Climb Intel · Applied ML · 2026

Predicting climbing difficulty
from hold geometry.

A stacked gradient-boosted ensemble grades Kilter Board routes across every wall angle — calibrated, evaluated without leakage, and aware of the climber's body.
Shain Carrasco · [email protected]
137K
Graded routes
86
Engineered features
%
Within 2 grades
<50ms
Live prediction
The Approach
1
Ingest & weight 137K routes
Synced from the Aurora API and BoardLib; each route consensus-weighted by log-ascents × quality so well-validated grades dominate the noise.
2
Engineer 86 features
Hold geometry, move sequence, and angle interactions, plus body pose imputed from 1.9M MediaPipe frames (33 landmarks each).
3
Stack & augment
XGBoost (900 trees) + LightGBM blended by a Ridge meta-learner — 357K training rows with mirror/reversal augmentation and hard-grade oversampling.
4
Calibrate & personalize
Isotonic calibration to a V-grade with a q10–q90 uncertainty band, then re-scored for the climber's height, ape index, and weight.
Accuracy by Grade
V0
75.3%
V3
72.1%
V4
72.5%
V5
69.6%
V6
67.5%
V8
58.6%
V9
51.0%
Within-one-grade accuracy is highest on the core grades V4–V6, where most climbing happens. The hardest grades (V9+) are the open problem — there simply aren't enough routes to learn from yet.
Under the Hood
Stacked GBM ensemble Ridge meta-learner Isotonic calibration q10–q90 quantile bands Consensus weighting Mirror + reversal augmentation Hard-grade oversampling Leakage-guarded frozen holdout Cross-model leaderboard Validation-gated retraining MediaPipe pose imputation Anthropometric personalization Per-board-size adaptation Autonomous scrape → retrain

Measured honestly, without leakage.

A Kilter grade is really a crowd average, and even expert setters routinely disagree by a full grade. So the fair question isn't "exactly right" — it's how close. Tested on routes held out of all training, % of predictions land within that one-grade margin of human disagreement.

Exact
%
±1 grade
%
±2 grades
%
What Drives a Prediction

Difficulty really is predictable from geometry. Board angle and its interactions dominate every other signal — the same model powers a live route creator, body-aware personal grades, and video pose coaching, and only ships an update when it beats the current model on held-out routes.

Board angle
Angle × reach
Wall steepness
Hold density
Angle × dyno
Relative feature importance by model gain
CLIMB INTEL · 2026 Stacked GBM · pose-aware · isotonic-calibrated Evaluated on held-out routes