We've added a new model to CS2PREDICT: SURGE. Unlike ELO, rank, or even SQUAD (which averages player ratings), SURGE looks at how fast individual players are improving — capturing a "hot-hand" signal that no other model on the site sees.

What SURGE measures

For each starter on a team, SURGE computes the difference between their K-D differential over their last 3 matches versus matches 4-6 before. This is the player's recent trajectory — are they trending up or down?

Then SURGE takes the maximum across the team's five starters. A team with one player who's significantly improved his K-D in the last week is more dangerous than ELO suggests — and SURGE captures that.

How we found it

We ran a grid search over 400 hypothesis variants on 10 years of historical HLTV data — varying window sizes (3/5/7/10/15 matches), features (rating, ADR, KAST, K-D differential), and aggregations (mean/median/max/min). The winner across all robustness checks: recent 3 vs prior 3, K-D differential, team max.

Backtest performance

On a 1,973-match holdout of top-tier matches:

• SURGE standalone: 73.9% accuracy
• SURGE + ELO combined: 77.0% accuracy (+11.8 pp vs ELO alone)
• Bootstrap 95% CI: [75.1%, 78.7%]
• Correlation with ELO: 0.11 — essentially independent signal

That 0.11 correlation is the key result. It means SURGE is providing genuinely new information that ELO doesn't see. When two independent signals agree, their combined prediction is significantly stronger.

Why it works

ELO captures team-level Win/Loss outcomes — but it's slow to update and averages across the roster. SQUAD captures the level of player ratings but not the trend. SURGE captures the trend specifically of the best-improving player, which is what often drives upsets at the top of CS2 (one star catching fire can carry the team).

Availability

SURGE is available to Pro subscribers. Predictions will start showing up on upcoming matches within minutes of this post — the refresh daemon has already picked up the new model.

As live data accumulates, you'll see actual production accuracy on the model's page.