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

It contains the analysis code and all derived artifacts (processed 1 Hz features, out-of-fold predictions, metric tables, and publication figures) needed to reproduce every number, table, and figure in the paper. The study audits how machine-learning models for short-horizon 5G downlink-throughput prediction are evaluated: the optimism of random-split versus leave-sessions-out evaluation, the dominance of autoregressive self-lag features, where error concentrates under absolute versus scale-invariant metrics, and whether lightweight transition-context features help (a null result).

All headline numbers carry per-session bootstrap confidence intervals. This record does NOT redistribute the raw data; the pipeline regenerates the processed features from the public Raca et al. (2020) 5G production dataset (github.com/uccmisl/5Gdataset). Requirements: Python 3.12, CPU-only (numpy, pandas, scikit-learn, scipy, matplotlib, pyarrow); all randomness is seeded.