Chatter Plots: Game Duration and Proxies in WCS Americas

Published on 04/25/2019 05:07 PDT by Firnafth

Now that WCS Winter has finished, we can look at some overall playstyle trends throughout the season!

Data Collection

We collected data by watching VODs of all games played for 2019 WCS Americas Winter (n=235 games). For each game, we notated the players, matchup, winner, and game duration; during the games we also notated the first four nonsupply buildings built by each player and any proxies attempted; and we tallied all attacks, their initiator, whether the attack was harassment or not, and its general unit composition as well as the unit composition used to defend against it.

For fully detailed methods, see this document.

For raw data, see this spreadsheet.


Game Duration

The average game lasted just over 12 minutes long, but the median game was just under 11 minutes long since there were a few very long games.

Statistically, we tested whether game duration differed between matchups. We took the average duration of all games between the same players to prevent the results from being skewed by players with more games (n = 80 average match durations). The unequal variances and sample sizes of the groups indicated use of a Kruskal-Wallis Test for the initial analysis and Dunn's Test to compare each pair of matchups (R package PCMPCplus).

We also tested whether game duration differed by map, averaging the few games between the same players that occurred on the same map (n = 217 average game durations). We normalized these data with a square root transformation and conducted an ANOVA (R package stats) to test for statisically significant mean differences.

The p-values we get from our tests are the probability that something is statistically insignificant given the data we have observed. If the p-value is low then it is highly unlikely that the difference in means we have observed and tested is statistically insignificant.


Game duration differed significantly between matchups  (p < 0.0001).

Specifically, TvZs lasted longer than every other matchup with statistically significant differences:

TvZ Duration Compared top-value

There was no significant difference in game duration between maps (p = 0.32)


These are the most commonly proxied buildings for each race. Zerg players threw out Nydus Worms in nearly a quarter of their games, but only rarely proxied hatcheries. Terrans proxied barracks in about 10% of their games. It's worth noting that the 6% of games with a proxy bunker include not only early rushes but also all-in two-base pushes. Protoss players proxied pylons in a third of their games, largely because they tend to rely on warp prisms for reinforcing much of the time.

On the left are the raw counts of buildings proxied throughout the entire tournament, and on the right the average number of each building created in a game where it was proxied, plotted against the percent of overall games that building was proxied in. Cannon rushers built (or attempted to build) just over 4 photon cannons per cannon rush on average; and when zerg players used Nydus Worms they tended to build nearly three total, although I suspect this is an average of a broad range where sometimes they only used one because they won instantly, or several because they were applying continual pressure.

This is just the first of multiple articles we are writing on this tournament, so keep your eyes out for more!