Expected Duel Win Probability

xDuelCS2v0.1.0 · baseline · validated 2026-07-16

What it measures

How likely a player was to win each fight before it happened, based on weapons, health, distance, flashes, smoke, and nearby support. A kill in a 20% situation is worth more than one in a 90% situation.

Formal definition

Logistic model P(attacker wins engagement) over pre-outcome features: weapon-class matchup, distance bucket, health differential, armor, blind fractions of both players, through-smoke indicator, and trade-support proximity for both sides. Evaluated per kill event from replay-derived state.

Inputs

duel events (replay)
position samples (support distances)
flash-effect events

Exclusions

team-kills
suicides
bot takeovers
events with confidence < 0.5

Sample & population

Minimum sample: 50 duels
Baseline population: All parsed duels in the same game and window
Direction: neutral
Unit: prob

Percentile interpretation

Not ranked directly; feeds Duel+ and Opening Impact.

Confidence = source confidence × model confidence. Values from baseline models are capped at 0.7. Below-minimum samples display the value grayed out with no percentile.

Example

An eco pistol player wide-swings a rifler holding an angle: xDuel ≈ 0.18 for the pistol player.

Known limitations

v0 coefficients are priors calibrated on synthetic fixtures, not fitted on pro data
Cannot observe crosshair placement or peeker's advantage directly

Model card

xduel_cs2 v0.1.0

MODELlogistic_regression · baseline

Training data — read this

Hand-specified directional priors, sanity-calibrated on the synthetic fixture corpus v1 (fictional matches). NOT fitted on production esports data.

Calibration

Reliability checked on synthetic corpus only; expect miscalibration on real data until re-fit. Confidence for downstream metrics is capped at 0.7 while status=baseline.

Features

weapon_advWeapon-class matchup advantage (attacker perspective, -1..1)
dist_closeEngagement below close-range threshold
dist_longEngagement beyond long-range threshold
health_diff(attacker HP − victim HP) / 100 at engagement
armor_advAttacker has armor
victim_flashedVictim blind fraction (blind ms / 1500, capped)
attacker_flashedAttacker blind fraction
through_smokeEngagement through smoke
attacker_supportedAttacker teammate within trade range
victim_isolatedVictim's nearest teammate beyond trade range

Coefficients (v0.1.0)

intercept0.00

weapon_adv1.80

dist_close0.10

dist_long-0.05

health_diff1.40

armor_adv0.25

victim_flashed1.60

attacker_flashed-1.60

through_smoke-0.70

attacker_supported0.15

victim_isolated0.20

Not intended for

Betting or wagering decisions · Cross-game comparisons · Absolute skill claims

Change log

v0.1.0 · 2026-07-16Initial baseline release

→ Rank players by xDuel