparameter values on Artificial Intelligence (AI) model performance. Results revealed that different hyperparameter settings lead to statistically significant variations in prediction accuracy. When compared with traditional methods, the models recommended by the ANOVASK algorithm outperformed both Random Search and Grid Search , with discrepancies not exceeding 2.2% in Mean Absolute Error (MAE), 5.1% in Root Mean Squared Error (RMSE), and 0.8% in Coefficient of Determination (R²). The proposed approach also demonstrated superior computational efficiency, requiring 14 to 16.4 h to obtain the models, compared to the 24 h taken by the other methods. Overall, the proposed methodology provides a robust, accurate, and efficient solution for RF hyperparameter tuning in vehicular networks.