How our AI predicts esports matches

The pipeline

For every upcoming CS2, League of Legends, Dota 2, or Valorant match on the tier-one circuit, we pull the two teams' identities and event context from PandaScore, plus each team's last ten finished series. That context — team names, game, event, best-of format, and each side's recent record and recent opponents — is packed into a structured prompt and sent to a large language model. The model returns a calibrated win probability between 50 and 95 percent, a one-line headline, and two-to-three sentences of reasoning.

Predictions are cached for 24 hours per match. When the cache expires, the model is asked again with fresh form data, so recent results and roster changes filter in daily rather than compounding stale reads.

What we optimize for

Calibration over confidence. A 65% pick should win around 65% of the time over a large sample. That is more useful than always leaning toward 85% — it lets you scale your interest to the actual matchup uncertainty. Our probabilities are capped between 50 and 95 percent to avoid overclaiming.

Recent form over historical results. Rosters, patches, and metas shift fast in esports. The most recent 30 days of series carry more weight than the previous season, which is why a hot streak matters even when a team's all-time record is average.

Opponent quality. A 4-1 run through weak qualifiers looks different than 4-1 through tier-one opposition. The model has that context because recent opponents are part of the prompt.

Where the model is weakest

Roster changes with fewer than five series played. A newly formed or newly rebooted lineup has almost no signal in the form data. The model will treat them at roughly the base rate for their region until they play enough.

Best-of-one variance. A single-map match between roughly even teams often gets a probability near 55/45 for the favorite. That is correct — the variance is real — but it means bo1s are less predictable by design.

Off-meta drafts and coach-driven tactics. The model does not read VODs. Any team executing something surprising at the draft or macro level will beat its probability until the results start showing up in the form window.

Market Insights & Trends data

The figures on /trends and /insights are not modeled by sport.gg — they are aggregated from public reporting on prediction-market trading volume. Every headline number links back to its primary source; we do not restate a claim without a citation.

Data sources. Annual and monthly trading volume figures come from industry press covering Polymarket, Kalshi, and adjacent venues — primarily Covers, Gambling Insider, The Block, and Binance Research. Forward projections ($740B by 2030, $1T by 2030) are analyst estimates from those publications, not sport.gg forecasts.

How we handle numbers. We quote figures verbatim from the linked report and cite the publisher and date. When two sources disagree, we show both rather than averaging. When a source revises a number, we update the page and bump its lastmod in the sitemap. We do not smooth, normalize, or extrapolate between reported data points on the charts — bars reflect what the cited source published for that period.

What the numbers do and don't mean. "Trading volume" is the notional value of contracts traded on prediction-market venues, not revenue, not user count, and not sportsbook handle. It is a useful proxy for interest and liquidity, not a measure of profit or accuracy. Projections through 2030 are opinions from named analysts and should be read as such.

Data refresh & update log

Every headline figure on /trends and /insights is re-verified against its primary source on the schedule below. The date in each row is the last time a human opened the cited report and confirmed the number still matches what we display. When a figure changes, we update the page, add a row to the changelog, and bump lastmod in the sitemap so crawlers re-index.

What this is not

This is not betting advice, and sport.gg does not accept wagers. Predictions are published for entertainment and analysis. Treat any number here as a fast baseline read, not a verdict.