Model Leaderboard
Prediction accuracy across premium tier matches · 1610/2751 settled
#1
65.6%
Nexus
META: 3-way consensus of Apex + Oracle + Eagle with avg confidence >55%.
86 ✓ / 131
45 wrong
#2
64.1%
Consensus
META: triggers when ≥75% of base models converge on same side.
59 ✓ / 92
33 wrong
#3
63.7%
Streak
Win/loss streak signal blended with Elo.
100 ✓ / 157
57 wrong
#4
63.4%
Vanguard
META: fires only when Elo and WR10 strongly agree (>60% same side).
26 ✓ / 41
15 wrong
#5
62.0%
Oracle
Conservative: Elo HQ dampened when streaks disagree.
85 ✓ / 137
52 wrong
#6
60.8%
Eagle
Blend: Elo + WR10 + form — weighted toward top-tier record.
124 ✓ / 204
80 wrong
#7
59.2%
Apex
Blend: Elo + form + streak — momentum-aware ensemble pick.
113 ✓ / 191
78 wrong
#8
58.9%
Fatigue
Dota-specific: penalty for back-to-back matches (<12h apart).
116 ✓ / 197
81 wrong
#9
58.5%
SOS
Strength of schedule: form weighted by opponent quality.
86 ✓ / 147
61 wrong
#10
57.8%
Roster Change
Penalty for unstable rosters — fresh lineups underperform Elo.
108 ✓ / 187
79 wrong
#11
57.3%
Elo H2H
Standard head-to-head Elo, scale=400.
114 ✓ / 199
85 wrong
#12
57.1%
Phantom
High-conviction: only commits when Elo + WR10 strongly agree.
40 ✓ / 70
30 wrong
#13
56.9%
WR10
Win-rate vs top-10 opponents, blended with Elo.
82 ✓ / 144
62 wrong
#14
56.6%
Elo HQ
High-confidence Elo: smaller K-factor, scale=2000 — slow learner.
69 ✓ / 122
53 wrong
#15
56.3%
Tier Elo
Elo computed only from matches against top-30 teams.
80 ✓ / 142
62 wrong
#16
56.2%
Form
Pure win-rate over last 10 matches.
86 ✓ / 153
67 wrong
#17
55.6%
Patch-Aware
Elo with patch-boundary decay — stale form loses weight after Valve patches.
84 ✓ / 151
67 wrong
#18
53.9%
Synergy
Hero co-occurrence winrates within recent team lineups.
83 ✓ / 154
71 wrong
#19
52.3%
Meta Strength
Team's recent hero pool weighted by current meta winrates.
69 ✓ / 132
63 wrong
Accuracy is measured per tournament tier — Premium is the clean, high-signal subset (mirrors cs2predict's top tier). Models with fewer than 20 settled predictions on the selected tier are calibrating. Coin-flip predictions (48–52%) are excluded. Click a model for confidence buckets, Brier score, and recent picks.