Eight of 13 selected Kepler timing curves give larger later errors with an early quadratic fit than with a linear fit

We test whether early timing curvature predicts later Kepler transit times. We use a verified subset of the Holczer 2016 catalog with one sequence per host. Of 921 sequences with sufficient coverage, 13 pass an early curvature-to-error threshold of five. Fits use times before BKJD 1000; forecasts use later times. Eight quadratic forecasts have larger weighted mean squared errors than linear forecasts. The median error ratio is 2.30, with a descriptive host bootstrap interval of 0.33 to 7.37. Nine late curvature estimates have the opposite sign. Annual terms and a residual cut keep eight larger errors, but a split at BKJD 900 gives five. The strongest negative early curve, KOI-525.01, has a known timing signal. A training-only sinusoid gives a later error close to the quadratic error. These tests do not establish orbital decay, a new planet, or a population failure rate. Correlated timing errors, selection, the short baseline and absent independent observations limit inference. Code, input rows, hashes, controls and figures are included.

astronomy · Kepler · transit timing · quadratic ephemeris · orbital decay · forecast validation