A Statistical Analysis of Draft Features in Ranked League of Legends
This analysis examined 805,355 ranked matches from all regions across the last four patches, corresponding to 8,052,743 participant rows spanning Iron through Challenger. The objective was to estimate how seven draft-related features contributed to win prediction using the same statistical framework across roles, ranks, and champions.
To separate lane-specific effects from broader team-composition effects, counter and synergy variables were divided into lane and cross-lane components.
The per-role analysis used a hierarchical mixed-effects model (GPBoost) with rank tier as a grouping variable. The per-champion analysis used Ridge-regularized logistic regression to stabilize coefficient estimates in smaller samples. Both analyses used 5-fold out-of-fold cross-validation at the match level to prevent data leakage.
The results are shown in charts and tables below, filterable by role and rank:
- Feature Weights by Role — stacked bars and data table, filtered by rank
- Feature Weights by Champion — stacked bars and data table, filtered by role and rank
What the features meant
Lane Ctr Gold measured the historical average gold difference at 15 minutes between the player's champion and its lane opponent.
Lane Counter WR measured the historical win rate of the player's champion against its lane opponent.
Game Counter measured the summed win-rate delta from 50% between the player's champion and each of the four non-lane enemy champions.
Game Synergy measured the summed win-rate delta from 50% between the player's champion and allied champions, excluding the bot-lane partner except where lane synergy was modeled directly.
Meta referred to the champion's overall win rate for that role and rank.
Lane Synergy WR applied only to bot lane and measured the historical win rate of the ADC–Support pair.
Lane Synergy Gold applied only to bot lane and measured the combined gold difference at 15 minutes for the ADC–Support pair.
Methodology
The analysis used a logistic link function (Bernoulli Logit) to model the probability of victory as a function of the seven standardized draft features above, with five features used for non-bot roles and seven for bot lane.
"Standardized" means that all features were scaled to zero mean and unit variance before model fitting. This made the fitted coefficients directly comparable even though the original features were measured on different scales.
Feature importance was quantified by normalizing the absolute values of the standardized coefficients, expressing each as a percentage of the total effect magnitude within the model.
For the per-role analysis, a hierarchical mixed-effects model was fitted per role using GPBoost, with rank tier as the grouping variable for random intercepts and random slopes. For the per-champion analysis, Ridge-regularized logistic regression (L2 penalty, C=1.0) was used to stabilize coefficient estimates in smaller champion-level samples.
To prevent data leakage, the analysis used 5-fold out-of-fold (OOF) cross-validation stratified at the match level, ensuring that all ten participants from a single match were assigned to the same fold and preventing intra-match information leakage. When computing features for matches in a given fold, historical statistics were derived only from the other four folds. P-values were computed at the population level using cluster-robust standard errors grouped by match, accounting for the correlation between the ten participant rows within each match.
Statistical significance and feature correlation
Statistical significance was reported through p-values. Results were marked as:
- p < 0.001
- p < 0.01
- p < 0.05
- not statistically significant
The analysis also used Variance Inflation Factor (VIF) to examine correlation between features.
A VIF of 1 indicates no correlation with other features. Values above 5 are commonly treated as a concern. All VIF values in this analysis were below 2.
Feature Weights by Role — Silver
Weights sum to 100% per row. Colors indicate statistical confidence: gold (p<0.001), green (p<0.01), blue (p<0.05), grey (not significant).
| Role | Lane Ctr Gold | Lane Ctr WR | Game Ctr | Synergy | Meta | Lane Syn WR | Lane Syn Gold |
|---|---|---|---|---|---|---|---|
| Top | 19.8% | 7.7% | 20.3% | 12.2% | 40.0% | — | — |
| Jungle | 6.4% | 4.5% | 23.9% | 15.9% | 49.3% | — | — |
| Mid | 11.6% | 10.2% | 18.5% | 16.9% | 42.9% | — | — |
| Bot | 1.8% | 9.4% | 26.2% | 16.8% | 27.0% | 15.6% | 3.3% |
| Support | 3.7% | 15.4% | 29.5% | 4.4% | 26.2% | 17.5% | 3.3% |
Conclusion
How important is winning the lane phase?
- Lane-phase gold difference carried a meaningful signal for top lane, and its contribution for mid lane became more visible at higher ranks.
- The same measure was not statistically significant for ADC in this model. Support showed some lane-phase contribution, but less than the solo lanes.
How important is picking a champion that fits your team composition?
- Cross-lane synergy contributed most for mid and jungle, while remaining part of the signal for every role.
- For bot lane, the historical win rate of the ADC–Support pair mattered more than the pair’s gold difference at 15 minutes. The duo signal appeared to extend beyond the lane itself.
How relevant is it to counter the enemy picks?
- Lane Counter WR remained meaningful beyond the lane-phase gold measure, suggesting that matchup value was not limited to the first 15 minutes.
- Cross-lane matchup effects were also substantial across roles and ranks; in many cells, Game Counter was the largest single contributor.
What about the meta?
- At Platinum and above, Meta contributed little for ADC in this model but remained relevant for several other roles.
- Meta was strongest in many lower-rank cells and weaker in much of Gold and Platinum. The pattern varied by role, so one tier-wide rule would overstate the result.
Does the above change from champion to champion?
- Champion-level estimates varied widely and smaller samples limited firm comparisons. Katarina, LeBlanc and Kassadin were notable examples where the lane-phase measure carried much more weight than it did for many other champions.
Why Meta did not appear in the per-champion table
Meta did not appear in the per-champion table for a statistical reason. When the analysis was restricted to a single champion-role-rank group, that champion's meta win rate was the same value for every match in the group. A feature that does not vary across observations cannot be estimated by logistic regression.
For that reason, Meta's contribution to winning was captured in the per-role analysis, where different champions with different meta win rates appeared in the same group, but it did not appear in the per-champion analysis.
Known bias
The analysis identified one important bias in the per-role analysis: champion pick rates were unbalanced. Popular champions contributed much more data than rare picks — in some cases as much as 20× more. As a result, the per-role regressions were disproportionately influenced by popular champions.
The per-champion analysis was not affected by that specific bias, since each champion was analyzed independently.