Writeup

Guess the rock along a drill path

×

The problem

A horizontal well is a hole that turns sideways through rock. TVT is a depth marker, true vertical thickness, that tells you which layer the bit is in. Guess it wrong and you steer into the wrong layer. The contest scores RMSE, root mean squared error, so lower is better.

Don't leak the well

I grouped the folds by well. If you split rows at random, the same hole appears in train and in test, and the model cheats by remembering that hole. GroupKFold keeps a well together. Five folds. LightGBM, CatBoost, and Ridge. A conservative gold profile, meaning I did not chase the loudest public number.

The well stays in one fold. That is the leak you are avoiding.

  1. 01The drill logeach well kept in one fold
  2. 02Three modelsLightGBM, CatBoost, Ridge
  3. 03Guess TVTlower RMSE wins
Drill log in, three models with well-grouped folds, TVT out.

The score

The LightGBM notebook scored 13.894 public and 13.880 private. A later notebook went to 15.923 / 16.374, which is worse. I kept the better one. Rank 4696 of 6125. The lower half of the board.

Later was worse. I kept the earlier number.

Two accepted RMSEs. LightGBM is the lower bar, which is the better one.

A title is not a score

There is a kernel titled ROGII [LB: 7.168]. That is a notebook name, not an accepted submission I can point to in the CLI. The accepted csvs I have are the 13.9 and 15.9 rows. I am not going to sell the title as a score.

What I won't pretend

What I would not claim. I did not invent well-log features. I did not beat the field. The useful bit is the split: treat a well as one object, not a bag of independent rows, and do not trust a later notebook just because it is later.

(competition)

The files

I put the files under the writeup. Take one, or take the zip.

writeup.md · 23 lines

# Guess the rock along a drill path
## The problem
A horizontal well is a hole that turns sideways through rock. TVT is a depth marker, true vertical thickness, that tells you which layer the bit is in. Guess it wrong and you steer into the wrong layer. The contest scores RMSE, root mean squared error, so lower is better.
## Don't leak the well
I grouped the folds by well. If you split rows at random, the same hole appears in train and in test, and the model cheats by remembering that hole. GroupKFold keeps a well together. Five folds. LightGBM, CatBoost, and Ridge. A conservative gold profile, meaning I did not chase the loudest public number.
## The score
The LightGBM notebook scored 13.894 public and 13.880 private. A later notebook went to 15.923 / 16.374, which is worse. I kept the better one. Rank 4696 of 6125. The lower half of the board.
### A title is not a score
There is a kernel titled ROGII [LB: 7.168]. That is a notebook name, not an accepted submission I can point to in the CLI. The accepted csvs I have are the 13.9 and 15.9 rows. I am not going to sell the title as a score.
## What I won't pretend
What I would not claim. I did not invent well-log features. I did not beat the field. The useful bit is the split: treat a well as one object, not a bag of independent rows, and do not trust a later notebook just because it is later.

Credits

KindWriteup
DateAug 2026
TagsKaggle, Geology, Trees