Agentic Model Selection and Tuning for Probability-Informed Network Traffic Forecasting

Mobile network traffic forecasting requires handling heterogeneous traffic regimes where no single predictive model performs optimally across all data segments. The probability-informed machine learning (PIML) framework addresses this by clustering training data with a Deep Gaussian Mixture Model (DGMM) and fitting separate regressors per cluster, but it leaves model selection and hyperparameter configuration fixed. We propose LM-Orchestrator, an agentic extension in which a local Small Language Model (SLM) autonomously selects and tunes the regressor for each cluster via a ReAct-style tool-calling loop with walk-forward cross-validation. Experiments on hourly network traffic data across six SLMs (3B–27B parameters) show that larger models achieve up to a 22.91% relative reduction in Mean Absolute Percentage Error over the best standalone estimator, while producing interpretable orchestration decisions unavailable in black-box AutoML approaches.