Notifications
My Performance
Knowledge Base

IGL Book

Empirical, round-deciding patterns from the CS2 match corpus

This book distills measurable, statistically robust patterns from ~2,000 CS2 demos — not opinions, but frequencies. Every number is an empirical round- or match-win probability over tens of thousands of rounds. Where a correlation is not a causal mechanism, it says so explicitly.

~2,000
Demos
88,000
Rounds
14,359
Players
317k
Kills
Methodology

Method in one sentence: we rigorously separate correlation from causation. A good predictor of winning (e.g. "the team plants") is not automatically a cause — planting is itself a mid-round success. Such confounds are conditioned out (e.g. "within planted rounds only"), and win probabilities are read side- and economy-adjusted. Sample sizes are noted on every table.

1 · Reading the Round

The man-for-man win probability is the foundation of every in-game decision — it tells you what a situation is objectively worth before individual skill enters the picture.

P(CT wins) by man-count — row = CT alive, column = T alive
CT\T54321
570%86%96%100%
431%49%71%90%99%
313%25%47%74%95%
23%7%19%45%80%
10%1%3%12%44%
Sample size per cell 8,000–24,000 rounds.

First blood = +20pp

5v4 → 70% round win, 4v5 → 31%. Side-independent — the first-blood side wins 69% whether it's CT or T.

The trade is worth ~20pp

5v4 → 4v4 (the enemy trades your entry back) drops from 70% to 49% — exactly back to a coin flip. This quantifies the value of trading.

Marginal value of the next kill — it grows as the round shrinks
TransitionWP swing
5v4 → 5v3+16pp
4v4 → 4v3+22pp
4v3 → 4v2+19pp
3v3 → 3v2+27pp
In late even fights, play for the pick — not for map control.

Even ≠ 50/50

It tilts toward T as numbers shrink: 4v4 49% → 3v3 47% → 2v2 45% → 1v1 44% (CT). The T holds initiative / post-plant pressure. In a 1v1 retake you're the underdog; in a 1v1 as T with the bomb down you're the favourite.

2 · Opening Duels & Trading

First contact decides more rounds than anything else. But the opening duel is not just about taking it — it's about winning it, and about whether your team can trade the loss.

First blood → ~70% round win — on every map

The opener's side wins 67–71% across all maps. CT takes first blood slightly more often (53–55%) because they hold angles. Winning first contact is universally decisive.

Opening-duel WIN rate is a real skill signal

Among entries, opening-duel win rate correlates 0.60 with overall frag output (KPR). Top-quartile openers average 1.04 KPR vs 0.77 for the bottom quartile. Scout who wins their first contacts, not who merely takes them.

Losing first blood is survivable — IF you trade it

When the opening victim's team trades the kill back, that side still wins 43%; un-traded, only 25% (+18pp). Yet only 30% of opening deaths actually get traded — the single biggest unrealised lever in the data.

An AWP opener survives more than a rifle opener

After taking the opening kill, an AWPer survives the round 49% of the time vs 41% for a rifler (both sides win ~77% after first blood). The AWP's edge is pick-and-survive, not just the pick.

Honest null result: dying "tradeably" is not a skill marker

How often a player's death gets traded does not predict their skill (top-quartile 0.88 KPR vs bottom 0.91) — it's positioning / role, not quality. We report it because separating signal from noise is the whole point.

3 · The AWP — why every team runs one

Every pro team fields an AWPer. The data explains why in concrete, measurable terms — and it isn't raw frag output.

AWPer vs the field — the role's fingerprint
MetricAWPerRifler
Median kill range1917
Headshot %39%50%
Opens the round0.410.30
Survives the pick49%41%
The AWPer owns long range and doesn't need headshots; the value is range control + survival.

Pick-and-survive + range dominance

The AWPer takes opening picks at distance the rifle can't contest, and survives them about half the time — turning a 5v5 into a 5v4 while staying on the board. That is the mechanism.

But AWP count does NOT predict winning

Teams with 0 AWPers win as often as teams with 1 (54% vs 53%). "Every team has an AWP" is about HOW you control space — not about whether the AWP itself wins more rounds. Build around the function, not the checkbox.

4 · Spacing & Positioning

Where you stand relative to your team is a measurable, causal lever. Both at the kill level and the team level the data shows the same inverted-U: too clustered is the worst, too scattered is bad, spread-but-connected wins.

Distance to nearest teammate at the kill → round win & survival
SpacingWin%Survive%
glued <300u65%42%
near 300–700u72%47%
spread 700–1200u73%49%
alone >1200u66%47%
n = 80k kills. Glued is worst (one nade kills two, wasted angles); alone is unsupported.
Team-level spacing (avg kill distance) → round win
Team playsWin%
tight <500u (stack)42%
medium 500–900u63%
loose >900u49%
Two independent cuts agree: optimal spacing ≈ 500–900u. "Spread but connected."
Fighting outnumbered — survival collapses with enemy density
Enemies near (<600u)Win%Survive%
0 (clean 1v1)69%49%
164%37%
265%30%
3+71%23%
Running into multiple enemies is near-certain death (23% survival at 3+) — only worth it if your death directly creates space/trades.

5 · Utility

Utility usage is one of the strongest controllable levers on the round, and how you deploy it matters as much as how much.

Utility investment in the first 30s (full-buy) → round win
Util volumeWin%
low39%
medium51%
high64%
+25pp from low to high. Most players under-invest. (Caveat: some team-skill confound, but counted only in the first 30s — mostly pre-engagement intent.)
Utility spread → round win (full-buy)
Util placementWin%
tight <700u (one site)35%
medium50%
spread >1100u (map-wide)63%
+28pp. Map-control / default with utility beats the rigid one-site execute in pugs.

Support sets up before contact

Supports throw 1.44 pieces of utility per round before first contact, vs 0.81 for riflers. The support's value is setup discipline — utility thrown to shape the fight, not react to it.

Don't waste utility on an anti-eco

Anti-eco rounds win 84% whether you throw lots of utility or little — the equipment edge already carries it. Save your nades and play retreatable; spend the round, not the kit.

6 · The Economy

The economy is worth ~36pp of round equity before any aim. The real question is rarely "can I buy" but "what is the +EV decision over the next two rounds".

Buy-tier matchup → win% (your tier vs enemy tier)
you\enemyEcoHalfFull
Eco50%13%17%
Half87%50%38%
Full83%62%50%
Equipment lead → round win
Equipment vs enemyWin%
behind > $4k32%
even (±$4k)~50%
ahead > $4k68%
A >$4k lead is +36pp. Economy management is a third of the round, not a soft skill.
Force vs save after a loss — 2-round EV (win this + win next)
ScoreSave ΣForce ΣFull Σ
leading (+2)607491
close (±1)647996
behind (−2)6785100
Force beats save everywhere — most when behind. (Caveat: money-depth confound — the save group is broker.)

Don't chain forces

Force→force → the round after is only 36%. Save→save (a real reset) → 49%. Force once opportunistically; if it fails, reset rather than force again.

Force only pays vs a weak enemy

A half-buy force wins 87% vs eco, 50% vs half, but only 38% vs a full buy. Never force into a fully-equipped enemy.

The first loss is the cliff

After a win you take the next round 62%; after one loss, 36% — then a plateau (~36%, it doesn't get deeper). The hole is stable, not bottomless.

Match point: buy a FULL

Closing rounds convert 58% on a full buy vs 45% forcing vs 19% on eco. Don't force or eco away match point.

7 · Match Structure (Bo13)

Not every round weighs the same. This is where the leverage lives in a best-of-13 — and why the pistol and the back half count disproportionately.

Lead → P(match win) — row = your rounds, column = enemy rounds
you\enemy0468
469%42%25%11%
686%51%39%20%
870%55%37%
6-0 → 86% match win; 6-6 → 51%; 12-6 → 94%; 12-11 → 52% (coin flip).

The pistol is worth 2 rounds, not 1

Win the pistol → win round 2 at 82% (the anti-eco conversion); lose it → lose round 2 at 89%. By round 3 the influence is gone (49%). The pistol is a 2-round package.

Late rounds carry escalating leverage

A round's match-WP swing: early ~15pp, R17 26pp, R21 30pp, R23 37pp, R24 57pp. The back half (R15–24) decides the match — an early loss is low-leverage. (Caveat: partly mechanical — late rounds only occur in close games.)

The post-pistol conversion rounds (R2–4) are high-leverage

Δ22–24pp — more than any mid-half round. Converting the pistol matters more than round 7.

8 · T-Side Strategy & CT Setups

T-side tactics revolve around one event: the plant. Because planting is itself a success, raw archetype win-rates are contaminated — only the causally clean, conditioned findings are below.

The plant is the pivot event

T plants → 69% win, doesn't plant → 23% (+46pp). T-tactics = get the bomb down; CT-tactics = deny the plant. ("Execute wins 84%" is mostly this confound — planting means you're already winning.)

Within planted rounds (clean): plant timing → post-plant win%
Plant timeWin%
rush <18s84%
execute 18–30s84%
slow 30–45s76%
very late >45s59%
Plant FAST — earlier = more time/players to defend it (+25pp vs late). All rounds here are planted, so the confound is controlled.

Save utility for the post-plant

Even after planting: lots of util (≥11) → 77% vs little (<8) → 58% (+19pp). Don't burn everything on the take.

Map control beats a tight commit

Teams that stay spread plant more (78%) and win more (67%) than teams that converge tightly (61%/67%). In pugs, default/spread beats the rigid one-site stack.

CT: don't over-stack

CT spread/default → 47% CT win vs CT stack → 42%. A stack leaves the other site open.

Bomb carrier ≠ planter

In 5,165 rounds the carrier changes (drops/passes). But "the lone carrier wins more" is a spacing confound (winning teams push forward), not a recipe to expose your carrier.

9 · Roles & Team Building

Roles are real and measurable, but no composition dictates winning. Build around function — range control, trade pairs, utility setup — not a role checklist.

The six roles — measurable signature
RoleSignature
Entrywins 60% of opening duels, KPR 0.97, dies (27% survival)
AWPerrange 19, HS only 39%, survives the pick (49%)
Support1.4 utility before contact, utility-volume leader
Refragger27% trade rate, plays off teammates
Lurker61% isolated kills, late (32s)
Riflerbaseline / all-rounder

Trade pairs are the biggest structural lever

Only 30% of opening deaths get traded; a trade is worth ~20pp. Every entry needs a trade partner in crossfire distance — this is the highest-ROI structural fix in the game.

Composition does not dictate wins

AWP count doesn't predict winning, a 5-rifle lineup ≈ an AWP lineup, and role diversity is flat. Field a team around function, not a roster of titles.

10 · Lurking

The lurk is the highest-variance play in the game. The data shows exactly when it pays — and when a man is worth more in the trade net.

Lurk risk/reward — when a lurk has impact
ConditionWin%
lurker survives the kill96%
lurker dies for the kill42%
late >40s (post-plant / flank)75%
early <20s70%
BEST: late + alone + survives95%
WORST: early + into the group66%
A surviving lurk pick (96%) is the highest-value play; a lurk that costs the lurker's life (42%) is the worst.

Lurk for free picks, never for trades

The entire value is in surviving. A bad lurk (early, into traffic, dies) wins 66% — worse than simply playing with the team (supported kills win 69%). Lurk only when it's late, isolated and survivable. (Causal caveat: surviving correlates with winning the fight anyway — but the actionable rule holds.)

11 · Per-Map Baselines

Map-specific baselines: side balance, plant frequency and tempo, utility volume — the foundation of pre-match prep.

Map baselines
MapCT winPlant%Plant timeUtil/rd
Anubis50%66%26s7.5
Dust251%62%25s7.4
Ancient52%67%22s7.7
Mirage53%59%24s7.2
Inferno53%66%28s7.7
Nuke53%60%26s5.7

Maps are CT-sided (50–53%)

Anubis is the most balanced (50/50). Ancient is the "plant map" (67%, fastest plants at 22s). Nuke is utility-light (5.7/rd, vertical).

12 · The Biggest DOs & DON'Ts

A summary of the largest measurable levers on the round, ranked by effect size.

Effect size on round win
FactorDODON'TGap
Take the opening kill75%25%+50pp
Hold a >$4k equipment lead68%32%+36pp
Fight supported (spaced, not stacked)73%65%+8pp
Trade the entry death immediately43%25%+18pp
Invest utility (full-buy, first 30s)64%39%+25pp
Spread utility, don't bunch it63%35%+28pp
First blood and economy dominate everything; trading and utility are the biggest controllable levers.

13 · Metric Reliability — know your numbers

Not every stat means something after one match. This is the statistical bedrock behind the role-fit engine — and a warning against reading single-game numbers.

How quickly each metric stabilises (variance decomposition)
MetricReliable afterVerdict
Utility / round~1 matchrock-solid (role identity)
AWP share~1 matchrock-solid (role identity)
Headshot %~3 matchesstable
Kills / round~5 matchesstable
Opening involvement~5 matchesstable
Multi-kills, survival~8–10 matchesneeds volume
Role-identity metrics (util, AWP) are trustworthy almost immediately; frag stats are mostly noise in a single match (ICC 0.20–0.32).

Why single-match stats mislead

After one game, ~70% of the variance in your frag stats is noise. That's why the player-fit engine shrinks noisy metrics toward the role baseline by reliability — and why a profile sharpens as you play more.

14 · Methodology & Limitations

Scientific honesty: the sample, the definitions, the de-confounding, and the deliberate limits of this analysis.

Data basis

~2,000 fully parsed CS2 demos (88,000 rounds, 14,359 players, 317k kills) from Faceit matchmaking. Win probabilities are observed frequencies; n is shown on every table.

Correlation ≠ causation

A predictor of wins is not automatically a cause. Confounds (planting, early kills and surviving are themselves mid-round successes) are conditioned out — findings hold "within" the controlled condition.

De-confounding, concretely

Tactics are judged only within planted vs non-planted rounds; economy effects are read side-separated; role metrics are relative to role expectation (a support's low KPR is not a deficit).

Limitations

The corpus is matchmaking, not structured pro teams → the mechanisms (first blood, trade, spacing, utility, economy) are universal, but set-play structure is pug-specific. Individual cell samples vary; very small n is not reported.

The player-fit engine

The role-fit (in the Analytics Hub) is reliability-weighted (shrinkage by split-half-validated stability) and measures deviation from role expectation. Split-half: best-fit role is stable Top-1 59% / Top-2 81% (chance 20%/40%) — a real, stable trait.

Empirical analysis of the Flashback demo corpus · win probabilities are observed frequencies · correlation ≠ causation unless noted.