Accurate short-term forecasting of downlink bitrate is important for proactive radio resource management and service assurance in 5G systems. This work presents a strict empirical comparison of a classical statistical baseline (SARIMAX) against two deep sequence models LSTM and Transformer for forecasting downlink bitrate on synthetic traces generated via simulation in an srsRAN-based testbed. The dataset consists of 10,000-time steps with exogenous features reflecting radio conditions and link adaptation, including carrier frequency offset (CFO), MCS (DL/UL), packet loss (PL), and iteration index, with the target variable being the downlink bitrate. Models are trained to predict the next-step downlink bitrate using fixed-length sequences (48-time steps) for the neural approaches, while SARIMAX is configured with exogenous regressors and selected as (0,1,1) by information criteria. On the held-out test set, the deep learning models substantially outperform SARIMAX: SARIMAX achieves MAE 133.15, RMSE 178.95, MAPE 6.84%, and R² 0.862, whereas LSTM reaches MAE 12.82, RMSE 19.65, MAPE 0.80%, and R² 0.998, and Transformer achieves MAE 20.77, RMSE 25.78, MAPE 1.14%, and R² 0.997. These results indicate that sequence-based deep models capture nonlinear temporal dependencies and exogenous interactions more effectively than linear SARIMAX on srsRAN synthetic traces, providing a strong basis for predictive control in simulated 5G environments.