Prediction Horizon

The prediction horizon specifies the exact future time step for which a model generates its output. In RAN load balancing, this can range from 10 milliseconds for near-RT RIC control loops to 24 hours for energy-saving sleep mode scheduling. The horizon is not a duration but a discrete point: a model predicts the load at t+100ms, not over the next 100ms. This distinction is critical for aligning forecasts with the execution latency of the actuating system.

Forecast accuracy degrades as the prediction horizon extends, following a non-linear decay curve. Key factors influencing this degradation:

• Channel coherence time: In high-mobility scenarios, CSI predictions become unreliable beyond a few milliseconds
• User behavior entropy: Individual traffic patterns are less predictable over longer windows
• Concept drift rate: The speed at which underlying statistical relationships change in the network Optimal horizon selection balances proactive action lead time against mean absolute percentage error (MAPE) thresholds acceptable for the use case.

Modern architectures often employ multi-horizon models that output predictions at several future time steps simultaneously. A single Transformer-based forecaster might predict PRB utilization at t+100ms, t+1s, and t+10s. This enables hierarchical decision-making:

• Short horizon (10-100ms): Near-RT RIC xApps for instantaneous resource scheduling
• Medium horizon (1-10s): Inter-cell load shifting and handover parameter adjustment
• Long horizon (minutes-hours): Energy-saving cell sleep/wake orchestration and capacity planning

The prediction horizon is intrinsically linked to the lookback window—the length of historical data fed into the model. A common heuristic is that the lookback window should be 3-5x the prediction horizon to capture sufficient temporal dependencies. For a 1-second horizon, a 3-5 second lookback is typical. However, LSTM and Transformer architectures can learn long-range dependencies, allowing shorter relative lookbacks. Poorly matched horizon-lookback ratios are a primary cause of underfitting in RAN forecasting models.

Different RAN optimization functions demand distinct prediction horizons:

• Mobility Load Balancing (MLB): 1-10 seconds—sufficient to adjust CIO parameters before congestion
• Dynamic Spectrum Sharing: 10-100 milliseconds—must react within channel coherence time
• Energy-saving sleep modes: 15-60 minutes—aligns with traffic tidal patterns
• QoE-aware traffic steering: 500ms-2 seconds—matches video buffer and TCP dynamics Selecting the wrong horizon leads to either premature action (oscillation) or delayed reaction (congestion).

Advanced systems implement dynamic horizon adjustment where the model or orchestrator modifies the forecast horizon in real-time based on:

• Network load volatility: Higher variance triggers shorter horizons for responsiveness
• Mobility state: High-speed UEs require shorter horizons due to rapid cell transitions
• QoS class: URLLC slices demand sub-10ms horizons; eMBB slices tolerate longer windows This adaptation is often governed by a meta-controller that monitors prediction residuals and adjusts the horizon to maintain a target confidence interval.

The fixed length of historical time-series data used as input for a forecasting model to make a single prediction. The lookback window defines the temporal context the model can observe and must be carefully tuned in relation to the prediction horizon.

• A longer lookback captures seasonal patterns but increases computational cost
• A common heuristic is a lookback-to-horizon ratio of 3:1 or 5:1
• For a 10ms prediction horizon, a 50ms lookback window is typical in Near-RT RIC applications

A statistical and machine learning methodology for predicting future values of a metric based on previously observed sequential data points. The prediction horizon is the defining output parameter of any forecasting model.

• Short horizons (ms to seconds): Used for real-time PRB scheduling and beamforming
• Medium horizons (seconds to minutes): Used for predictive load balancing and handover optimization
• Long horizons (hours to days): Used for capacity planning and energy-saving mode activation

A phenomenon in online learning where the statistical properties of the target variable change over time, rendering the model less accurate. The risk of concept drift increases proportionally with the length of the prediction horizon.

• Sudden drift: A flash mob event causing an abrupt traffic spike
• Gradual drift: Seasonal changes in user mobility patterns over weeks
• Mitigation requires continuous monitoring and shorter horizons for volatile environments

The implementation of predictive load balancing logic as an xApp running on the Near-Real-Time RAN Intelligent Controller. This architecture enforces a prediction horizon constraint of 10ms to 1 second for control loop execution.

• xApps must complete inference within the horizon window to be actionable
• The E2 interface provides the streaming telemetry needed for short-horizon predictions
• Longer horizons are delegated to the Non-RT RIC for policy guidance

The specific application of predictive algorithms to forecast the future resource utilization and user demand on an individual cellular base station. The prediction horizon directly determines whether the forecast is used for reactive or proactive resource allocation.

• A 100ms horizon enables just-in-time PRB allocation
• A 5-second horizon enables seamless inter-cell load shifting before congestion
• A 15-minute horizon enables energy-saving cell sleep mode activation

A machine learning model that continuously updates its parameters incrementally as new streaming telemetry data arrives. This approach is essential for maintaining accuracy over extended prediction horizons without full retraining.

• Adapts to changing traffic patterns in real-time
• Eliminates the cold-start problem for newly deployed cells
• Requires robust model drift detection to prevent degradation over long horizons