Cross-Layer Predictive Resource Allocation and Slicing in 5G Systems: An Evidence-First Re-Examination of the Time-Scale, Drift, and Imbalance Gaps

Abstract

Ultra-Reliable Low-Latency Communication (URLLC) slices in 5G networks remain vulnerable to transient packet drops. Predictive-scheduling proposals in this space typically assume (i) that physical-layer indicators such as RSRP/RSRQ directly signal impending drops, (ii) that rare-event imbalance in drop labels is uniform across traffic classes, and (iii) that a sequence predictor’s forecast horizon can gate a scheduler without an explicit time-scale reconciliation. We test all three against a 5,000-record, 20-cell, 3.4-day multi-slice 5G KPI trace covering eMBB, URLLC, mMTC, and a healthcare-oriented (HC) slice, and independently audit our own modeling pipeline for the same class of unverified assumption. Assumption (i) is rejected: packet loss correlates near-zero with RSRP/RSRQ, linearly and by rank. A pooled correlation between load/latency and packet loss (r = 0.53, r = 0.50) looks like a replacement causal story but collapses to near-zero (r = 0.01, r = −0.02) once slice identity is partialled out, since active-user count is itself almost fully determined by slice (R2 = 0.99) – a compositional artifact, not a within-slice mechanism. Assumption (ii) is partially rejected: SLA-violation imbalance is slice-conditional (1.2%–14.3%), not uniform. Our own pipeline audit found two further errors: every model had been denied the target row’s own slice/cell identity, a scheduling label known in advance, and four GRU configurations had shared one uncontrolled randomnumber stream. Correcting both changes the empirical picture substantially: a gradient-boostedtree baseline’s regression error falls roughly fivefold (365.7 to 73.0 Mbps) and now leads every GRU configuration; a pooled-α GRU reverses an earlier, confounded reading to clearly outrank its slice-conditional counterpart at the pooled level (ROC-AUC 0.70 vs. 0.58), while on the one adequately-sampled slice (URLLC) the point estimate favors slice-conditional weighting, a reversal a bootstrap test cannot confirm. Feeding the corrected forecast – now with real tracking skill (r = 0.24, up from r = −0.01) – into a Lyapunov-inspired scheduler simulation yields a mixed result: latency improves (−3.0%) while drop rate still worsens (+39.9%) relative to a reactive baseline, both moving toward or past parity as the boost coefficient increases. A leave-one-cell-type-out probe now shows real transfer, explained entirely by slice identity being available as context, while the harder leave-one-slice-out probe remains at chance by design.We report both the scientific corrections and the two pipeline bugs that produced them, on the view that disclosing a wrong intermediate result is more useful than a clean narrative that never had one.