An exposure phase error model increased mean later log loss in nine screened Kepler timing records

We tested whether exposure phase improves error forecasts for local Kepler transit timing contrasts. The parent sample is the 15 suspected sampling cases from Szabo et al. (2013). Ten have Holczer et al. (2016) timing data. Nine meet fixed count rules for separate groups of three consecutive transits. Each contrast cancels a linear ephemeris. Recorded times and cadence numbers from 168 Data Release 25 files set phase at the expected transit date. Quarters 0 to 8 supply 595 fit groups. Quarters 9 to 17 supply 678 test groups. A constant variance multiplier and a phase dependent multiplier use the same early fitted timing mean. The phase model has two sine and cosine harmonics and a fixed coefficient penalty. It increased equal object mean Gaussian log loss by 0.04290 nats per later contrast. The descriptive 95 percent quarter block interval was 0.01313 to 0.07669 nats. Two of nine objects improved. Mean coverage of nominal 95 percent intervals was 92.83 percent for the constant model and 91.39 percent for the phase model. A zero mean check also favored the constant model. With catalog outlier rows retained, the mean difference was unstable. The primary result is conditional on full mission catalog screening and the specified fit. It does not show that phase has no effect or establish a physical cause. The package preserves source provenance, code, input data, decisions, predictions, simulations, figures, and repeat instructions.

Kepler · transit timing · error calibration · exposure phase · local timing contrasts · later data test · reproducible astronomy