The Backtest: 2,131 Pro Matches, Three Predictors
Before launching this page we tested the obvious approaches on 2,131 professional CS2 matches (April–July 2026, all tiers) where a final result, live bookmaker odds and an AI prediction were all available. The results decided our methodology:
| Predictor | Accuracy | Flat-stake ROI at listed odds |
|---|---|---|
| Bookmaker favorite | 82.1% | −3.6% |
| CS2WH Model v3 (map-Elo + form + head-to-head + roster changes) | 65.4% | — |
| BO3.gg's AI | 62.0% | −22.1% |
| Simple Elo baseline | 62.8% | — |
| AI's underdog picks only (736 bets) | 20.8% | −59.6% |
The "value bet" test — and why we publish it anyway
We also simulated the strategy everyone asks for: bet whenever our model's probability exceeds the bookmaker's implied probability. On the 6,000-match holdout it lost at every threshold— −62.9% ROI at a 5-point edge (3,713 bets) down to −86.2% at a 20-point edge. Model-vs-market divergence has historically meant the model was missing something the market knew. We still flag these divergences live ("Value Watch" above) and grade them automatically — if the live record ever contradicts the backtest, the data will show it here first. Until then, treat every "value bet" claim you see — ours included — as unproven.
What this means
- The market is the best public predictor. Odds-implied probabilities were well calibrated in every band we tested — when the market said 70–80%, favorites won 85%; at 40–50%, they won 47%.
- No simple model beats the bookmaker.An esports AI and a classic Elo both landed ~20 points below the market. Every "value betting" strategy we tested lost money after the margin.
- Underdog "value picks" were the worst bet of all — following the AI when it disagreed with the bookmaker lost 60% of everything staked.
Our methodology, therefore
We lead every match page with the market-implied win probability(live odds, margin removed) and surround it with the context odds can't show — six-month form, T/CT and pistol-round stats, head-to-head — plus the CS2WH Model: map-level Elo per team, recent form, head-to-head record and roster-change signals, blended by a logistic layer trained on 24,920 professional matches. It re-rates every team automatically as each new result and roster move lands, and it is well calibrated (when it says 70–80%, teams win 78%). We only add factors that prove themselves on unseen matches: head-to-head (+0.2 pts) and roster changes (+0.4 pts) made the cut; rest days, win streaks and round-margin momentum were tested and rejected. Every prediction is snapshotted at match start, graded automatically, and reported above unedited.
The honest takeaway for bettors: if you bet, do it for entertainment on sites we've vetted, expect the market to already know what you know, and never stake more than you can afford to lose (responsible gambling resources). 18+.
Data & reproducibility
Match data, odds and AI predictions come from BO3.gg's public API. The backtest counted a prediction as correct when the picked team won the match; matches decided by forfeit or missing any of the three signals were excluded. ROI assumes a flat 1-unit stake at the listed decimal odds.