MIRAGE v6 DEPLOYED
Per-Map Match Outcome · Per-map Gradient Boosting

Eight separate models. One for every CS2 map.

Mean AUC (8 maps)
0.615
held-out test set

About this Model

MIRAGE is a collection of eight independent LightGBM models — one per active CS2 map. Each model is trained exclusively on historical results for its specific map, capturing map-specific team specializations that a single global model misses. Features include map Elo history, CT/T win rates with economy stratification, player lineup quality (CT/T per-player rolling win rates from demo data), and side-specific momentum. MIRAGE v6 introduces lineup-level features: per-player rolling CT/T win rates aggregated to team level, enabling detection of individual map specialists.

Top Feature Drivers

Map-specific Elo differential 28%

Per-map team Elo from match history (elo_before — no leakage)

Overall Elo differential 18%

Baseline quality gap (map-independent anchor)

CT full-buy win rate differential 14%

Which team dominates CT rifle rounds on this map

T execution rate delta 11%

Bomb-plant conversion gap — T-side aggression edge

Lineup CT mean win rate 9%

Average CT win rate of each player's history on this map

SoS-adjusted CT/T win rates 8%

Opponent-quality-weighted side win rates

Eco management delta 7%

CT eco vs T eco efficiency gap

CT anti-execute win rate 5%

CT wins on T full-buys — defensive specialist signal

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

Per-Map AUC Breakdown

nuke
0.638
dust2
0.63
inferno
0.627
ancient
0.611
overpass
0.613
mirage
0.607
anubis
0.577
vertigo
N/A

AUC measured on chronologically held-out test split per map.

How to Interpret

MIRAGE win probability for a specific map tells you how likely each team is to win that exact map based on all available historical signal for that map only. Use the per-map scores during veto analysis: a large MIRAGE edge (>10pp) on a map means it's a priority pick for the stronger team and a priority ban for the weaker one. Nuke currently has the strongest discriminative power (AUC 0.638); anubis is noisiest.

No settled predictions recorded for this model yet.

Live Performance

No live predictions recorded yet.

Selector Role

Sub-model — per-map win probabilities feed into TITAN and the veto advisor.

Version History

v2 2026-Q1

First per-map LightGBM; mean AUC 0.603

v3 2026-Q2

+CT/T side win rates from demo_rounds; mean AUC 0.608

v4 2026-Q2

+per-map regularization, exp-decay CT/T, auto-pruning; mean AUC 0.612

v5 2026-Q2

+economy buy-type stratification; nuke AUC 0.641 (best ever)

v6 2026-Q3

+player lineup CT/T win rates; ancient/overpass new bests; mean AUC 0.615

All Models