ROGII - Wellbore Geology Prediction

I entered this competition with Haruki Kasamatsu and Bossaerts Vincent. The task was to predict true vertical thickness (TVT) along horizontal wells. Each well has measured depth, coordinates, a gamma-ray (GR) log, a known TVT prefix, and a vertical reference log. Neighboring rows are not independent samples. They are points on one trajectory.

Validation

The center of the solution was well-grouped cross-validation. No well appeared in both training and validation, which avoids the leakage of a random row split. Seed 42 was for hypotheses and choices. After those choices were frozen, seeds 43 through 46 were used only for confirmation. We looked at the tail of well-level error and at geographic groups, not only mean RMSE, and we did not adopt a change that was unstable outside the mean.

The officially selected model

The officially selected model combines a GR reference built from horizontal wells in the same typewell group, a likelihood learned from GR at several scales, and a small posterior-mean correction after that pipeline. That submission has a Private RMSE of 8.044.

A submission whose scoring finished after the deadline

A later submission added hierarchical selection among candidate trajectories and spatial and geological corrections. Validation treated it as the stronger candidate, but scoring started at 23:10 UTC, ran past the 23:59 UTC deadline, and finished at 00:33 UTC, so it was not part of the official selection. The late candidates sat within 0.024 ft on the Public leaderboard and 0.585 ft apart on the Private leaderboard. We trusted the leakage-resistant validation over small Public differences.

The full solution is in the Solution Writeup.