Research · Aug 2026
Polymarket is sharp at League of Legends. We have the data to prove it — because it's beating our model.
131 professional series · June 11 – August 5, 2026 · every prediction frozen pre-match and published
TL;DR: We built a Bayesian model that predicts pro LoL series and tracked it against Polymarket's pre-match odds for two months, series by series, in public. The market beat our model overall (70.2% vs 63.4% accuracy). When we disagreed with the market, the market was right 2-to-1. And yet our real-money bot — every bet verifiable on-chain — is up +$163 on $2,054 wagered. This post explains why all three facts can be true at once.
Methodology
- Window: Jun 11 – Aug 5, 2026. Every pro series with both a frozen model probability and a Polymarket pre-match price (degenerate 0%/100% quotes excluded): 131 series, 12 leagues (LPL, LCK, LEC, LCS, CBLOL, LCP, MSI, EMEA Masters, Prime League, LFL, LES, NACL).
- Model: Bayesian, 7 signals — champion win rates on patch, lane matchups, duo synergy, team form, head-to-head, Elo, season win rate. Accuracy tracked publicly at /model/accuracy.
- Scoring: accuracy (did the >50% side win) and Brier score (lower = better calibrated).
1. The market wins overall
| Accuracy | Brier | |
|---|---|---|
| Polymarket | 70.2% | 0.2022 |
| Our model | 63.4% | 0.2247 |
No excuses: over 131 series the crowd with money on the line beat our seven-signal model on both accuracy and calibration.
2. When we disagree with the market, we lose 2-to-1
In 104 of 131 series, model and market picked the same winner (both right 71.2% of the time). In the 27 series where they disagreed, Polymarket's pick won 66.7% of the time vs our 33.3%. If you think you've spotted something an esports moneyline missed — two times out of three, you haven't.
3. Except in the LCK, where everyone drowned — and the market drowned harder
| League | n | Model acc. | Market acc. | Model Brier | Market Brier |
|---|---|---|---|---|---|
| Prime League | 27 | 59.3% | 63.0% | 0.254 | 0.238 |
| LPL | 22 | 77.3% | 86.4% | 0.203 | 0.147 |
| LES | 12 | 75.0% | 91.7% | 0.188 | 0.075 |
| LCK | 11 | 36.4% | 18.2% | 0.249 | 0.432 |
| LEC | 11 | 63.6% | 81.8% | 0.185 | 0.127 |
| EMEA Masters | 9 | 44.4% | 66.7% | 0.274 | 0.264 |
| LFL | 8 | 37.5% | 50.0% | 0.287 | 0.227 |
| LCP | 8 | 62.5% | 62.5% | 0.209 | 0.250 |
| MSI | 8 | 50.0% | 75.0% | 0.239 | 0.160 |
| NACL | 6 | 83.3% | 66.7% | 0.218 | 0.234 |
The wildest stat in the dataset: during this stretch Polymarket favorites lost 9 of 11 LCK series (18.2% accuracy, Brier 0.432 — worse than a coin flip priced with confidence). An upset wave this violent punishes confident pricing hardest, and the market was more confident than we were. Small sample — but if you bet the LCK this summer, you felt it.
4. Our model knows what it doesn't know
| Model confidence | n | Actual win rate |
|---|---|---|
| 50–60% | 71 | 54.9% |
| 60–70% | 43 | 69.8% |
| 70–80% | 13 | 84.6% |
| 82–87% | 4 | 75.0% |
Monotone and close to the diagonal: when the model says 60–70%, the pick wins 69.8% of the time. Less accurate than the market — but honest about its own uncertainty, which is what makes it usable for edge detection rather than pick'em.
So why is the real-money bot up?
Separately from this series-level comparison, we run a bot that bets real USDC on Polymarket's game-level LoL markets — only when the model's probability diverges from the price by a minimum edge, with flat sizing. Its wallet is public and every bet is verifiable on-chain: draftlol.ai/bot.
Since going public on July 24: 141 settled bets, 61.7% win rate, +$163.12 on $2,053.94 wagered (+7.9% ROI).
The resolution of the apparent paradox: you don't need to out-predict the market on average to bet profitably. You need to (a) know how good your own probabilities are, (b) only act when price vs. probability leaves margin after fees, and (c) pass on everything else. The model loses the accuracy contest and still finds enough mispriced spots to stay ahead — for now. If that changes, the wallet will say so before we do.
Data & reproducibility
Every prediction was frozen pre-match and is listed at /predictions with its outcome; model accuracy history at /model/accuracy; the bot's live balance, positions and full bet log at /bot. Questions or want a cut of the data we didn't publish: @draftlolai.