CLUTCH v1 DEPLOYED
Clutch / 1vX · Gradient Boosting (sequential)

When it's 1vX — who comes through?

AUC
0.877
held-out test set

About this Model

CLUTCH specialises in 1v1, 1v2, 1v3, 1v4, and 1v5 clutch situations parsed from 472,000+ pro CS2 rounds. It predicts clutch success probability from the kill sequence state: how many enemies remain, which weapons each side holds, and kills accumulated before the clutch entry point. AUC 0.877 on held-out data. CLUTCH identifies players whose clutch performance significantly exceeds model expectation — the real specialists under pressure.

Top Feature Drivers

Enemies remaining 38%

1v1 vs 1v5 has dramatically different base rates

Kills before clutch entry 26%

Momentum — player already fraggin or cold entry

Weapon advantage 20%

Clutching with AWP vs pistol changes everything

Economy state (clutcher) 10%

Can the clutcher full-buy the next round if they win?

Map position indicator 6%

Some maps have clutch-favorable positions

Feature importances reflect normalized split gain across the training set. Values are approximate — the model uses all features simultaneously.

How to Interpret

CLUTCH probability is the model's estimate for a 1vX outcome given the current game state. A 1v2 clutch typically has ~30% base rate; a player who wins 45%+ is a genuine clutch specialist. The leakage fix (v1 rewrite) is critical: the original 0.943 AUC included post-game information — the real number (0.877) is more honest and still exceptionally strong.

No settled predictions recorded for this model yet.

Live Performance

No live predictions recorded yet.

Selector Role

Live sub-model — activates during 1vX situations; feeds into SPECTER.

Version History

v1 (rewrite) 2026-Q2

Fixed leakage: removed total_round_kills (post-game); AUC 0.943→0.877 (genuine)

All Models