Where 5G Throughput Prediction Actually Fails: A Transition-Aware, Leakage-Audited Evaluation on Real Drive-Test Traces

Machine-learning models that predict short-horizon cellular throughput underpin adaptive video streaming, cloud gaming, and edge orchestration. On the widely used Raca et al. 5G drive-test dataset, such models can report low error. We show that, on this dataset, these headline numbers can be optimistic and, more importantly, that aggregate metrics conceal where and how the models fail. Using the saturated file-download sessions, where throughput reflects deliverable capacity rather than application idling, we (i) quantify the optimism of random-row evaluation relative to a deployment-realistic leave-sessions-out (LOSO) protocol, negligible at a 1 s horizon but reaching ∼32% at 8 s, and show that a trivial persistence baseline significantly beats the learned models at the two shortest horizons; (ii) show by ablation that most reported accuracy is carried by autoregressive throughput features rather than radio/context indicators and that the indicator-only mapping generalizes poorly across sessions; (iii) find, contrary to the intuition that handovers dominate error, that absolute error is lower around transitions and concentrated in the high-throughput regime (Gini 0.71), but that, in scale-invariant terms, the error is largest at low throughput and that transitions carry only a modest, noisy under-prediction tendency (event-aligned trough ≈$-6$ Mbps) that nearly vanishes under linear-space training (≈−1 Mbps), i.e., largely a log-space retransformation artifact; and (iv) find, using paired per-session bootstrap tests, that lightweight transition-context features give no significant global gain, a (mostly) null result. All headline numbers carry per-session bootstrap confidence intervals; the self-lag finding further replicates on an independent HSDPA drive-test dataset from Australia. We recommend leave-sessions-out evaluation, a persistence baseline, self-lag ablation, and scale-invariant, bias-aware reporting as a minimal honest protocol.