Accurate classification of Bangla news has become a crucial need for effective information retrieval and automated content management, but research in this area is commonly affected by limited resources and uneven model performance. This research constitutes a systematic comparative study between machine learning algorithms in single-label and multi-class categorization of Bangla news in 5 domains: Economy, Entertainment, International, Sports and State. Using a curated balanced dataset comprised of 2,500 articles (500 per class), we set the question of reproducibility in our benchmarks by testing four classical models (Logistic Regression, Random Forest, K-Nearest Neighbors and Multinomial Naive Bayes), one lightweight LSTM, and two ensemble approaches (Voting and Stacking). Our results prove the superiority of the Stacking ensemble (including RF, NB and KNN with the Logistic Regression as meta learner) as it achieved the greatest accuracy of 91.20 % and the highest macro F1 score compared to all the single baselines and the Voting ensemble (90.00 %). Category-wise results show good performance (F1=0.85−0.96) with 'Sports' being most separable and 'State' being the most challenging, supported by further validation through ROC analysis (macro-AUC 0.97-0.99). These outcomes offer a benchmark for Bangla news classification using small data sets and underscore the effectiveness of ensemble learning in low-resource settings, the basis for pursuing transformer-based learning architectures in the future.