Sequential drift fits gave less timing error than rolling lines for 38 of 49 selected Kepler objects

We tested whether a local timing drift model improves prediction of the next observed Kepler transit. A standard integrated-Wiener model was fitted to the first 60% of each catalog time span. Process and error scale parameters were then fixed. Each later measurement updated the state only after its prediction was saved. The sample had 929 eligible confirmed objects and candidates from the Holczer timing catalog. An early likelihood rule selected 49 objects at 47 stars, with 5144 later targets. Drift error was smaller than an early-chosen rolling weighted line for 38 of 49 objects. The median paired weighted error ratio was 0.973, with a 95% host-star bootstrap interval of 0.951 to 0.984. The geometric mean ratio was 0.943; its interval included one. Against an expanding line, 31 of 49 objects improved and the median ratio was 0.878. Nominal 95% drift intervals covered 93.2% of later measurements, with objects weighted equally. Nominal 68.27% intervals covered 76.2%. Strict quality cuts, a 70% early split, confirmed-only objects, a finer grid, and gaps of at least 10 days are reported. The result supports a modest median gain against the tested rolling line in the selected sample. It does not establish calibrated phase intervals or a new planet. This is a retrospective measurement test: source extraction and eligibility use full-record information, and observation variances use target catalog errors. The package preserves source numeric data, code, all later predictions, provenance, and offline reproduction checks.

Kepler · transit timing · sequential prediction · Kalman filter · integrated Wiener process · prediction intervals · reproducible benchmark