After 128 hours of DotA 2, 350,513 creep kills, 9067 hero kills, 3281 tower kills and 161 courier kills; The International 2017 is finally concluded. Our predicted favourites Team Liquid clinched the title with a lower-bracket run of epic proportions.
At Stratz we successfully anticipated Liquid being the strongest team present, yet how did our other predictions fare?
Main Event Predictions
We scored 9 / 22 in total, including correctly predicting the champions, Team Liquid. 9 / 22 does not sound impressive, however very few people expected to see Liquid and EG both fall out of the upper bracket on day 1.
These upper bracket games are of vital importance to the other predictions. As you can see, as soon an upper-bracket prediction fails, all the lower bracket predictions are thrown off as well.
Applying our predictions to individual games, rather than the bracket as a whole, would give you 73% accuracy, with only 6 incorrect series.
Where can we improve?
Our TI7 team predictions were based on quite a simplistic model. There is considerable room for growth. We previously illustrated the main weaknesses of all professional DotA 2 team rankings or predictions is region agnosticism. By ignoring whether a team is Chinese or European, we risk undervaluing all teams of some region.
It’s quite clear that nearly everybody undervalued Chinese DotA heading into TI7. Three of the top four TI7 teams were Chinese. We correctly predicted LFY to be up there, however Newbee and LGD surprised us. Incorporating region status into our model will be top priority for the next major DotA 2 event.
Another idea that we need to investigate is ‘lower bracket momentum’. Based on TI7, it seems that despite their apparent skill, the existing lower bracket teams have a mental advantage going into the next series. Evil Geniuses, TNC, Invictus Gaming, Virtus Pro and LFY were all eliminated the very next series after slipping into lower bracket.
Only two teams fell from upper-bracket and didn’t immediately lose:
- Team Liquid (Even here Liquid lost game 1 to Secret and came very close to elimination. Liquid had to struggle harder to win their first lower bracket game than the grand-final)
- LGD-Gaming (Their signature Magnus offlane was 1st phase banned by VP for both upper-bracket games. When DC allowed LGD to have their beloved Magnus pick, perhaps this countered the ‘upper bracket elimination hangover’)
Over TI5, TI6 and TI7 lower-bracket teams have a 15–6 record against recently relegated upper-bracket teams. This is quite frankly astonishing, given that upper-bracket teams should technically be the strongest. Whilst we have not yet quantified this effect of lower-bracket momentum, it will definitely be under consideration next event.
Compendium Predictions
Tournament Predictions
Following our advice would have yielded 50% correct tournament predictions. An impressive number with multiple categories available for each.
We highlighted our reasoning for our picks in our previous article, therefore will not revisit correct ones. However we are very proud of the “Total Number of Games Played at Main Event” prediction. Many people guessed half best-of-3’s end 2–0, half 2–1; giving a prediction of over 50 main event games. We tested this, proving it false, when we went through past major DotA 2 events and found that 2–0 is actually more likely than 2–1, enough to make 45–49 the likeliest scenario.
“Shortest Game” and “Combined Total Kills” were so close that they were still reasonable educated guesses. “Longest game” is another very hard to predict category.
How can we improve?
We failed to predict both the number of unique heroes both picked and banned. Our flaw was using predominantly non-TI events to estimate the expected pick/ban numbers. With TI5, TI6, and TI7, we’ve seen a definite pattern of more hero variety than any other tournament. 104 unique heroes were picked at TI5, 105 unique heroes were picked at TI6, and 107 at TI7!
Why do more heroes get picked at TI than at Majors?
There are two causes
- Higher number of total games
- Deeper team preparation
The first reason is fairly obvious. More games equals more drafts for seldom picked heroes to excel in.
The second reason is due to the stakes at TI being far higher than any other tournament. If your team can work out how to incorporate an overlooked hero into your draft, other teams will not be prepared to face that hero, or understand how to beat it. Each team strives to gain any possible edge over foes at TI, leading to even more experimentation pre-TI and a greater pool of heroes each player feels comfortable with.
Player and Team Predictions :(
For team and player compendium predictions we only got two correct. However this shouldn’t be atypical. 90 unique players and 18 teams at TI7 make it hard to be spot-on.
We correctly predicted tournament winners Liquid, arguably the most important compendium prediction, rewarding 2500 points.
We also anticipated “player with highest assist average” to be LFY.Ahfu. We were just one position off from picking the “player with lowest death average” [Ame narrowly beat Ramzes by only 0.1 deaths!].
How can we improve?
In retrospect, while we feel many of our player and team predictions failed due to random chance, we think our “Team/Player that picks the most different heroes” entries were flawed.
Following VP’s disappointing performance at Epicenter, many fans accused them of only being able to play Magnus, Warlock, Troll, Ursa and a handful of other heroes. So at the Summit 7 VP proved people wrong by never picking the same hero twice throughout the whole tournament. They won the tournament despite picking 50 unique heroes.
Due to our predictions analyzing games at The Summit 7, it assumed VP would continue this wacky trend. In hindsight it’s clear this was a one-off from VP, not representative of their drafting style at other tournaments. If we had removed Summit 7 VP games from our models, Newbee correctly come out as the “Team that picks the most different heroes”.
We must utilize our human knowledge of the professional DotA 2 scene, to override our computer-generated model, where necessary.
Overall, our compendium prediction logic appears sound. We got many tournament predictions correct, and were close enough on player and team predictions that, with slightly different luck, would have appeared exceptional.
We cannot wait for the next DotA 2 major to bring you even more accurate statistically-based compendium predictions.
