How the draft grade is calculated
The grade and the win chance shown after a draft come from a small statistical model. This page lists what goes into it, how it was checked against real ranked matches, and where it falls short.
What the grade means
After a draft is complete DotaPicks estimates how likely each side is to win and shows the result as a win chance, for example 54.2%. The letter grade ranks that estimate against real ranked drafts.
A figure such as +4.2% next to a hero suggestion is the change in your win chance if you add that hero. For a ban it is the change in the opponent's win chance if they could not have that hero.
This is a statistical estimate, not a prediction. It describes how drafts like yours have tended to do, and it does not know how well anyone plays.
What goes into the number
The win chance combines three inputs, each measured in percentage points from your side's point of view and weighted as below. The weighted sum is turned into a probability with a logistic curve, and it is scaled down while fewer than ten heroes are picked.
| Input | Weight |
|---|---|
| Counter advantage Average matchup edge of your heroes against theirs | 0.060 |
| Synergy difference Your team's synergy minus the enemy team's | 0.050 |
| Hero win rate difference Your heroes' average win rate minus theirs | 0.080 |
How it was calibrated
The model was checked on 341,947 ranked All Draft matches played after the hero data snapshot of Oct 2, 2026. Its accuracy is measured on the 102,585 newest matches, which it was never fitted on.
| Measure | Value |
|---|---|
| Matches (fit and holdout) | 239,362 + 102,585 |
| Match window | Oct 2, 2026 – Oct 3, 2026 |
| Hero data snapshot | Oct 2, 2026 |
| AUC | 0.589 |
| Log-loss | 0.6809 (always guessing 50%: 0.6931) |
| Calibration error (ECE) | 0.031 |
Matches are split by time, and every figure above is measured on the 102,585 newer ones. Only matches that started at least 12 hours after the snapshot were used, so the snapshot cannot already contain their results. The weights in use were set by hand, not fitted on these matches. A refit on the 239,362 older matches was compared against them and not adopted.
Refitting the weights on the older matches gives an AUC of 0.595 against 0.589 for the current weights, and a log-loss of 0.6793 against 0.6809. A bootstrap puts the AUC gain between 0.0051 and 0.0065.
In the refit the counter advantage carries all of the weight and the other two inputs get none.
A refit is adopted only if its log-loss is no worse, its AUC is not lower by more than 0.001 and its calibration error is no worse. The refit did not meet all three, so the weights listed above stay in use.
Calibration curve
Each point groups holdout matches by predicted Radiant win chance. Its height is how often Radiant really won. A perfectly calibrated model sits on the diagonal. Groups with very few matches are noisy.
Show the numbers
| Predicted range | Mean predicted | Observed | Matches |
|---|---|---|---|
| 10–20% | 19.8% | 0.0% | 1 |
| 20–30% | 28.0% | 34.3% | 423 |
| 30–40% | 36.8% | 39.8% | 10,419 |
| 40–50% | 45.6% | 48.6% | 40,768 |
| 50–60% | 54.4% | 57.4% | 40,414 |
| 60–70% | 63.2% | 66.2% | 10,087 |
| 70–80% | 72.0% | 80.9% | 471 |
| 80–90% | 80.1% | 100.0% | 2 |
The letter scale
The grade compares your predicted edge, your win chance minus 50 points, with the edges of real ranked drafts from the holdout.
| Grade | Beats this share of real ranked drafts | Minimum edge in points |
|---|---|---|
| A+ | 97% | +14.7 |
| A | 92% | +11.1 |
| A− | 85% | +8.2 |
| B+ | 75% | +5.4 |
| B | 62% | +2.4 |
| B− | 50% | −0.0 |
| C+ | 38% | −2.5 |
| C | 25% | −5.4 |
| C− | 15% | −8.3 |
| D | 5% | −13.0 |
| F | The rest | |
An A means a larger predicted edge than 92% of real ranked drafts.
How mistakes are measured
The cost of a move is how much win chance the final 5v5 would have gained if you had made the better choice at that turn and everything else stayed the same.
The better option is the first hero the suggestions offered at that turn that could fill an open position and is either the hero you chose or in neither final lineup. Following the top suggestion therefore never costs anything.
- Pick cost: the win chance of your final lineup with your hero swapped for the better option, minus its win chance as drafted, never below zero.
- Ban cost: for the hero you banned and for the better option, find the largest win chance the opponent's final lineup could have gained by swapping one of its heroes for that hero. The cost is how much larger the better option's gain is, never below zero. It measures what the opponent could have gained, not what they would have picked.
Only moves that cost at least 0.5 percentage points are listed as mistakes. Costs are measured on the final lineups so that every comparison is a complete 5v5 inside the range the model was calibrated on, and heroes already in either final lineup are never counted as the better option.
Limits
- The model was calibrated on ranked All Draft matches, not on Captains Mode, so a grade for a Captains Mode style draft is an extrapolation.
- An AUC of 0.589 is a modest edge over the 0.500 of a coin flip. A draft is only one part of a match, and the model says nothing about how the match is played.
- Win chances are shown between 32.1% and 67.8%, the range the model was calibrated on. A value outside it is capped and marked.
- The calibration window is short, so the weights should be refitted as the game and its patches change.
- A shared draft link holds the picks in the link itself and is not tamper-proof. Anyone can edit one, so a grade you receive is not proof of how a draft was played.
- Beating the Easy bot, which picks at random from its top 15 suggestions, inflates the grade. The grade does not adjust for the bot's level.