Adding a prior sender-change term did not lower test loss in fixed conditional models of two message records

We tested whether the previous change of sender improves the next sender forecast in Copenhagen SMS and a pinned CollegeMsg record. Each target was the next event on the same pair within one hour. Inputs used events through the current message only. We fitted fixed logistic models on the first half of elapsed time and tested the second half. The baseline used past time gaps and message counts. The added model also used the previous sender change. For 7,071 Copenhagen test events, log loss rose from 0.549891 to 0.549924 natural log units per event. The descriptive 95% three-day block range for baseline minus added loss was -0.000113 to +0.000093. For 1,322 CollegeMsg test events, loss rose from 0.415213 to 0.435351, or 4.85%. Its block range was -0.029306 to -0.009762. All planned model checks kept negative mean loss reductions. A later diagnostic found a large change over time in the CollegeMsg target rate after same-sender transitions. The result concerns one added term above a model that already uses history counts. It does not establish an absence of communication memory or a causal reply effect. The package contains code, source provenance and hashes, source retrieval instructions, aggregate scores, all primary block draws, figures, and checks. A separate feature calculation and fit agreed. A fresh-source repeat matched all 12 core numerical files. CollegeMsg event data are excluded because their reuse licence was not confirmed.

temporal networks · human communication · sender order · chronological prediction · conditional forecast · log loss · model comparison