A model for tomorrow's CS2 matches.
After months in private development, Project Leggy is now publicly released — a CS2 prediction model finally open for the world to use. Every day, Leggy publishes the next day's slate of predictions and grades itself once the results come in — match by match, map by map.
Project Leggy is a next-generation CS2 prediction engine — fine-tuned on every recorded match in 2026 and powered by millions of daily simulations. Per-map win probability, predicted scores, veto modeling, and the driving features behind every call are all surfaced live as the slate publishes. Confidence is shown raw — no rounding to a nicer story.
All-time accuracy
How a prediction is built
Every published prediction starts from a snapshot of both teams' recent professional matches — fetched daily from the underlying database — and reduces them to a few dozen features per side: map-level win rates over a trailing window, recent series form, head-to-head history, veto tendencies, and player-level rating signals. Features are encoded so that recent matches count more than older ones, and the same window is used at training time.
Map-level win probability is the model's primary output. From there, the predicted veto is simulated: each side bans and picks under the policy implied by their veto history, producing a distribution over which maps are likely to play. Series-level outcomes are then constructed by simulation — millions of runs daily — sampling map results under the predicted veto and aggregating into the series score distribution shown on each match card.
Confidence is calibrated against historical predictions, so a stated 65% win probability has historically resolved as a win roughly 65% of the time. The number you see on a match page is the raw model output after calibration — never rounded toward a nicer story, never nudged for editorial reasons.
What's surfaced on each match
For every match the model predicts, the detail page exposes the full reasoning surface: per-map win probability and predicted score, the veto scenarios with their probabilities, the top features driving each map's call, both teams' recent results on that specific map, and — once the match resolves — a graded view that highlights which maps and scenarios were correct. Nothing is hidden behind a "premium" tier; the entire model output is public.
What this isn't
Project Leggy is not a betting service, not a tipster, and does not publish "locks" or odds-based recommendations. It's a model — its outputs are probabilities, and probabilities are wrong by definition some of the time. The accuracy and calibration figures on this page are the only honest claim the project makes about quality. A match isn't predicted at all when either team has fewer than seven recorded matches in the database, because below that threshold the features carry too little signal to produce a probability worth reporting.
Frequently asked questions
How accurate are Leggy's predictions?
Currently 86.6% on match winners across 149 graded matches since launch. Every prediction is logged and graded against the actual result, and the running figure is recomputed on each daily build.
How is the model trained?
Leggy is trained on every recorded professional CS2 match in 2026, with features covering map-level win rates, recent form, head-to-head results, veto tendencies, and player-level statistics. Map-level win probabilities are produced by simulation over the predicted veto outcome.
How often are predictions refreshed?
Once per day. The full slate for the next day publishes overnight (UTC), and the previous day's results are graded against the predictions in the same run.
Why aren't all matches predicted?
The model only publishes a prediction when both teams have at least seven recorded matches in the database — fewer than that and there isn't enough signal to produce a probability worth reporting. Matches without a prediction are still listed but marked as such.
Is this betting advice?
No. Project Leggy publishes model output for entertainment and research. Confidence is shown raw, with no rounding to a nicer story, and the site is not a betting service.