About

Pre-match CS2 predictions from team ratings, map strength and form, checked against the market.

About

cs2edge predicts the outcome of professional CS2 matches before they start. Match data and team rankings are scraped from HLTV, then fed into a Glicko-2 rating system that tracks each team's strength and uncertainty over time. A logistic-regression model trained on these ratings produces calibrated win probabilities for every predicted match.

When bookmaker odds are available and market coverage is sufficient, an optional market blend(controlled by α) mixes the model's probability with the de-vigged market consensus. At full market weight (α = 1) the output is purely market-derived; at zero it is purely model-derived. α is auto-tuned weekly once coverage crosses the threshold.

The Performance page tracks historical accuracy, Brier score, and calibration across all resolved predictions.

Glossary

Win probability: The model's estimated chance that team A wins the match. Values above 50% favour team A.
Glicko-2: Opponent-adjusted skill rating. Higher is stronger. Each team also carries an RD, the uncertainty of its rating.
Strength: A team's validated rating on a specific map: 75% overall Glicko-2 plus 25% map-specific evidence. Ranks the map leaderboards.
RD: Rating deviation: how uncertain a rating is. Smaller means more confident. 350 is the default for a team without history.
Uncertainty: Combined uncertainty on a map rating, increased when the team has been inactive.
CI: 95% confidence interval on the win probability.
Edge: Model probability minus market probability, in percentage points. Edges under 2pp are noise and are not shown.
Market, implied: The market's own win probability with the bookmaker margin removed, the fair line the model is compared against.
Closing line: The final pre-match price. Predictions are scored against it.
Market blend, alpha: Weight that mixes model probability with the market: 0 is pure model, 1 is pure market. Auto-tuned weekly once coverage crosses the threshold.
Market coverage: Upcoming matches with bookmaker lines, against the threshold needed to enable blending.
Brier score: A proper scoring rule for probabilistic forecasts. Lower is better; always guessing 50% scores exactly 0.25.
Calibration: How well predicted probabilities match observed win rates: matches called at 70% should be won about 70% of the time.
Form: The last ten results. A filled segment is a win, an outlined one a loss.
Win rate: Maps won over decided maps, on a specific map leaderboard.
Bo: Best-of series length: Bo1, Bo3 or Bo5.
Pick: A map confirmed in the veto, so it will be played.
Provisional: A team with fewer than five results in the database. Its rating carries a high RD.
Diamond call: A match where the model is at 58% or more and the market at 71% or more on the same team. Marked with a diamond glyph.

Methodology

Ratings are computed via a full Glicko-2 replay over all historical match results, updated continuously as new results arrive. Per-map ratings are also maintained for each team-map combination, feeding the map-series blend that sharpens predictions once the veto is known.

The logistic-regression model uses Glicko-2 features (rating difference, RD, head-to-head history, recent form) as inputs. It is retrained periodically on the full resolved-match history.

After the map veto is published, the system can incorporate a post-veto blend: per-map win probabilities for the picked maps are combined into a Bo-series probability and blended with the base prediction. This captures map-pool asymmetries that the base model averages over.

Honest disclaimer

Predictions are pre-match only. No live-score or in-play data is used. Past accuracy does not guarantee future results. This tool is for informational purposes only; it is not financial or betting advice.