Time-Series Forecasting with SARIMAX for Intent Prediction

Abstract

By converting high-level user objectives into work able settings, intent-based networking makes autonomous and f lexible network administration possible. However, without the ability to proactively predict network intents across different temporal granularities, its full potential is still constrained. Using practical operational datasets, this study examines the use of SARIMAX (Seasonal AutoRegressive Integrated Moving Average with eXogenous variables) for multi-scale intent prediction. With a fixed 1-day forecast horizon, we perform extensive trials under two training regimes—8-day and 10-day historical data—across prediction windows of 5, 10, 15, 20, and 25 minutes. Our f indings show a consistent trade-off between granularity and stability: longer windows (20–25 minutes) produce smoother forecasts at the cost of increased lag and decreased sensitivity to sudden changes, while shorter windows (5–10 minutes) offer greater responsiveness to real-time fluctuations but are prone to noise. There were clear bias-variance trade-offs between the two training durations, with the 5-minute window achieving the lowest MAE and the 25-minute window minimizing RMSE. The 10-minute setup reliably balanced responsiveness and stability despite regularly high MAPE values (over 280%), making it op erationally appropriate for IoT service orchestration and intent driven 5G slice management. The study lays the groundwork for future machine learning and hybrid model integration to improve intent prediction in dynamic network contexts and emphasizes the crucial role temporal aggregation plays in forecast success.

Keywords