CS2 Betting Model: Glicko-2 Ratings + Per-Map Monte Carlo

Olympus Bets models Counter-Strike 2 betting with Glicko-2 team ratings and per-map Monte Carlo simulation. Each map is rated and simulated independently, a veto-capture poller records the actual pre-match map veto roughly 15-30 minutes before a match starts, and per-map win probabilities are combined into series probabilities for BO1 and BO3 formats. The engine covers match winner, map handicap, and total maps markets. A round-level simulation engine is being validated in shadow mode via replay backtests before it replaces the current per-map approach. Free daily CS2 projections are on the site and via the MCP server.

Engine Overview

Counter-Strike 2 is a newer addition to our engine lineup (current tag: cs2_glicko2_v0.1), and we're upfront that it is earlier-stage than our longer-running team-sport engines. The production approach combines Glicko-2 team-strength ratings with a per-map Monte Carlo simulation, rather than trying to predict an entire best-of series as one unit.

Esports rosters and map pools change faster than traditional sports rosters — a team can be excellent on one map and mediocre on another, and that gap can shift after a single roster move. The engine is built around that reality: rate teams, simulate map by map, and let the actual veto sequence — not an assumed map distribution — determine which maps get priced.

How the Simulation Works

1. Glicko-2 Team Ratings

Each team carries a Glicko-2 rating, a rating deviation (how uncertain we are about that rating), and a volatility term. Ratings update after every observed match and map result pulled from provider adapters. The rating deviation widens during layoffs or after a roster change and narrows as a team plays consistently, so the model's confidence in a rating is itself data-driven rather than fixed.

2. Real Pre-Match Veto Capture

Map vetoes are not cosmetic — they determine which maps are actually played, and teams veto around their own strengths and weaknesses. A veto-capture poller records the real ban/pick sequence roughly 15-30 minutes before the match starts. That captured map pool feeds directly into the simulation, so map handicap and total-maps pricing reflects the maps that were actually selected rather than an average across the full map pool.

3. Per-Map Monte Carlo Simulation

Once the map pool is known, each map is simulated independently using map-specific team ratings. This produces a per-map win probability for every map in the pool, rather than a single series-wide number that would smooth away map-specific skill differences.

4. Series Probability Construction

Per-map win probabilities are combined into series-level probabilities for both BO1 and BO3 formats. Because each map is priced independently first, the series probability correctly reflects cases where a team is a strong favorite on some maps in the pool and an underdog on others, instead of a single blended rating standing in for the whole series.

5. Round-Level Engine (Shadow Mode)

A round-level simulation engine (internally tagged cs2_roundsim_v1) is currently running in shadow mode, accruing validation history through replay backtests. It is not the production engine yet — the live engine remains Glicko-2 plus per-map Monte Carlo. We're building the round-level engine to eventually model economy state and round-to-round momentum directly, but it only replaces the current approach once its replay validation clears our bar.

Data Sources

• Match and map result providers — HLTV-style match and map results via provider adapters, updated on a rolling basis
• Veto-capture poller — real pre-match ban/pick sequences captured shortly before match start
• Roster tracking — staleness-aware roster state so lineup changes are reflected in ratings rather than carried forward silently

What the Model Outputs

For each CS2 match, the pipeline produces:

• Match winner probability — series-level win probability for BO1 or BO3 formats, built from combined per-map probabilities
• Map handicap probability — the probability a team covers a given map-count handicap, informed by the captured veto pool
• Total maps probability — the probability a BO3 series goes over or under a given total maps line

These outputs feed the same profitability zone and Kelly-sizing pipeline used across every other league, so CS2 picks are gated by the same self-learning zone system rather than a separate esports-only threshold.

CS2 Engine Facts

Why Per-Map Simulation Matters for Betting

A series-level model has to summarize an entire team into one number and hope that number holds across every map in the pool. That assumption breaks down whenever a roster's map-by-map skill is uneven, which is the norm in CS2 rather than the exception.

By rating and simulating each map independently — and by feeding in the actual veto sequence rather than a generic map distribution — the model can answer map handicap and total-maps questions with the specific maps that will be played, not an average across maps that were banned out before the match even started.

Explore Other Models