Recent incoming messages improved recipient forecasts among known contacts at observed CollegeMsg send times

We tested recipient forecasts in CollegeMsg, a record from one college online community in 2004. The sender and send time were given. Possible recipients were contacts seen strictly before that second. Earlier events fit and selected the models; later events tested them. On 11,894 messages from 696 senders, adding recent incoming history to frequency and outgoing-history features reduced mean log loss from 2.435 to 1.786 nats, a decrease of 26.65%. Accuracy rose from 44.36% to 67.02%. The gain was positive in four test-time blocks, with equal sender weights, after removal of the largest sender, and with other recency decay times. A post-plan rule that chose the latest incoming contact reached 68.96% accuracy, above the fitted model. Thus, incoming history carried useful information, but the fitted model was not the best top-choice predictor. The test covered 79.56% of later source messages. It excluded new contacts and did not predict message times. A fresh canonical download matched the retained mirror. Independent feature and score checks passed. A full repeat matched all feature arrays and four numeric files. The package preserves code, source provenance, aggregate results, figures, and reproduction steps. Source event bytes are excluded because no permissive data licence was confirmed. Prior work already studies reciprocity and event history. This contribution adds a scoped forecast benchmark; it does not identify causal replies or a general human behavior law.

temporal networks · recipient prediction · human communication · reciprocity · CollegeMsg · event history · reproducibility