Start only when your team can leave cleanly.
The objective is not worth turning the pit into a losing fight.
Objectives2,214 ATTEMPTS · 617 REPLAYS REVIEWED
When is starting Dragon worth the risk?
Securing Dragon helps. Dying around it gives much of the edge back. Here is what to protect before, during, and after the call.
Which Dragon fits your composition?Reset before forcing the next play.
Keeping the gold moving in your favor mattered more than immediately forcing a tower.
Trade the map when the contest is already bad.
A missing attack on Dragon can still mean farming, vision, pressure, or a cross-map play.
eventual win · 1,774 observed attempts
eventual win · 366 observed attempts
eventual win · 74 observed attempts
attempts include opposing Dragon-targeted attacks seen in the replay
WHAT HAPPENED AFTER THE CALL
Finishing Dragon is only the first step.
These are the results after every qualifying attempt Manalytics could review, not only the successful ones.
68.3% final win
The starting team secured Dragon.54.9% final win
The starting team left without either team taking Dragon.47.3% final win
The other team secured after the starter committed.THE COST OF DYING AT DRAGON
Two deaths erase most of the secure’s advantage.
Casualties are inferred from the complete minute snapshots bracketing the attempt, not exact death locations seen in the replay.
final win · 876 attempts
- Opponent structure @ 3m
- 30.8%
- Own structure @ 5m
- 67.8%
final win · 675 attempts
- Opponent structure @ 3m
- 35.0%
- Own structure @ 5m
- 66.8%
final win · 642 attempts
- Opponent structure @ 3m
- 43.9%
- Own structure @ 5m
- 61.7%
The observed penalty: the starter’s final win rate falls 19.1 points from a deathless attempt to two-plus deaths, while the opponent’s immediate structure conversion rises 13.1 points.
CONTEST OR TRADE THE MAP
Not hitting Dragon does not mean doing nothing.
Manalytics can see attacks on Dragon, but it cannot fully see comms, threatened engages, or every zoning decision.
The sample is too small to give reliable advice yet.
The other team still won 31.8% and took a structure within three minutes 35.7% of the time.
No attack on Dragon may mean a cross-map play, zoning, farming, late arrival, or a true concede.
AFTER THE SECURE
Leave the pit with the advantage.
Dragon creates an opening. Whether it lasts depends on who died, the reset, the next wave, and the next safe structure.
Teams that kept gaining gold and took a structure reached 79.7% modeled win chance. When the gold swung back and no structure fell, that dropped to 42.7%.
THE CLEAN EXIT
Secure → reset → take what is safe → set up again.
The best Dragon play continues after the objective dies.
Look at lanes, health, numbers, and what the enemy can contest with.
Protect the gold and health advantage from the next bad skirmish.
Take the structure the map gives you instead of forcing one.
Spend, fix waves, and make the next Dragon easier.
FOUR WAYS OUT OF THE PIT
Keeping the gold lead matters more than forcing a tower.
These modeled win chances compare teams that entered the five-minute window in similar situations.
14,562 captures · 45.4% of covered cases
4,718 captures · 14.7% of covered cases
6,316 captures · 19.7% of covered cases
6,495 captures · 20.2% of covered cases
The player read: gaining gold without a tower performed 20.5 points better than taking a tower while the gold state got worse. Do not force the visible reward if the fight or siege is bad.
THE NEXT 60 SECONDS
The minute after Dragon tells you whether the call worked.
These are patterns in similar games, not a guarantee that one action caused the result.
clean follow-up · 71.0% final win
clean follow-up · 54.9% final win
clean follow-up · 73.7% final win
clean follow-up · 47.9% final win
clean follow-up · 80.1% final win
clean follow-up · 43.0% final win
How we know — models, confidence ranges, and limitations
THE CONTRIBUTION UNDERNEATH
Capture rate still answers the wrong question.
The old 69.8% capture win rate bundled together prior gold, XP, kills, structures, lane economy, composition, Dragon score, objective timing, side, patch, region, and the objective outcome itself. Calling that Dragon value overstated the contribution.
Among capture-like states with comparable waiting examples, taking Dragon now contributes +7.8 final-win points relative to leaving it alive for the next minute. The translation paths above explain whether that advantage survives.
SELECTION → CONTRIBUTION
Where the other eleven points went.
This decomposition is the report’s core result. Adjustment changes the magnitude, not just the disclaimer.
Capture teams versus all eligible wait states.
Gold, XP, kills, structures, lanes, score, timing, composition, side, patch, and region.
Capture now versus waiting one minute on common support.
HETEROGENEOUS EFFECTS
Contribution changes with the Dragon state.
Cohort before captureCapture statesObserved captureModeled waitContribution95% intervalRead
Rows use the global cross-fitted nuisance models and are suppressed below 100 supported captures. “Inconclusive” means the match-clustered interval crosses zero.
WHAT SURVIVES ADJUSTMENT
Early Dragons, Soul, and Elder still move the outcome.
The fourth global Dragon is not automatically Soul; treating ordinal and team stack as the same variable creates another confound.
95% interval +3.7 to +11.1
- Observed capture
- 84.1%
- Modeled wait
- 76.7%
- Interpretation
- Durable
95% interval +7.0 to +25.1
- Observed capture
- 81.5%
- Modeled wait
- 65.6%
- Interpretation
- Large, noisy
ADJUSTMENT CHECK
The estimate shrinks when you see more of the state.
Coarse models overstate Dragon. Full state pulls it down. Missing vision and positions mean the leftover is still not pure causation.
Timing, score, level, CS, and item state only. Leaving out economy and map control inflates the estimate.
Adding gold, XP, kills, structures, lane economy, recent slopes, composition, patch, region, and side removes another 3.8 points.
Vision, exact champion positions, cooldowns, health, and comms are incomplete across the corpus. The remaining 7.8 points cannot be called purely causal.
NOT ENOUGH DATA YET
Starts, contests, and elemental flavor stay unpublished.
These comparisons do not have enough like-for-like games to publish a number.
Only 32.1% of observations sit in usable overlap. Publishing an effect would be misleading.
Only 1.8% overlap and 180 detected contests. No fair comparison yet.
Observed type win rates span 68.5%–70.7%. Type is tangled with ordinal and Soul-rift state, so this is not proof that Hextech wins more.
THE DEEPER RESULT
Dragon still helps after comparing similar game states.
Stronger teams naturally take more Dragons. After accounting for the game state we can measure, taking Dragon now still adds roughly eight modeled win-chance points in capture-like positions.
The result is not proof that forcing Dragon is always correct. Vision, exact positions, cooldowns, health, and comms remain incomplete.
METHOD NOTES
What the models compare
dragon-translation-v1.2 · queue 420 · September 2026
Capture now versus wait one minute in similar game states · clean five-minute follow-up = positive gold movement plus a structure