Machine learning is increasingly applied in business analytics to support predictive decision-making for client classification and sales forecasting. This study evaluates multiple models on a synthetic financial dataset that preserves the statisticalstructure, feature relationships, and business logic of real-world enterprise sales data, ensuring full privacy and reproducibility. A unified preprocessing and feature engineering pipeline enables fair comparison across tasks: multi-class client categorization (with business-priority threshold tuning), weekly sales forecasting for operational decisions, and monthly revenue forecasting for strategic planning. Baseline, classical, regularized, ensemble, and decompositional time-series models are compared.Hyperparameters are tuned using Optuna, with performance evaluated using recall-focused metrics for classification and MAE/RMSE/MAPE for forecasting, supplemented by SHAP interpretability analysis. Results show optimized XGBoost as the strongest classifier (after threshold adjustment for priority classes), ElasticNet outperforming others for weekly horizons, and Prophet delivering the best monthly predictions. The work highlights systematic model comparison, interpretability, and practical applicability in privacy-constrained enterprise settings, offering a reproducible reference for similar business analytics applications.
Asma Bouach (Wed,) studied this question.