Methodology
Endereye uses Monte Carlo simulation to estimate each player's probability of surviving the next elimination round, and to surface the specific conditions observed most often that determine the outcome.
In a 60+ player lobby, simulating exact permutations becomes computationally impossible. The 'Safe' and 'Needs #' labels bypass player variance by using point calculations to surface guaranteed survival thresholds.
Summary
Each time a seed completes, the model runs 20,000 simulated playthroughs of the remaining seeds. In each simulation, player performance is sampled from a model built on Elo rating and season ladder completion stats (average time, best time). The fraction of simulations in which a player survives the next elimination becomes their displayed survival probability.
Survival and threat paths are derived from the same simulation batch. After all simulations run, the model identifies which opponent placement patterns most reliably separate "survived" outcomes from "eliminated" outcomes for each player, and surfaces the most frequent ones.
Invariants
Predictive Accuracy
Model Comparison
Macro-AUC scores each event on its own field only (does the model rank that event's qualifiers above its non-qualifiers), then averages across events, so a small event and a large one count equally. The confidence interval comes from resampling which of the backtested events are included; it measures how much this lift depends on which events happen to be in the dataset, not on match-outcome randomness within a given event.
Probability Calibration
A well-calibrated model should be right as often as it says it will be: when it assigns 70% survival odds, the player should survive roughly 70% of the time. The chart below shows predicted vs. actual survival rates across all historical LCQ and MSS events. Grey bars are the predicted rate; colored bars are actual outcomes. Blue means well-calibrated; green means the model was conservative; red means it was overconfident.
Scenario Path Accuracy
To validate survival and threat paths, I check historical events against the model's predictions. For each player the model expected to survive who was actually eliminated (threat cohort) or vice versa (survival cohort), I check whether the opponent the model flagged as pivotal actually placed as predicted.
The grey bar shows what hit rate you'd get by picking that opponent randomly. The blue bar is the model's actual hit rate.
Threat paths: n=12 surprise eliminations across all events
Survival paths: n=7 clutch survivals across all events
Accuracy by Round (Brier Score)
As the tournament progresses, the model's accuracy changes because the pool gets smaller. This chart tracks the average Brier error as we approach the final seed.
Limitations
- The model does not account for player momentum, fatigue, or meta-game dynamics within an event.
- Elo and completion time metrics are based on the season ladder. A player having an unusually good or bad day is not reflected.
- Scenario path validation is based on a small number of historical events. Sample sizes will grow as more seasons are archived.
- DNF probability is estimated from historical completion rates and does not account for known circumstances like technical issues or scheduling conflicts.