Noise assumptions change the sign of the season prediction gain for HD 114853 at 74.0266 days

Prior surveys report a 74-day velocity signal and different amplitudes before and after 2015 in HD 114853. We test whether separate era sine coefficients improve predictions of complete omitted observing seasons. We rebuild 135 nightly means from 145 public HARPS spectra. We hold two published periods fixed. We test six Gaussian noise settings, with and without a common drift. At 74.0266 days, a 1.5 m/s white noise term gives a total prediction log score gain of +6.394 for the separate fits. Adding a 2 m/s exponential noise term with a 10-day decay time changes the gain to -0.788. This sign-reversing noise setting lowers the full-data split-model log likelihood by 15.147. The settings do not have equal support. The last season has 53 nights and supplies +4.979 of the white-noise gain. At the second period, 74.5271 days, all 12 tested gains are positive, from +0.004 to +13.939. The local best period also shifts from 74.0535 to 74.5462 days when a drift enters the model. Noise and period assumptions control the evidence from this stability test. We do not estimate the noise parameters or show that correlated noise fits best. The test integrates linear coefficient uncertainty, but periods and covariances stay fixed. The star and periods were selected from the same historical data. The score sums are conditional prediction comparisons, not Bayes factors or planet false-alarm probabilities. This work does not establish or reject a planet. We preserve source rows, provenance, code, all model results, figures, and offline reproduction instructions.

HD 114853 · HARPS · radial velocity · stellar activity · predictive validation · correlated noise · reproducible astronomy