This paper presents an applied study of the problem of short-term forecasting of the daily number of requests to a computer service center. For a two-year time series covering 2023–2024, calendar, lag-based, and meteorological features were constructed, and a hybrid ensemble of base models ˗ Ridge, ElasticNet, SVR, Random Forest, XGBoost, and LightGBM ˗ was implemented. To prevent data leakage, strict temporal feature shifting, logarithmic transformation of the target variable, and a TimeSeriesSplit scheme with five splits were employed. Ensemble weights were optimized as a constrained optimization problem using SLSQP, Particle Swarm Optimization (PSO), Differential Evolution (DE), and Simulated Annealing (SA). The best performance was achieved on the baseline feature set: the hybrid ensemble with PSO-based weight optimization reached RMSE = 2.0847, MAE = 1.5988, MAPE = 20.62%, and R² = 0.6315. Compared to the best single model in terms of MAPE (SVR, 21.29%), the improvement amounted to 0.67 percentage points. It is shown that increasing feature space complexity through trigonometric time encoding and additional derived features does not improve generalization performance. The obtained results confirm the practical applicability of compact ensembles for forecasting noisy operational demand affected by chaotic factors.