Online gaming has become increasingly popular in recent years.
Currently, the most common business model of online gaming is
based on monthly subscription fees that game players pay to obtain
credits, which allow them to start or continue a journey in the
game's virtual world. Therefore, from the perspective of game
operators, predicting how many players will join a game and
how long they will stay in the game is important since
these two factors dominate their revenue.
This paper represents a pilot study of the predictability of online
gamers' subscription time. Specifically, we study the gameplay hours of
online gamers and investigate whether strong patterns are embedded in
their game hours. Our ultimate goal is to provide a prediction model of
online gamers, which takes a player's game hours as the input and
predicts whether the player will leave in the near future. Our study is
based on real-life traces collected from World of Warcraft, a famous
MMORPG (Massively Multiplayer Online Role-Playing Game). The traces
contain the gameplay histories of 34,524 players during a two-year
period. We believe that our study would be useful for building a
prediction model of players' future game hours and unsubscription
decisions; i.e., decisions not to renew subscriptions.
Online gaming has become increasingly popular in recent years.
In , it is reported that over 55% of Internet users
are now also online gamers. Currently, the most common business
model of online gaming is based on monthly subscription fees
that game players pay to obtain credits, which allow them to
start or continue a journey in the game's virtual world.
Therefore, from the perspective of game operators, predicting
how many people will join a game and how long they
will stay in the game is crucial, since these two factors
dominate their revenue.
Predicting how many gamers will join a game before a game's
launch is very difficult, if not impossible, since it involves
many non-game-related factors, such as the release date of the
game (whether it is launched during the summer vacation), the
artistic design (whether it is comic-like or realistic),
cultural issues (whether it is Eastern- or Western-style), and
even advertising strategies. Predicting how long players will
stay once they join a game is more feasible, as it should
correlate with the extent of users' involvement in the game's
virtual world. Usually, this can be inferred from the players'
external behavior, such as how quickly their avatars advance to
new levels and how long they spend in the game each day.
This paper presents a pilot study of predicting online gamers'
subscription times. A player's subscription time denotes the
length of time since he/she first joined the game to the time of
his/her last login, i.e., the player has not logged in since then.
Specifically, we study the gameplay hours of online gamers and
investigate whether strong patterns are embedded in their game
hours. Our ultimate goal is to provide a prediction model of
online gamers that takes a player's game hours as input and
predicts whether the player will decide not to continue in the
game once his/her current subscription expires. In this paper, we
use the term unsubscription decisions to describe such
decisions. Predictions about players' unsubscription decisions are
important to game operators because the decisions affect the
operators' revenue directly. Our rationale is that, if we can
predict the subscription time of players before they actually
leave a game, the game operator can take remedial action to
prevent the players' departure and improve the game based on
feedback provided by those players.
Predictions about players' subscription time can provide the
Players usually quit a game because they are dissatisfied
with the game's design or content, or even other players'
cheating activities. Thus, to some degree, player
unsubscriptions should indicate low user satisfaction. In other
words, if we can predict which players will leave the game in
the near future, we may have a chance to stop them leaving, or
at least understand their reasons and make future improvements.
To this end, operators could conduct surveys to determine the
causes of player dissatisfaction and improve the game
accordingly. However, it is more likely that operators would
receive useful comments because dissatisfied players who have
been totally disappointed with a game may reluctant to take
surveys from game companies.
The predictions about players' subscription times also
facilitate predictions about the number of future players. Even
though we can predict the number of players directly by using time
series modeling , subscription time prediction provides
more information because we can predict "which" players will
leave the game rather than just "how many" players will leave.
With such information, game operators can plan their network and
server allocation beforehand and optimize resource arrangements in
Our study is based on real-life traces collected from World of
Warcraft , a famous MMORPG (Massively Multiplayer
Online Role-Playing Game). The traces contain the game play
histories of 34,524 players during a two-year period. Our
results indicate that, although short-term prediction is
feasible, long-term prediction is much more difficult because
players may become more involved in the game or lose interest
The remainder of this paper is organized as follows.
Section 2 contains a review of related works. In
Section 3 we summarize our traces and describe
the collection methodology. We analyze how much time gamers spend
playing the game in Section 4 and when they play
the game in Section 5. In Section 6,
we evaluate the the feasibility of using players' short-term game
hours to predict their long-term gameplay behavior. Then in
Section 7, we summarize our findings and discuss
possible avenues of future research.
2 Related Work
In a previous work that focused on an MMOG called RockyMud
, the authors collected a set of traces of session
inter-arrival times, session lengths, avatars' transition probabilities
between different regions, and region stay times. The authors' analysis
showed that the inter-arrival times of game sessions follow an
exponential distribution. In addition, the transition of avatars between
different regions can be well modelled by a first-order Markov chain,
while the region stay time and session length can be described by a
Pearson distribution and a Pareto distribution respectively.
Based on a set of World of Warcraft traces, Pittman et al
attempted to propose a realistic, empirical model for simulating
users' gameplay behavior and the fluctuations in game servers'
popularity over time . The authors conjectured that at
least four types of information are required to establish a
prediction model: 1) the server's population changes over time; 2)
the arrival rate and session duration of players; 3) the spatial
distribution of avatars in the virtual world; and 4) the movements
of avatars over time (how many distinct regions the avatars visit
and how long they stay in a region). They observed that the number
of players fluctuated in a diurnal pattern and there can be an
approximate 5-fold increase in the number of players between 4 am
and 6 pm. In addition, they found that session times appeared to
follow a power-law distribution where approximately 50% of the
gamers remain online for 10 minutes or less. They also discovered
that the number of players versus the rank of each zone, from the
most populated to the least populated, exhibited a power-law
Chambers et al  conducted a user behavior study of
Counter-Strike, a famous FPS game. Their work focuses on two
issues: users' satisfaction with a game, and the predictability of
the game server's workload. They analyzed the number of connection
attempts and session times, and found that it is extremely
difficult to satisfy users. If a game server is not stable, gamers
tend to go elsewhere without considering "loyalty". Chambers et
al. also found that users have short attention spans, and users'
session times are usually shorter than one hour. They also
analyzed the popularity of game servers and found that the number
of users on different servers follows a power-law distribution.
Moreover, the server workload exhibits predictable patterns in
terms of day and week scales, but the predictability diminishes
with larger time scales.
3 Data Description
In the following, we introduce World of Warcraft, and describe how
we collect players' game hours in an automated fashion. We
conclude this section with a summary of the collected traces.
3.1 World of Warcraft
World of Warcraft is the fourth game set developed by Blizzard
Entertainment Incorporation, and it is currently the most popular
MMORPG in the world. According to MMOGChart , the
game's 10 million subscribers accounted for 62% of the MMOG
market in May 2008 . Because of its high popularity, it
has become a field for researchers to study psychology
, social behavior [8,12], and game play
3.2 Data Collection
We used the who command, which is publicly available to
every player in the game, to collect our traces. The command asks
the game server to reply with a list of players who are currently
online. Thus, anyone can obtain the gameplay history of all the
users on a server by issuing the who command with a
regular interval. To do so, we create a character on a World of
Warcraft server and keep it online all the time. Our character is
controlled by a program and automatically collects a list of the
online users every 10 minutes. If a player logins and logouts
within 10 minutes, we may not be able to observe his/her re-login
activity in consecutive snapshots. However, we do not think this
problem is significant because most WoW session times are much
longer than 10 minutes .
For scalability consideration, the World of Warcraft server restricts the
number of users returned by a query to a maximum of 50 accounts.
Thus, we have to narrow down our query ranges by dividing all the
users into different races, professions, and levels. For example,
we need to first ask the server to list all the users with the
"Fighter" class with the first query, and then ask the server to
list all the users with the "Wizard" class with the second
query, and so on. This technique allows us to systematically list
the entire set of online players despite the restriction of the
3.3 Trace Summary
We collected our traces from Dec. 2005 to Oct. 2007. During the
monitored 664 days, 34521 accounts are observed, as shown in
Table 1. However, only 7043 of those accounts
remained active for more than 30 days, which indicates that most
accounts were never used after the free trail period expired. As
we focus on the long-term gameplay patterns of WoW players, in our
analysis, we only use the 7043 accounts whose subscription periods
are longer than 30 days.
Table 1: Trace description
4 How Long Do Gamers Play?
In this section, we examine "how long" gamers play from various
in terms of the overall subscription time, consecutive gameplay
days, and daily gameplay activity.
4.1 Subscription Time
Figure 1: The survival curve of players' subscription
In this study, we consider that a player has quit a game if he/she
does not login into the game for three months. Note that some
players' subscription periods are censored, i.e., some players
started playing WoW before our measurement started, and some
continued playing after our measurement ended. Thus, we cannot
directly estimate the distribution of players' subscription times
by a cumulative distribution function (CDF). Instead, we use the
Kaplan-Meier estimator , which takes account of the
censored status of each subscription period, to estimate the
distribution of players' subscription times. The Kaplan-Meier
estimator's output is called the survival function, which
reduces to the cumulative distribution function if none of the
subscription periods are censored.
The survival function of players' subscription times is shown
inbscribe to WoW continuously for longer than one
Fig. 1. We observe that, probabilistically, 60%
of users will suyear after their first visits, while 50% of users
will subscribe for longer than 500 days. This result indicates
that the game is indeed very attractive game, and most players
seem to become addicted to its fantasy world once they become
creatures in it.
4.2 Consecutive Game Play Days
Intuitively, if users regularly play the game every day for a long
period, they are probably addicted to the game. Thus, we consider
the distribution of consecutive game play days in order to
understand the extent of addiction of WoW gamers. We define an ON
period as a group of consecutive days during which a player joins
the game everyday, and an OFF period as the interval between two
Figure 2: Cumulative distribution functions of
ON/OFF periods and season lengths
Fig. 2(a) is the cumulative
distribution function of the length of ON and OFF periods. We
observe that OFF periods are slightly longer than ON periods on
average, but the difference is insignificant. In addition,
probabilistically, around 80% of the gamers' ON and OFF periods
are shorter than 5 days. In other words, players tend to alternate
between ON and OFF periods shorter than 5 days. This might be due
to MMORPG's addictive characteristics; that is, a player may not
like to leave the game for a long time, as doing so may cause him
to lose the sense of playing a role in the game world, and become
less familiar with the virtual world. Therefore, players tend to
come back to the game world frequently to continue their onward
journey or simply to keep company with their partners or guild
We observe that some OFF periods are extremely long; for example,
3% of OFF periods are longer than 1 month, and 1% are longer
than 3 months. This may be due personal reasons that force gamers
to stop playing the game for a long period, such as preparing for
exams, beginning a new job, or running out of money to purchase
subscription credits. In addition, we find that, even after a long
OFF period, gamers may come back and play the game as seriously as
before. Hence, we need to divide a player's subscription time into
a number of active periods, where two adjacent active periods are
separated by a long rest period from the game. We call each active
period a season, and a long rest period between two seasons
a vacation. More specifically, we define a vacation as an
OFF period that is longer than 30 days, and a season as an active
period between two vacations.
The cumulative distribution functions of the lengths of seasons
and vacations are shown in
Fig. 2(b). From the graph, we find
that vacations are generally longer than seasons, but the
difference is not significant. Furthermore, we find that around
50% of the seasons are longer than 60 days. This indicates that
WoW gamers tend to become addicted to the game, so it is common
for them to spend longer than 2 months without a vacation during
their adventure in the game's virtual world. In addition, we can
see that less than 20% of vacations are longer than 180 days,
which indicates that, after a vacation longer than half a year,
only 20% of the gamers will return to the game. We also observe
that around 20% of the seasons are shorter than 10 days, which
indicates that some gamers will come back from a vacation to join
the game for only a few days and then take another vacation.
4.3 Daily Activities
Here, we consider the characteristics of users' daily behavior,
including the average daily playtime, average daily session count,
and average session playtime. Note that if a gamer does not play
the game for some days, we do not include those days in his
average daily playtime. For example, if a gamer's subscription
time is a year, during which he only played for 200 days, then his
average daily playtime will be his overall playtime divided by 200
Figure 3: CDF of daily playtime and session
The CDFs of the average daily playtime and the average session
playtime are shown in Fig. 3(a). We
find out that 75% gamers play longer than 1.9 hours per day on
average, and 25% longer than 4.9 hours per day, which indicates
that the game is very attractive for its gamers. If we analyze the
average session playtime, we find significant "knees" around 1
hour and 5 hours, which indicates that after logging into the
game, there is a high probability that players will stay for at
least one hour, but usually no longer than 5 hours. Because of the
long session property, players probably do not login into the game
too many times a day; hence the daily session count is not large,
as shown in Fig. 3(b), where more
than 80% of gamers' session counts are less than 2 per day on
average. We summarize the quantiles and averages of the average
daily playtime, average session playtime, and average daily
session count in Table 2.
Quantiles (5%, 25%,
50%, 75%, 95%)
Session time (hr)
(0.4, 1.0, 1.8, 3.0, 5.5)
Daily session count
(1.0, 1.1, 1.4, 2.1, 3.3)
Daily playtime (hr)
(0.5, 1.6, 3.1, 5.1, 8.8)
Table 2: Summary of daily activities
5 When Do Gamers Play?
Figure 4: Bar chart of login count versus day and
bar chart of login count versus hour
We now consider the question: When do gamers play?. Our analysis
is based on the day scale and week scale, i.e., whether gameplay
occurred during the night or in the daytime, and whether it
occurred on weekdays or weekends. The results are shown in
Fig. 4. Intuitively, we might think that the
average daily playtime of the gamers on weekends would be higher
than that on weekdays, and playtimes of each weekday would be
similar to each other. However, the results do not support our
intuition. The average daily playtime on the weekends is indeed
higher than that on weekdays, but the difference is not
significant. This might be due to two reasons: 1) WoW is such an
attractive game that users play every night, even if they have to
work the next day; and 2) it is much more fun to play a MMORPG
like WoW with partners. WoW encourages multi-party gameplay by
providing many missions and dungeons that are difficult so that
only teams of players can conquer. For example, for a strong
"boss," players often need to gather at least two fighters, one
wizard, and one priest to defeat it. The fighters concentrate on
attacking the boss, the priest takes care of the damages caused by
the boss, and the wizard keeps casting protective magic on
partners and damaging magic on the enemy.
Furthermore, Fig. 4 shows that the average daily
playtimes for each weekday are significantly different. This may
be because, as the weekend draws closer, gamers start to extend
their playtimes, so the average daily playtime begins to increase
from Thursday. After the weekends, the game's attraction
continues, so gamers cannot concentrate on their work, and play
the game whenever they can, even during working hours. This effect
is the lowest on Wednesday, and starts to increase as the weekend
With regard to playing hours a day, we observe that 1) there is an
obvious difference between the number of gamers during night hours
and morning hours. The number of gamers begins to increase most
rapidly around 6 pm, which indicates that most gamers begin to
play immediately after they finish work. The number of gamers
reaches a peak from 10 pm to midday, and is the lowest from 5 am
to 7 am. 2) The number of gamers increases from 6 am to 10 pm;
hence, even during working hours, players continue to join the
game. This may be because students skip classes to play the game,
or workers play secretly during business hours.
6 Predictability Analysis
Figure 5: Predictability of daily playtime, daily session
count, and session time
In this section, we investigate whether users' gameplay behavior
is predictable, i.e., can we predict players' future game hours
based on their gameplay history. Our analysis is comprised of two
parts. In the first part, we analyze whether players' short-term
behavior can be used to predict their long-term behavior. In the
second part, we assess whether temporal dependence exists between
consecutive time periods in four different time scales, namely
days, weeks, ON periods, and seasons.
6.1 Predictability of Short-term Behavior
To determine whether players' short-term behavior is a reliable
indicator of their long-term behavior, we use the average session
time, average daily session count, and average daily playtime as a
summary of players' short-term behavior. We expect that some
variables, such as the average length of ON periods, the average
season length, and the overall subscription time may correlate
with players' long-term behavior.
Fig. 5 shows the plots of the correlations
between the three short-term behavioral factors and the three
long-term behavioral factors. We observe that the lengths of the
average ON periods are moderately correlated with all the
short-term behavioral factors, and the average daily play time has
the strongest predictability. Fig. 5(c) shows
that, if players' average daily game time is shorter than 1 hour,
then their average ON periods will probably be less than 2 days,
i.e., these players tend not play the game for three consecutive
days. On the other hand, the average daily playtime of highly
addicted players can be as high as 10 hours, and they may play the
game for more than 20 days without interruption. However, it is
clear that the average length of seasons and the overall
subscription time do not correlate with all the short-term
behavioral factors. Since this indicates that players' interests
may change significantly over time, we cannot simply use an
overall average of players' short-term behavior to predict their
long-term gameplay behavior. Instead, we need to monitor the
evolution of players' game hours over time and keep track of their
interest in the game  in order to accurately predict
when unsubscription will occur. We will consider this issue in our
6.2 Players' Game Hours in Consecutive Periods
Figure 6: Strength of the temporal dependence in day, week,
ON period, and season scales
We also consider the temporal dependence of players' game hours in
consecutive periods. In other words, we examine whether players'
gameplay behavior in one time period will be carried over to the
following period. As shown in Fig. 6, five types
of time periods are considered: session, day, week, ON period, and
season. Not surprisingly, the overall playtime between consecutive
weeks exhibits the strongest auto-correlations among all the time
scales we consider. Session time and daily playtime are also
strongly auto-correlated; however, the magnitude is not as strong
as that of weekly playtime. The reason may be that the weekly
patterns are the most regular for most people, while session times
and daily playtimes are more easily affected by events and the
different schedules on weekdays and weekends. On the other hand,
the auto-correlation of ON period playtime is also moderate,
although the length of consecutive ON periods is less regular. The
season length has no auto-correlations at all, which we consider
reasonable as consecutive seasons are actually separated by a rest
period longer than 30 days. Moreover, a season might be long
enough to affect or change players' interest in the game. This
implies that the prediction of players' unsubscription should be
performed in a time scale shorter than a season.
Table 3 summarizes the findings discussed in this
ON period length
Table 3: Predictability of daily playtime, session time, session
count, and temporal dependence in day, week, ON period, and season
[*] The symbols represent the correlation strength.
∗∗∗: strong correlation (cor ≥ 0.8);
∗∗: medium correlation (0.8 > cor ≥ 0.5);
∗: weak correlation (0.5 > cor ≥ 0.3);
×: no correlation (0.3 > cor).
7 Summary and Future Work
In this paper, we study players' game hours for a famous MMORPG,
the World of Warcraft, during a 2-year period. We analyze when
gamers join the game's virtual world and how long they stay in the
game. In addition, we investigate whether players' future game
hours can be predicted by their previous behavior. Our results
indicate that although short-term prediction is feasible,
long-term prediction is much more difficult as players' interest
in the game may increase or decrease significantly over time.
In spite of the difficulties involved in prediction, we will
continue with prediction modelling. Our goal is to construct a
model that can predict whether a player will leave a game in the
near future. Predicting players' future behavior (in terms of
leaving or staying in a game) would be advantageous to game
operators as it would help them prevent the loss of subscribers
and enable them to determine how to improve the game by surveying
players that lose interest in it.
Blizzard Entertainment Inc.
World of Warcraft.
Blizzard Entertainment Inc.
World of Warcraft Reaches New Milestone: 10 Million Subscribers.
Estimating A Distribution Function for Censored Time Series Data.
J. Multivar. Anal., 78(2):299-318, 2001.
C. Chambers, W. chang Feng, S. Sahu, and D. Saha.
Measurement-based Characterization of A Collection of Online Games.
In IMC'05: Proceedings of the 5th Conference on Internet
Measurement 2005, pages 1-14, 2005.
W. chang Feng, D. Brandt, and D. Saha.
A Long-term Study of A Popular MMORPG.
In NetGames '07: Proceedings of the 6th ACM SIGCOMM Workshop on
Network and System Support for Games, pages 19-24, 2007.
Z. Z. Eric Wan, Xin Xu.
2006 Online Game Report.
In Pacific Epoch Red Innovation Report Series, 2006.
T. Henderson and S. Bhatti.
Modelling User Behaviour in Networked Games.
In MULTIMEDIA '01: Proceedings of the Ninth ACM International
Conference on Multimedia, pages 212-220, 2001.
V. H. hua Chen and H. B.-L. Duh.
Understanding Social Interaction in World of Warcraft.
In ACE '07: Proceedings of the International Conference on
Advances in Computer Entertainment Technology, pages 21-24, 2007.
J. Kim, J. Choi, D. Chang, T. Kwon, Y. Choi, and E. Yuk.
Traffic Characteristics of A Massively Multi-player Online Role
In NetGames '05: Proceedings of 4th ACM SIGCOMM Workshop on
Network and system Support for Games, pages 1-8, 2005.
M. Kwok and G. Yeung.
Characterization of User Behavior in A Multi-player Online Game.
In ACE '05: Proceedings of the 2005 ACM SIGCHI International
Conference on Advances in Computer Entertainment Technology, pages 69-74,
Total MMOG Active Subscriptions.
Collaborative Play in World of Warcraft.
In LA-WEB '06: Proceedings of the Fourth Latin American Web
Congress, page 3, 2006.
D. Pittman and C. GauthierDickey.
A Measurement Study of Virtual Populations in Massively Multiplayer
In NetGames '07: Proceedings of the 6th ACM SIGCOMM Workshop on
Network and System Support for Games, pages 25-30, 2007.
P. Svoboda, W. Karner, and M. Rupp.
Traffic Analysis and Modeling for World of Warcraft.
In ICC '07: IEEE International Conference on Communications
2007, pages 1612-1617, June 2007.
Expert: 40 Percent of World of Warcraft Players Addicted.
In tom's GAMES, 2006.
1. This work was supported in part by
National Science Council of the Republic of China under the grant
Sheng-Wei Chen (also known as Kuan-Ta Chen) http://www.iis.sinica.edu.tw/~swc
Last Update August 01, 2019