Methodology & backtest

It's tempting to think a hard counter-pick (Renekton into Shen, Ashe into Teemo) should let a draft-only model predict much better than ~55%. We tested that directly, with a leak-free walk-forward harness (each model trained only on games strictly before the games it's scored on), and none of these beat the shipped model on both accuracy and calibration:

• Solo-queue counter-picks. A dense champion-vs-champion lane-matchup prior from ~9,900 solo-queue pairs (emerald+, far more than pro data ever sees). Result: no change (54.7% → 54.7%). Pro lane outcomes are dominated by team coordination, not the solo matchup.
• Off-meta / off-role picks. Flagging picks played out of their usual role (the “Ashe top” signal). Nudged accuracy up ~0.3pp but hurt calibration — directionally real, too noisy to trust as a probability.
• More history.Training on 3–4 seasons instead of 2. No gain: recency decay already fades old seasons to near-zero weight.
• League effects & pick order. Per-league blue-side bias, region-local metas, and pro pick-order (who counter-picked whom). All within noise or accuracy-up/calibration-down.

The draft-only mode now has its own dedicated fit. Previously the no-teams prediction reused the full model's coefficients with the team terms zeroed — but those weights were fit with team features present, and without them the pairwise-synergy term (which is estimated in-sample) gets over-trusted. Measured on the walk-forward harness, adding synergy to a draft-only fit makes it worse than champion strength alone (55.7% vs 56.0% current-season), while the lane-matchup term helps. The shipped draft-only fit therefore uses champion strength + lane matchups with synergy excluded: 56.2% vs 56.0% current-season and 55.1% vs 54.6% overall against the old zeroed-coefficients behaviour, with better calibration on both slices. (The draft builder's synergy-pick suggestion still ranks through the full model's synergy weight — it's a relative comparison among candidates, not a calibrated probability.) We also tested a champion “dominance” rating — margin-of-victory averaged per champion, the champion-level analogue of the team-Elo margin weighting. Rejected: log-loss identical to four decimals in every configuration; dominance alone predicts 55.2%, nearly as well as win rates do, confirming the two measure the same underlying champion strength.

The honest conclusion: at the pro level, which champions get drafted is a genuinely weak predictor once you can't see who's piloting them. That's exactly why the “teams + players” model jumps ~10 points to ~66% — the predictable signal lives in team and player identity, not the champion select screen. Disaster drafts are real but rare (a handful per season), so even predicting every one perfectly moves aggregate accuracy under a point.

Should it train on the current season only? We measured that too. Training on this season alone (walk-forward within it) scores 63.7% with teams, ~2 points BELOW the shipped model that also uses last season as history — the early-season games starve without it. More recent-weighted data doesn't beat the current recency decay; less data just loses signal.

International events (MSI / Worlds / EWC) are the hardest. Held-out accuracy on cross-region events runs well below regional play — ~57% aggregate (EWC 66%, Worlds 58%, MSI ~52% on a tiny 80-game sample) vs 65% inside a single league. The cause is structural: team Elo only bridges regions through the handful of inter-region games, so when the best of two regions meet the rating gap is genuinely uncertain. We help it where we can — inter-region games (an international league, or any game where the two teams' home leagues differ) move Elo 3× as much, since they're the only games that calibrate strength across regions. That lifted held-out international accuracy 55.7% → 56.5% and improved its calibration (log-loss 0.715 → 0.714) with no cost to regional games — a real but modest gain; cross-region prediction stays fundamentally data-starved. The draft builder flags these matchups so you can weight them accordingly.

What about wombo combos, all-AD/all-AP teams, or a low-damage/all-squishy comp? This is a real gap: the model's only inter-champion signal is pairwise synergy (exactly 2 champions at a time, shrunk toward 0) — there is no feature that looks at the 5-champion whole. We built the best measurable proxies available and tested each: team damage output and tankiness (from real per-game damage/tanking stats), damage-share concentration (one carry vs a balanced comp), and AD/AP balance and engage-champion count (from champion attributes, since Oracle's Elixir has no crowd-control-time column to measure a true “wombo” signal from). Draft-only, one candidate (AD/AP lean) showed a small, consistent gain across every test fold. But once combined with team and player Elo — the model actually used once you pick teams — that gain vanished into noise (currentSeason 65.71% → 65.76%, well within run-to-run variance) while overall accuracy and log-loss both got marginally worse. Team/player identity is simply too dominant a signal for these subtler structural effects to move the needle on real pro outcomes. None of these features ship.

Why not just use Riot's official Global Power Rankings? We tested it properly rather than assuming. The GPR history is real and, importantly, not retroactively rewritten — we verified 7,032 overlapping historical observations fetched through different query windows and found zero conflicts, so using it wouldn't leak future results into past predictions. Added as a feature it changed held-out log-loss from 0.6125 to 0.6125: nothing. The reason is that GPR is itself an Elo-style rating computed from the same match results our own Elo already consumes — and ours covers all 612 teams in the data, where GPR ranks only ~58 tier-1 teams (about a quarter of games, with no academy or regional sides at all). It is a less complete version of a signal the model already has.

What about form, streaks, rest days, or head-to-head records? Also tested, also negative. Recent form correlates 0.75with team Elo — it is a noisier restatement of what the rating already encodes. Win/loss streaks actively hurt (−0.5 points of current-season accuracy). Days of rest and games-in-the-last-week have essentially no relationship with the outcome at all (correlation ~0.015). Head-to-head history, series game number, playoffs-vs-regular-season, and roster changes all landed inside the measurement noise. The honest summary: once you know who is playing and how dominant they have been, schedule and momentum narratives add nothing measurable.

Every prediction shows fair decimal odds (1 ÷ probability) for each side. If a bookmaker offers longer odds than the fair number on a side, the model sees positive expected value on that side — before costs.

Be honest about the bar: a bookmaker's implied probabilities contain a built-in margin (typically ~4–7% across both sides), and closing lines on major leagues are sharp. To profit you need the model's calibration edge to exceed that margin consistently — check the log-loss and calibration table above, not just accuracy, and treat small-sample leagues (see the breakdown below) with extra skepticism. One structural caveat: team Elo is anchored across leagues only by the few inter-league games (MSI, Worlds, EWC), so ratings are most trustworthy for matchups WITHIN a league or at international events — an isolated league's ratings can drift high or low as a block. Nothing on this page is betting advice; it's a measured, walk-forward-validated probability estimate with known error bars.