Selecting periods in training data lowers median prediction gains for 12 selected HARPS signals with sine models and independent errors

We test how period selection changes predictions for 12 selected HARPS radial-velocity signals. Each host has one source orbit and at least four eligible observing seasons. We compare a published period, a sine period selected from all velocities, and a sine period selected within each training set. Every training set fits instrument-era offsets, a trend, and sine coefficients. Seasons use gaps above 90 days. Gain is one minus the ratio of summed signal and baseline test errors. With a 3 m/s independent error floor, median gain falls from 0.4667 for full-data selection to 0.1228 for training selection. The median paired decrease is 0.1909. The two searches give 12 and 8 positive gains, respectively. A denser grid changes one marginal sign. Error-floor tests change other signs. Six signals retain positive gains in all four planned settings and all added checks that apply. We repeat both searches in 128 noise trials and 128 coherent sine injections per star. Full-data and training selection give positive gains in 1507/1536 and 79/1536 noise trials. These controls do not repeat candidate selection. We verify all 2,904 primary source rows and rebuild all 1,984 selected nightly inputs. The package contains code, provenance, all folds, control velocities, results, and figures. Earlier candidate selection, cuts, and calibration still use full-series information. The results apply to the stated sine models and independent errors. They do not confirm planets or give candidate false-positive probabilities.

HARPS · radial velocity · period selection · prediction error · planet candidates · public data