timesift (Q7368801)

Fits and compares representations of time-varying data against a prediction target. Given a table of targets and a table of time-stamped series belonging to them, it builds each candidate representation, from the record unreduced through a calendar grain such as a week or a month to a lookback anchored on each target, fits the requested learners on each, scores every candidate on one set of held-out folds, and stacks the out-of-fold predictions into an ensemble. Calendar-aware binning keeps a bin a real week or month rather than a fixed block of hours. Learners, response heads and metrics are registered rather than hard-coded, so adding one is a registration and not a fork of the fitting code. The penalised baseline is an elastic net fitted by cyclic coordinate descent along a warm-started path, following Friedman, Hastie and Tibshirani (2010) <doi:10.18637/jss.v033.i01>. The shipped default is presence-absence with a joint multi-label head, the setting used for species distribution modelling from microclimate loggers, scored by the area under the receiver operating characteristic (ROC) curve with the true skill statistic of Allouche, Tsoar and Kadmon (2006) <doi:10.1111/j.1365-2664.2006.01214.x> beside it.