End-to-end latency remains a pivotal performance constraint in fifth-generation (5G) wireless systems, especially for Ultra Reliable Low Latency Communication (URLLC) services requiring sub-1 ms delay and 99.999% reliability. This study surveys quantitative advances from latency prediction to reduction in 5G networks. Analytical and learning-based techniques such as mixture models, Extreme Value Theory (EVT), and Deep Reinforcement Learning (DRL) are reviewed. Reported results indicate accuracy gains up to 85% in tail-latency prediction, 6.3 ms average delay reduction, and 31% packet-loss improvement across 22 studies. Numerical examples reveal that predictive latency modeling using Gaussian Mixture Extreme Value Models (GMEVM) achieves 99.9999th-percentile prediction within ±0.7% error, while optimization schemes reduce mean latency from 8 ms (4G) to <1 ms (5G MEC). This synthesis identifies open challenges in integrating probabilistic latency forecasting with adaptive control for deterministic 5G backhaul.