Predicting Success of
Steam Games
MICHAL TRNĚNÝ MASARYK UNIVERSITY, 2016
Task
Investigate how well can be post-release success of games predicted from basic information
about games (available before release)
No data regarding previews, early reviews, user activity on social networks…
Steam
Largest platform for selling
PC games
Since 2004
Over 10,000 games in total
Steam Charts
Tracks how many players are in
game every hour
Monthly statistics
Since July 2012
Not very accurate measure of
success (?)
(however, Steam Charts is the
best source with sufficient
history)
SteamSpy
Tracks number of
owners
Since April, 2015
No access to history
Manually collected
data for about 1,400
games
Still not a good
measure of success?
Data
About 4,700 games from July, 2012 – July, 2016
Completely omitted Early Access titles and Free-to-play titles
Info such as genres, price, release date, descriptions, languages, features
+ Thumbnail and screenshots of every game
K-means used to extract 5 most dominant
colors
First screenshot and thumbnail
Screenshots
Subsetting
Pick only games whose publisher has a history of at least 2 games
Covers 30-40 % games
Notable attributes
Previous games: how many, max and min players of any game, Gini index
Developer, Publisher (limited to max 53, rest is “other”)
text descriptions
Price
HW requirements: disk space, RAM, GPU
Languages
Number of colors on thumbnail and first screenshot + dominant thumbnail color
Results
Classification (10 classes) Random Forest Regression (values 0-15) SVM
Correlation coefficient 0.80
MAE 1.24
RMSE 1.67
Accuracy 46 % / 44 % within +-1 from actual 53 % / 52 %
within +-1 from actual 87 % / 84 % within +-2 from actual 82 % / 77 %
Training, validation, test set (60 %, 20 %, 20 %)
Classification
Reference
Prediction 1 2 3 4 5 6 7 8 9 10
1 79 32 8 4 0 0 0 0 0 0
2 42 43 21 5 3 0 0 1 0 0
3 0 2 1 0 1 0 0 0 0 0
4 1 0 4 3 5 0 1 1 0 0
5 1 0 3 1 2 0 0 0 2 0
6 1 1 0 1 3 2 2 0 0 0
7 0 0 0 1 1 2 0 0 1 1
8 0 0 0 0 0 0 0 1 0 0
9 0 0 0 0 0 0 0 0 0 0
10 0 0 0 0 0 0 0 0 0 0