Two direction measures reduced 14-day loss by 0.14% in email and raised it by 0.46% in college messages for tree models selected on past data

We tested whether two pair-direction measures improve forecasts of a later recorded event. We used public institution email and CollegeMsg records. Each main row had at least two events in a 14-day history. The target was any pair event in the next 14 days. Tree models used traffic count, timing, older pair history, and endpoint activity. Models were selected on past data and kept fixed for later tests. Adding a two-direction indicator and direction balance reduced email log loss from 0.564441 to 0.563675 across 24,152 test rows, a 0.14% reduction. It raised college-message loss from 0.496398 to 0.498675 across 1,148 rows, a 0.46% increase. Combined person and time weights gave descriptive 95% gain ranges of -0.000502 to +0.001986 and -0.011538 to +0.006842 nats per row. These ranges keep fits fixed and are not validated population confidence intervals. All planned email point gains were positive and below 0.32%. College signs changed across checks; fixed seven-leaf models gave a small positive gain. The tests do not establish an absent benefit or a general cost of direction information. A source audit excluded the email tail after a 271-day empty interval. A canonical download confirmed the retained CollegeMsg bytes. The package gives code, source hashes, all aggregate results and resampling draws, figures, checks, and repeat instructions. Source events remain external under their original rights.

communication networks · tie return · reciprocity · temporal prediction · email · CollegeMsg · reproducible benchmark