Season means reduce the aggregate depth forecast error for 193 selected Kepler signals in Q13–Q16

We tested whether four Kepler season means predict later catalog transit depths. We fitted Q1–Q12 and compared Q13–Q16 errors with constant and linear models. The saved rules select 193 signals at 186 stars with confirmed labels in the NASA cumulative KOI table. We give equal weight to quarter means. The primary score is the geometric mean RMSE ratio, with equal weight for each star. Season means give a ratio of 0.885 against a constant, with a conditional 95% host-bootstrap interval of 0.816–0.956. Only 89 of 193 signals improve, and the median ratio is 1.037. A planned secondary rule fits Q1–Q8 and requires at least 20% smaller error in Q9–Q12. It selects 44 signals at 44 stars. Of these, 35 improve in Q13–Q16, and the aggregate ratio is 0.512, with interval 0.419–0.624. The full-sample ratio against a linear trend is 0.742. Five fixed changes to filtering and weights retain a lower aggregate error against a constant. Independent numerical checks pass, and a full offline rerun gives the same numerical result files. The package contains code, provenance, frozen object labels, predictions, exclusion records, sensitivity results, and figures. A download step checks the external source table hashes. This is a retrospective test of derived measurements. The source transit template, some flags, and current object labels use later data. Shared spacecraft effects and sample bias are outside the bootstrap intervals. The result adds a reproducible prediction benchmark for known seasonal depth effects. It does not establish a new planet or a physical cause.

Kepler · transit depth · seasonal variation · prediction · exoplanets · reproducibility