Machine Learning versus Deep Learning for Public Financial Report Classification and Popularity Prediction: A Feasibility Study on Ghana's Controller and Accountant-General's Department

Abstract Government agencies increasingly publish financial reports online, but small public-sector portals rarely have the labeled data needed to test automated classification tools. This study evaluates the feasibility of these tools using a corpus of 131 official reports from Ghana's Controller and Accountant-General's Department (CAGD). We compare classical machine learning models (Logistic Regression, Random Forest, and eXtreme Gradient Boosting [XGBoost] with Synthetic Minority Over-sampling Technique [SMOTE]) and character-level deep learning models (bidirectional Long Short-Term Memory [LSTM] and Text Convolutional Neural Network [TextCNN]) across three tasks: category prediction, binary popularity classification, and log-download forecasting. Models were evaluated using a stratified 70/15/15 train, validation, and test split, randomized hyperparameter search, and a final fine-tuning phase. The fine-tuned XGBoost classifier achieved 1.00 accuracy on the five-class category task. Logistic Regression reached 0.75 accuracy (Area Under the Receiver Operating Characteristic curve [ROC-AUC] of 0.80) on the binary popularity task, and Random Forest reached an R-squared of 0.85 on log downloads. An ablation study removing twenty-five keyword features dropped category accuracy from 1.00 to 0.80, a stable drop across five randomized data splits. This indicates keyword features inflated initial results, though remaining performance confirms the models learn non-tautological structure. This work serves as an honest feasibility check, showing that thorough feature ablation is essential before deploying these models on live government portals.