Priority for repeated contacts gave mixed gains at a 10 percent forecast limit in three school campaigns

We tested whether priority for pairs seen on both of two past school days improves forecasts of contact on the next observed school day. We used three public sensor records from preparatory classes at one school in France. Each rule selected an expected 10 percent of pairs seen in the two input days. We compared priority for repeated contacts, with total time as a tie rule, against total time alone. Mean precision changed from 64.58 to 65.63 percent in 2011, from 59.21 to 58.84 percent in 2012, and from 69.31 to 74.06 percent in 2013. There were two, five, and three target days. Six of ten days had gains. The sign and size of the change depended on the selection limit and target contact length. A simple recent-contact rule had higher mean precision than the repeat rule in two campaigns. Between 36.21 and 56.45 percent of next-day pairs were absent from the two input days. These records do not establish a gain across schools or a disease-model benefit. The package has code, source hashes and retrieval instructions, aggregate results for every setting, the analysis plan, figures, and numerical checks. Restricted source data stay at the original provider and in retained storage under their source terms. Fresh downloads gave an exact replay of the analysis tables.

contact forecasts · temporal networks · school contacts · repeated contacts · reproducibility · selection budget