Master’s Programme in Mathematics and Operations Research

Win probability estimation for strategic

decision-making in esports

xPetu

Master’s Thesis
2024
© 2024 xPetu.
This work is licensed under a “CC BY-NC-SA 4.0” license.
Copyrights of elements related to League of Legends™
in this work are owned exclusively by Riot Games, Inc.
Author xPetu
Title Win probability estimation for strategic decision-making in esports
Degree programme Mathematics and Operations Research
Major Systems and Operations Research
Supervisor Prof.
Advisor D.Sc. (Tech.)
Date 26 August 2024 Number of pages 38 Language English
Abstract
Esports, i.e., the competitive practice of video games, has grown significantly during
the past decade, giving rise to esports analytics, a subfield of sports analytics. Due
to the digital nature of esports, esports analytics benefits from easier data collection
compared to its physical predecessor. However, strategy optimization, one of the focal
points of sports analytics, remains relatively unexplored in esports. In traditional
sports analytics, win probability estimation has been used for decades to evaluate
players and support strategic decision-making.
This thesis explores the use of win probability estimation in esports, focusing
specifically on League of Legends (LoL), one of the most popular esports games in
the world. The objective of this thesis is to formalize win probability added, i.e., the
change in win probability associated with a certain action, as a contextualized measure
of value for strategic decision-making, using mathematical notation appropriate for
contemporary esports. The proposed method is elaborated by applying it to the
evaluation of items, a strategic problem in LoL. To this end, we train a deep neural
network to estimate the win probability at any given LoL game state. This in-game
win probability model is then benchmarked against similar models.
Keywords Esports analytics, League of Legends, win probability estimation, match
outcome prediction, in-game win probability, win probability added
Preface
Greetings, dear reader.

This thesis is the culmination of an adolescent obsession with optimizing video game
strategies. While the version you are reading is complete in its academic content, I have
chosen to omit any personal data to preserve my privacy. Since beginning my journey
as a YouTube content creator in 2016, I have maintained a clear distinction between my
real identity and my online persona, as I have no interest in public recognition—only
in sharing my passion for video games and technology.
I am publishing this thesis with the hope that it may inspire a young League
of Legends player somewhere in the world to pursue academic studies, discovering
the same sense of fulfillment and curiosity that I have found. More importantly, I
encourage everyone to follow their passion, whatever it may be, for life is too short to
spend chasing someone else’s dreams. Finally, I would like to express my heartfelt
gratitude to all my supporters—you have made it possible for me to do what I love,
and for that, I am forever thankful.

San Francisco, 24 September 2024

xPetu

4
Contents
Abstract 3

Preface 4

Contents 5

Symbols and abbreviations 6

1 Introduction 7

2 League of Legends 9

3 Background 14
3.1 Win probability estimation in traditional sports . . . . . . . . . . . 14
3.2 Win probability estimation in esports . . . . . . . . . . . . . . . . . 17
3.3 Artificial intelligence and esports analytics . . . . . . . . . . . . . . 20

4 Methodology 22
4.1 Win probability in a zero-sum game . . . . . . . . . . . . . . . . . 22
4.2 Problem definition and challenges . . . . . . . . . . . . . . . . . . 23
4.3 Contextualized evaluation of actions . . . . . . . . . . . . . . . . . 24

5 Case study 27
5.1 In-game win probability model . . . . . . . . . . . . . . . . . . . . 27
5.2 Contextualized evaluation of items . . . . . . . . . . . . . . . . . . 30
5.3 Summary of the results . . . . . . . . . . . . . . . . . . . . . . . . 33

6 Conclusions 34

References 35

5
Symbols and abbreviations
Symbols
A action space
X state space
𝐷 data set 𝐷 = {𝑑 (1) , . . . , 𝑑 (𝑁) }
𝑑 (𝑖) 𝑖-th data point 𝑑 (𝑖) = (𝑎 (𝑖) , 𝑥 (𝑖) , 𝑧 (𝑖) ) in 𝐷
𝑎 (𝑖) 𝑖-th action in 𝐷
𝑥 (𝑖) 𝑖-th initial state in 𝐷
𝑧 (𝑖) 𝑖-th final state in 𝐷
𝑦 match outcome (1 = win, 0 = loss)
𝑤 estimated win probability
𝑊 initial win probability
Δ𝑊 win probability added

Abbreviations
AGV added goal value
AI artificial intelligence
API application programming interface
DNN deep neural network
ECE expected calibration error
ETM end-of-game tactics metric
LoL League of Legends
MOBA multiplayer online battle arena
RNN recurrent neural network

6
1 Introduction
Esports, i.e., the competitive practice of video games, has grown significantly in
popularity during the past decade [52]. The growing body of research on esports is
highly interdisciplinary [18]. Esports analytics is defined as the use of esports-related
data to assist with decision-making processes arising both in-game and outside of
it [42]. Its physical predecessor, sports analytics, has produced insights that extend
beyond the domain of sports [3, 58]. Due to their digital nature, esports benefit from
easier data collection, a core challenge of sports analytics in the big-data era [30].
Esports analytics has clear potential, but the field is still in its infancy [18, 42].
By definition, the optimal strategy in sports, esports, or any game where two teams
compete against each other, is to maximize the probability of your team winning [35].
Although difficult to quantify, this win probability is implicitly estimated by players,
coaches, and fans during a match [26]. Computational win probability estimation
is one of the cornerstones of modern sports analytics, with research beginning in
the 1960s [24, 36]. Win probability estimates are now regularly used in sports to
guide decision-making [26, 28, 40]. In esports, win probability models have recently
started to gain traction through their use in broadcasts of major events such as the
2023 League of Legends (LoL) World Championship [37]. Despite their prevalence
[17, 27], esports win probability models have not yet been widely utilized in strategy
optimization, a focal point of sports analytics [1].
In this thesis, we study a recurring decision-making problem in esports: Given a
game state and a finite set of actions, which action should a player take to maximize
their team’s win probability? This allows us to consider both high-level decisions (e.g.,
in Counter-Strike: Should we take site A or B this round?) and low-level decisions
(e.g., in LoL: Which item should I buy in this situation?). In both examples, the set of
decision alternatives (actions) is finite; there are two sites in the former and roughly
two hundred items in the latter. The problem definition excludes continuous gameplay
decisions such as movement in real-time games or aiming in shooters.
As any LoL player would tell you, the choice of items cannot be made without
game-specific details and is thus context-dependent. This is often the case with
strategic decisions in esports, which makes it interesting yet difficult to evaluate and
compare the decision alternatives. Moreover, some actions are only taken by players
to secure wins when they are already ahead or as a last resort in desperate situations.
Such biases cloud esports data, which makes simple aggregate statistics such as win
rate unreliable for comparing actions. Thus, a debiased, i.e., contextualized [53],
method of evaluating actions is required.
This thesis explores the use of win probability estimation in esports and proposes
the mean win probability added as a contextualized metric for evaluating actions.
Win probability added, i.e., the change in win probability associated with a certain
action or player, is a commonly used metric in sports analytics for evaluating players
and decisions [28, 40, 53]. The objective of this thesis is to formalize the use of win
probability estimation to support esports decision-making in a general, game-agnostic
manner. Additionally, this thesis aims to elaborate the proposed method by applying it
to the evaluation of items, a strategic problem in LoL, one of the most popular esports

7
games in the world [56]. To this end, we train a deep neural network to estimate the
win probability at any given LoL game state. This in-game win probability model is
then benchmarked against similar existing models.
This thesis is structured as follows. Chapter 2 provides an overview of the core
concepts and terminology of LoL. Chapter 3 establishes a background for the thesis by
reviewing the relevant literature on sports and esports analytics. Chapter 4 describes
the method of evaluating decision alternatives using win probability estimates. In
Chapter 5, this method is applied to real esports data through a LoL case study. Finally,
Chapter 6 concludes the thesis with a summary of the thesis and an outline of future
research avenues.

8
2 League of Legends
League of Legends (LoL) is a competitive computer game developed by Riot Games.
This chapter describes the game to the extent necessary to understand the rest of this
thesis. All elements related to LoL in the thesis (e.g., characters, items, graphics)
are the exclusive property of, and provided courtesy of, Riot Games. Because LoL is
updated periodically, every two to three weeks, some of the details will eventually
become outdated. Nevertheless, the core game concepts provided here remain relatively
reliable. Current and precise game information can be found on the LoL Wiki [23].
At the time of writing, the most recent game update is Patch 14.15.
LoL can be played on different maps, i.e., virtual battlegrounds, such as Summoner’s
Rift and The Howling Abyss. Furthermore, LoL includes multiple game modes, each
with its own set of rules. This thesis, however, focuses solely on the standard
competitive map and mode combination, Classic Summoner’s Rift 5v5. Throughout
the rest of this thesis, the term LoL refers specifically to this game format.

Figure 1: A simplified visualization (minimap) of Summoner’s Rift, the standard

League of Legends map. The gray areas in the bottom-left and top-right corners are
the Blue Base and Red Base, respectively.

In LoL, two teams of five players compete against each other with the objective of
destroying the opposing team’s central structure, the Nexus. The teams are assigned—
either by a matchmaking system or according to tournament rules—a side of the map
to play on. These sides are represented by the colors blue and red. The teams are thus
referred to as Blue Team and Red Team. As seen in Figure 1, each team has its own
Base, which contains the team’s Nexus. Blue Team wins by destroying the Red Nexus,
and vice versa. Players can perform a multitude of actions in LoL, but the outcome of
a match is ultimately determined by the destruction of a Nexus.
In order to reach the enemy Nexus, the Blue Team must first destroy the turrets
protecting the Red Base. Naturally, the Red Team tries to prevent this from happening,

9
while simultaneously attacking the turrets and base of the Blue Team. In standard
play, each player serves a specific role in a team. The five roles are called top, jungle,
mid, bot, and support. These names stem from the allocation of players to specific
parts of Summoner’s Rift (see Figure 1). There are three lanes (top, middle, bottom)
that lead from one base to the other. The area between the lanes is the jungle. The
support usually begins the game in the bottom lane with their team’s bot but eventually
transitions into roaming around the map. This standard lane allocation ensures that all
turrets are protected from the beginning of a match.
Before the game begins, each player must select a virtual character, known as a
champion, to play. This pre-game process is called champion selection, or simply draft,
and it sets the groundwork for the match. With 10 players and 168 unique champions
to choose from, the number of possible draft permutations is 168!/158! ≈ 1.3 · 1022 .
In practice, some champions are picked more frequently than others, resulting in
a distribution of drafts that is concentrated on a relatively small subset of these
permutations. Nonetheless, the draft produces inherent pre-game variation, making
every LoL match unique.
In the game, each player controls their champion—typically using a keyboard
and a mouse—until a Nexus is destroyed or a team forfeits the match. The player is
synonymous with their champion; champions cannot be drafted twice in the same
match. In addition to the five human-controlled champions, each team has an army of
minions, which are comparatively weak, computer-controlled units that automatically
attack any opponents they encounter.

Figure 2: A screenshot from LoL demonstrating champion movement. The blue

circle on the left indicates a movement command for the champion in play, Shen.

The most elementary control mechanism in LoL is a right-click of the mouse,

which causes the champion to move or perform a basic attack, depending on the
location of the cursor. This fundamental pattern of LoL is illustrated in Figure 2. In

10
addition to moving and attacking, players cast their champion’s abilities to interact
with their opponents, their teammates, and other components of the game. Most
champions have five abilities in total: one passive ability, three basic abilities (typically
assigned to the Q, W, and E keys), and one ultimate ability (typically assigned to
R). Unlike basic attacks and movement commands, which can be made virtually
continuously, most abilities have a cooldown, a post-cast timer that limits the use of
the ability. In addition to unique abilities and a distinct appearance, every champion
has characteristic statistics, which dictate, e.g., how fast they move and how powerful
their attacks and abilities are.
The standard lane allocation combined with opposing motives inevitably leads to
combat between the teams. Since LoL is a real-time game, player combat is fast-paced
and requires precise keyboard and mouse control. At the core of combat are two
gameplay elements: health and damage. Basic attacks and most abilities deal damage,
causing their targets to lose health. Once a unit’s (e.g. champion, minion, turret)
health reaches zero, the unit dies. The champion dealing the final instance of damage
to a dying unit is rewarded a kill. Any other champions participating in the kill are
rewarded an assist. Dead champions are temporarily suspended from the game; they
have a death timer ranging from 10 to 60 seconds, increasing in duration as the game
proceeds. During this timer, the dead champion cannot affect the game, creating a
window of opportunity for their opponents to destroy turrets or even the Nexus.
To prevail in combat, players strive to increase their relative strength and gain an
advantage over their opponents. The term strength does not refer to any particular
attribute in LoL; rather, it encompasses the overall power of a champion. A player can
increase the strength of their champion through two types of fundamental resources:
experience and gold, which are gained by, e.g., killing minions, participating in
champion combat, and destroying turrets. The benefits of these resources are indirect;
they translate into increases in champion strength only when certain thresholds are
met. Collecting enough experience will cause a champion to level up. Every level
amplifies a champion’s statistics and abilities. Collecting enough gold allows a player
to purchase items that increase specific statistics and grant unique effects. Some items
provide an active effect, an additional ability for the player to use. Deciding which item
to purchase is difficult because the provided effects are often impossible to compare
directly.

Figure 3: A screenshot of the LoL HUD. Information appearing from left to right
and top-down: champion, level, experience, abilities, summoner spells, current and
maximum health, current and maximum energy (usually mana), items, and gold.

11
 Figure 3 presents a screenshot of the in-game heads-up display (HUD), which
contains critical information about the state of the player as well as the actions available
to them. The following explanation provides an example of LoL terminology in action.
In the case of Figure 3, the played champion is Shen, he is level 14 and has roughly
a third of the required experience for the next level-up. Most of Shen’s abilities are
available (Passive, Q, E, R), except for Spirit’s Refuge (W), which is on cooldown for 4
seconds. Shen has 2471 health out of his maximum of 3003. He has 250 energy out of
his maximum of 400. Shen’s summoner spells are Ignite (D), which is on cooldown,
and Flash (F), which is available. Shen has five items and a Stealth Ward (4) in his
inventory. One of the items, Titanic Hydra (1), grants an active effect that is available.
The item Deadman’s Plate (2) is granting additional effects with 100 stacks. Shen can
also cast Recall (B) and has 448 unspent gold.
In addition to kills and deaths, numerous other events can occur in a LoL match.
The most impactful events and their consequences are listed in Table 1.

Table 1: Overview of the most impactful events in LoL.

Event Consequence

Kill (champion) Killer receives gold and experience.

Assist Assistant receives a share of kill gold and experience.
Death Dead champion is temporarily suspended from the game.

Voidgrub (kill) Killing team deals more damage to turrets.

Rift Herald Killer can summon the Herald to destroy a turret.
Dragon Killing team becomes slightly stronger.
Baron Nashor Killing team becomes stronger for 3 minutes.
Elder Dragon Killing team becomes extremely strong for 3 minutes.

Turret (destruction) Destroyer receives gold and the lane opens up.
Inhibitor Destroying team gets more minions for 5 minutes.
Nexus Destroying team wins the match.

Like most modern esports, LoL has a steep learning curve. For a novice player, the
first step is to learn the abilities and nuances of one champion. However, playing the
game at any decent level requires internalizing the abilities of all 168 champions, the
numerous champion-specific interactions, and the essential game mechanics. Unlike
perfect information games [59] such as chess, the game state in LoL is partially
observable [4]; a team can only see the parts of Summoner’s Rift occupied by their
units, and the rest is obscured by the fog of war. Moreover, a player’s in-game view,
the camera, can only focus on a small, specific area of the map at any given time. The
amount of observable information is thus limited by the player’s ability to move the
camera while performing hundreds of gameplay actions every minute.

12
 While not comprehensive, the above overview covers: 1) the essential concepts
and terminology of LoL, and 2) the complexity and motivation behind analyzing
decision-making in esports. The next chapter reviews past analytical work, covering
studies specific to LoL, as well as broader research on esports and traditional sports.

13
3 Background
In order to understand and appreciate the potential of win probability estimation in
esports, one must acquaint oneself with the existing applications of win probability
estimation in traditional sports. Therefore, this chapter begins with an overview of the
relevant sports analytics literature. Then, we look at recent developments in esports
analytics, focusing on win probability estimation, specifically for League of Legends.
Accurate win probability estimation requires the ability to draw conclusions from the
state of a game. This ability is also crucial in developing artificial agents that can
play games. While not the subject of this thesis, the research on such game-playing
systems provides valuable findings that can be applied to win probability estimation.
Thus, this chapter concludes with a discussion on the parallels between game-playing
artificial intelligence and esports analytics.

3.1 Win probability estimation in traditional sports

Sports analytics is a multidisciplinary field comprising the collection, analysis, and
interpretation of sports-related data [1, 29]. The field is also commonly referred to as
statistics in sports [44]. The primary objective of sports analytics is to gain insights
that provide a competitive advantage on the field of play [1]. On-field sports analytics
include, e.g., the assessment of player performance [41] and the comparison of
coaching tactics [28]. Sports analytics also serves to inform off-field decision-making
in various business-oriented aspects of the sports industry, ranging from optimal ticket
pricing of sports matches to economic assessment of large-scale sports events like the
Olympic Games [30].
The impact of sports analytics can be seen in phenomena such as the three-point
revolution in basketball, where analysis of shot effectiveness resulted in a dramatic
strategic shift at the professional level [35]. The findings in sports analytics also
extend outside the domain of sports [30]. For example, sports data has been used to
exhibit human biases [3, 15] and to support game theoretical predictions of human
behavior [8, 58]. The advent of the big-data era has led to promising advancements in
sports analytics [30]. However, the collection and standardization of data remain core
challenges [41].
Estimating the win probabilities of the competitors is one of the central ideas
in sports analytics. Win probability estimation probably emanates from baseball
analytics, with research dating back to the early 1960s [24]. In addition to baseball,
win probability estimation has been applied to most major sports, with examples
from basketball [28, 49], hockey [6, 34], American football [26, 33], and soccer
[11, 40]. Recently, the advancements in sports data collection [41] and an increase
in the popularity of sports betting [48] have motivated research on win probability
estimation. Applications of win probability estimation vary from descriptive (e.g.
performance evaluation and prediction) to prescriptive (e.g. the recommendation and
optimization of strategies). The shift from descriptive to prescriptive analytics is a
recent trend in sports analytics [41].
Lindsey [24] used statistical methods to model the progression of the score of

14
a baseball game based on a data set of roughly two thousand professional baseball
matches from 1958 and 1959. By analyzing the score difference data, Lindsey [24]
presented estimates for the teams’ win probabilities based on the score of previous
innings. This analysis aims to support baseball managers’ decision-making during a
match. By having estimates of their win probability, a winning team could rest their
star players to avoid injuries when the comeback probability of the opposing team is
at a satisfactorily low level [24].
In a more recent book on baseball analytics, Tango et al. [53] employed a Markov
chain approach to analyze the game. Markov chain modeling assumes that the
future state of a stochastic process only depends on its current state and is otherwise
independent of its history [32]. Tango et al. [53] described baseball game states as
permutations of the half-inning (unit of play), score difference, occupied bases, and
the number of outs. The state-to-state transition probabilities can then be estimated
from historical data to obtain a discrete-time Markov chain describing the game of
baseball. Tango et al. [53] used this stochastic model to compute the win probability
of the home team in each game state. The resulting win expectancy matrix enables the
estimation of win values for each basic event in baseball (single, out, home run, etc.)
[53]. These win values describe the average change in win probability when an event
takes place, quantifying the value of each event in terms of win probability. Tango
et al. [53] used these win probability estimates to analyze coaching decisions, debunk
common myths, and generate strategic guidelines for baseball.
A similar Markov chain approach was taken by McFarlane [28] to evaluate end-of-
game decision-making in the National Basketball Association (NBA). McFarlane [28]
used logistic regression to estimate win probabilities during the last three minutes of
a basketball game. These estimates were used to construct an end-of-game tactics
metric (ETM) for comparing tactical decisions such as choosing between a two-point
and a three-point field goal attempt. ETM is defined as the difference between the win
probability of the actual decision and that of the theoretically optimal decision for the
game state [28]. McFarlane [28] evaluated the end-of-game decisions made by NBA
teams during the 2015–2016 season and showed that the mean ETM difference of the
teams correlated well with their actual winning percentage in close games.
Win probability estimates have been used to evaluate tactical decisions in other
sports as well. Lock and Nettleton [26] used random forests to estimate the play-by-play
win probability in National Football League (NFL) games. Their random forest model
included variables describing the past performance of the competing teams, allowing
accurate win probability estimation even in the case of unevenly matched teams
[26]. Lock and Nettleton [26] provided examples in American football where the
estimated change in win probability could support the decision-making process of a
coach. This focus on changes in win probability is similar to the win value analysis in
baseball by Tango et al. [53]. Lock and Nettleton [26] experimented with multiple
adjustments to their win probability model, including an attempt to account for the
effects of momentum in American football. The momentum adjustment resulted in
added complexity, but no improvement in performance [26], in line with the findings
of other NFL studies [13, 19]. In addition to their in-depth coverage of win probability
estimation in American football, Lock and Nettleton [26] proposed a general binning

15
method for measuring and visualizing the quality of win probability estimates. While
not stated in the original paper, this binning method is a special case of reliability
diagrams [31], which are used to evaluate the calibration of machine learning models
[16]. Moreover, Lock and Nettleton [26] also describe the assessment of variable
importance in a win probability model.

Table 2: Evaluation methods for win probability models.

Method Description

Accuracy [16] Proportion of correctly predicted outcomes.

ECE [16, 40] Average difference between estimated win
probability and true proportion of wins.
Reliability diagrams [31, 40] Visualization of estimated win probability
versus true proportion of wins.

In their research on in-game win probability estimation in soccer, Robberechts

et al. [40] developed a Bayesian win probability model with an innate measure of
prediction uncertainty. The frequent occurrence of ties and the low-scoring nature of
soccer pose modeling challenges that are not present in the sports discussed above [40].
Robberechts et al. [40] overcame these challenges by modeling the future goal-scoring
probabilities of both teams as independent Poisson distributions. These goal-scoring
estimates are then mapped to corresponding win-draw-loss probabilities. The win
probability model included both pre-game features and in-game features. These
features are dynamically weighted, so that the in-game features rise in importance as
the game progresses. This approach allowed the model to produce more accurate win
probability estimates during the closing moments of a game [40].
Robberechts et al. [40] used reliability diagrams and expected calibration error
(ECE) [16] to evaluate the calibration of their win probability model. Table 2
summarizes these methods of evaluating win probability models. Robberechts et al.
[40] also analyzed feature importance in their model, following the footsteps of Lock
and Nettleton [26]. Robberechts et al. [40] defined an added goal value (AGV) metric
for evaluating soccer players based on their average contribution to their team’s win
probability. AGV relies on estimated changes in win probability, similar to win values
[52] and ETM [28]. These win probability metrics are summarized in Table 3.
The analytical work mentioned in this section is merely an overview of the vast
body of research on win probability estimation. Novel win probability models are being
developed for previously relatively unexplored sports, such as cricket [2]. Based on
the research described in this section, one can conclude that win probability estimates
are a commonly used basis and globally accepted for evaluating players and decisions
in sports. In the next section, win probability estimation is shown to have similar
potential in the realm of esports.

16
 Table 3: Win probability metrics used in traditional sports.

Metric Description Use

Win values [53] Average change in win proba- Evaluating the impact of basic
bility from game events. events in baseball.
ETM [28] Win probability distance to an Comparing tactical decisions
optimal decision. at the end of a basketball game.
AGV [40] A player’s contribution to the Evaluating player performance
team’s win probability. in soccer.

3.2 Win probability estimation in esports

Despite the popularity of esports, the field of esports analytics is still in its early stages
[18, 42]. In contrast to traditional sports analytics [30], esports analytics does not
face the same data collection challenges due to the inherently digital environment
of esports [42]. Multiplayer online battle arena (MOBA) is among the most popular
esports genres, with LoL being its biggest title [56]. This section, along with the rest of
the thesis, focuses primarily on LoL due to the popularity of the game and the author’s
experience with it. Nevertheless, notable analytical work has been published for other
major esports titles, such as Dota 2 [10, 17, 21] and PlayerUnknown’s Battlegrounds
[14, 25]. Due to major differences between the core game mechanics, it is pertinent to
focus on the literature specific to LoL [12, 22, 27, 43, 60].
Win probability estimation is a relatively novel topic in LoL analytics, with few
research papers dedicated specifically to it [22, 27]. However, extensive work has been
done under the term match outcome prediction, which is typically treated as a binary
classification task [5]. A binary match outcome classifier determines which outcome
(win or loss) is more likely, without necessarily assigning explicit probabilities to the
outcomes [20]. However, most binary classifiers include some numerical measure of
confidence that can be mapped to a probability [20]. Thus, win probability estimation
can be seen as an extension of match outcome prediction.
LoL match outcome prediction can be divided into two interesting subproblems:
pre-game prediction [12, 60] and in-game prediction [22, 43]. Pre-game predictions
are made after the draft phase of a match, using only information available before
the actual gameplay commences. This includes information on the players’ past
performances and their chosen champions. In-game predictions are made using all the
available information up to time 𝑡 ∈ [0, 𝑇), where 0 and 𝑇 denote the start and end of
a match, respectively. Note that LoL matches vary in duration, with typical matches
lasting anywhere from 15 to 40 minutes. End-of-game prediction, i.e., outcome
classification using all information available at time 𝑇 is a comparatively uninteresting
task, equivalent to asking which basketball team won when you know the final score.
Do et al. [12] trained and compared various machine learning models for pre-
game match outcome prediction. Their novel contribution was to use measures of

17
player-champion experience as features in the models. The final experience measures—
player-champion win rate and champion mastery points—were chosen using Pearson’s
correlation test. Do et al. [12] collected and used a data set of 5 000 ranked LoL
matches from various skill levels. They achieved upwards of 72% pre-game prediction
accuracy with multiple machine learning methods, including support vector machines,
k-nearest neighbors, decision trees, and deep neural networks (DNN). Their DNN was
throned the best due to its high validation accuracy (75.1%) and comparatively low
standard error (0.6%) [12]. Significantly, these high pre-game prediction accuracies
were achieved for matches governed by a fair matchmaking system, aimed at giving
equal odds for both teams [38]. These results indicate that the draft phase, where
each player selects a champion to play, is paramount for the outcome of a match.
Furthermore, to maximize their win probability, players should only pick champions
they are well-practiced on [12].
White and Romano [60] approached the task of pre-game prediction using logistic
regression and a data set of 87 743 ranked LoL matches. Their research focused on
the effects of psychological momentum in multi-match play sessions. The logistic
regression model achieved 72.1% accuracy in pre-game prediction. The momentum
effects of the players’ previous matches were slight, increasing the pre-game prediction
accuracy by 0.1–0.3% compared to a baseline model [60].
Silva et al. [43] published one of the first research papers on in-game outcome
prediction for LoL matches. They approached the problem using recurrent neural
networks (RNN) due to their suitability for prediction tasks involving time series data
[43]. The choice of RNNs for match outcome prediction implies an assumption of
momentum effects [13] in LoL. This is in contrast to the previously discussed Markov
Chain modeling [28, 53], which relies on an assumption of independence of the game
states [32]. Silva et al. [43] did not explicitly consider these assumptions in their paper,
citing the sequential nature of the data as the primary reason for using RNNs. They
used a data set consisting of 7 621 professional LoL matches played between 2015 and
2018. This match data is multimodal, including both categorical pre-game information
and numerical in-game time series. The best RNN trained by Silva et al. [43] achieved
in-game prediction accuracies ranging from 63.9% at the start of a match (𝑡 = 5 min)
to 83.5% at the later stages of a match (𝑡 = 25 min).
The models discussed above were evaluated solely based on prediction accuracy.
However, in order to obtain meaningful win probability estimates, a match outcome
prediction model must be well-calibrated [16] in addition to being accurate [9]. Kim
et al. [22] were the first to address the problem of confidence calibration in LoL match
outcome prediction. They trained a DNN for in-game prediction and then calibrated
the model using a novel method designed specifically for this task. The proposed
method, data uncertainty loss, is a loss function [20] that aims to minimize calibration
error by accounting for the inherent uncertainty in the matches [22]. The calibrated
in-game prediction model achieved an accuracy of 73.8% (aggregated over all game
times) and an ECE of 0.57%. This marks a significant improvement over the baseline;
a similar uncalibrated model had an accuracy of 73.0% and an ECE of 4.47%. Using
reliability diagrams and ECE, Kim et al. [22] showed that the data uncertainty method
outperformed other, commonly used calibration methods. Temperature scaling [16]

18
was the second-best calibration method, yielding only slightly worse results than the
data uncertainty method [22]. The details of all the prediction models discussed above
are outlined in Table 4.
Table 4: Summary of LoL match outcome prediction models.

Prediction task Method Matches [#] Accuracy [%]

Pre-game [12] DNN 5 000 75.1

Pre-game [60] Logistic regression 87 743 72.1
In-game [43] RNN 7 621 63.9–83.5
In-game [22] DNN 83 875 73.8

Maymin [27] developed an in-game win probability model to refine existing

LoL gameplay metrics and introduce new ones. These advanced gameplay metrics
were incorporated into a novel player-evaluation framework, which is designed to
aid player improvement [27]. Riot Games provides public access to LoL match data
through an application programming interface (API) [39]. However, this public data
is quite limited in granularity, especially for advanced in-game analytics. Maymin
[27] implemented custom software to extract more granular and comprehensive match
data. The in-game win probability model used logistic regression and was trained on
millions of matches [27]. To prevent highly correlated game states in the training data,
Maymin [27] included only a random minute of data from each match. Maymin [27]
also investigated the correlation between an individual’s performance and their team’s
win probability. To this end, an end-of-game outcome prediction decision tree was
constructed using the advanced gameplay metrics of an individual player. The decision
tree achieved 80% validation accuracy, indicating that the advanced metrics succeed in
measuring player performance [27]. The research paper did not include any measure
of prediction accuracy or calibration error [16] for the in-game win probability model.
Although still limited in quantity, the research on win probability estimation
in esports is promising. Recent papers [9, 22] have highlighted the importance of
confidence calibration in providing useful win probability estimates. This is crucial
because modern deep learning models—while attractive due to the ease of data
collection in esports—are inherently poorly calibrated [16]. In theory, esports games
make it possible to collect perfect information. In practice, however, the quality of
the available data is often constrained by the game APIs. To overcome this issue
and obtain higher-quality data, custom data collection methods can be developed, as
demonstrated by Maymin [27]. Despite these advancements, strategy optimization,
one of the focal points of sports analytics [1], remains relatively unexplored in esports.
Fortunately, a shift to such prescriptive analytics can be expected as the field of esports
research matures [41]. To conclude the background of this thesis, the following section
discusses how esports analytics serves to benefit from the research on game-playing
artificial intelligence.

19
3.3 Artificial intelligence and esports analytics
Games have long been a playground for artificial intelligence (AI) research. Game-
playing AI systems were first developed for classic games like backgammon [54] and
chess [7]. Recently, AI systems have matched and even exceeded the skill of top
human players in video games such as Gran Turismo [61], StarCraft [57], and Dota 2
[4]. These video games pose challenges that mimic the complexity of the real world:
high-dimensional environments, partial observability, and long time horizons [4, 57].
Increasingly complex video games help bridge the gap from the study of abstract
games to useful applications in real-world domains [4].
Google DeepMind’s AlphaGo algorithm [46] achieved notoriety when it defeated
Lee Sedol, a world champion in the game of Go [47]. The first versions of AlphaGo
used a combination of value and policy networks to evaluate and select moves
[47]. The policy networks were initially trained using supervised learning on a data
set of Go matches played by human experts [46]. AlphaGo’s successor, AlphaGo
Zero, completely abandoned the dependence on human knowledge, relying solely
on reinforcement learning through self-play yet vastly surpassing its predecessors
in performance [47]. Among many other technical challenges, Silver et al. [46]
highlight the problem of successive game states being strongly correlated, which
leads to overfitting of the value network if not properly addressed. This problem
was mitigated by training the value network on millions of independent game states,
each sampled from a unique match of self-play [46]. The same problem and solution
appeared five years later in esports analytics [27].
The AlphaGo Zero algorithm was generalized and applied to chess and shogi
(Japanese chess) under the name AlphaZero [45]. AlphaZero achieved superhuman
performance [7] in both games within 24 hours of tabula rasa reinforcement learning
through self-play [45]. A few years later, DeepMind developed a multi-agent rein-
forcement learning algorithm for StarCraft, a notoriously difficult real-time strategy
game with a large, combinatorial action space [57]. The algorithm, named AlphaStar,
reached Grandmaster level in StarCraft II, ranking above 99.8% of competitive human
players [57]. This marked a significant milestone for AI research, as mastering the
complex domain of StarCraft can be seen as a stepping stone towards even more
difficult real-world applications [57].
There have been no notable published attempts at developing a superhuman AI
agent for LoL, perhaps due to the absence of an official interface to the game engine
that would facilitate the training of reinforcement learning models. Dota 2, however,
includes an official scripting API designed for building game-playing programs [55].
This API was used by Berner et al. [4] in developing OpenAI Five, the first AI
system to defeat the world champions at an esports game. Like LoL, Dota 2 is a
five-on-five MOBA game where team coordination is vital for performance [51]. The
OpenAI Five model consisted of five near-identical DNNs, each controlling one of the
five heroes (the Dota 2 equivalent of champions) on the team [4]. These networks
demonstrated collaborative behavior by concentrating the team’s resources in the hands
of its strongest members, a strategy seen in expert human play [4]. The observation
embedding system of OpenAI Five [4] was hand-designed for the nuances of Dota 2,

20
similar to the StarCraft-specific architecture of AlphaStar [57].
As game-playing AI systems conquer more esports, we should expect a rise in
the level of human play, similar to the historical effect of chess engines [7, 45]. The
presence of these AI agents opens up a unique opportunity to learn from near-perfect
players. This opportunity is specific to esports, at least for now, as traditional sports are
physically constrained and require advanced robotics to mimic human play. Moreover,
esports analytics serves to benefit from the vast body of game-playing AI research, as
both fields face similar technical challenges and opportunities [27, 46].

21
4 Methodology
This chapter addresses the following decision-making problem: Given a game state
and a finite set of actions, which action should a player take to maximize their team’s
win probability? The proposed method relies on having access to a large data set of
matches, which is used to train a win probability estimation model and then compute
statistics to evaluate the decision alternatives. The chapter attempts to formalize win
probability added as a contextualized measure of value for strategic decision-making,
using mathematical notation appropriate for contemporary esports.

4.1 Win probability in a zero-sum game

Consider a zero-sum game with two competing teams. Furthermore, assume that the
game has only two possible outcomes (win and loss); there are no draws. Since the
game is zero-sum, the teams have opposite goals [59]. Therefore, an increase in win
probability for one team results in an equal decrease for the other. The ground-truth
win probability cannot be determined for complex games like modern esports [9]; the
win probability must be estimated.
Due to the symmetric nature of zero-sum games, it suffices to only consider the
win probability from the perspective of one of the competing teams. Let 𝑤(𝑥) denote
a win probability estimate at game state 𝑥 ∈ X, where X is the set of all possible states
of the zero-sum game. For every game state 𝑥, there exists a mirrored state 𝑥 ′ ∈ X such
that 𝑤(𝑥 ′) = 1 − 𝑤(𝑥). The game states 𝑥 and 𝑥 ′ represent the same situation from
the perspectives of each team. Let T𝑥 denote the team from whose perspective 𝑥 is
given. Now, let 𝑦 ∈ {0, 1} be a dependent random variable representing the outcome
of a match. From the perspective of team T𝑥 , 𝑦 = 1 denotes a win and 𝑦 = 0 a loss.
Formally, 𝑤 : X → [0, 1] estimates the conditional probability of team T𝑥 winning
the match given a game state 𝑥, i.e.,
𝑤(𝑥) ≈ Pr(𝑦 = 1 | 𝑥). (1)
By symmetry, the estimated win probability of the other team is 1 − 𝑤(𝑥). From now
on, 𝑤(𝑥) is referred to as the win probability at state 𝑥, remembering that the 𝑤(𝑥) is,
firstly, an estimate, and secondly, given from the perspective of team T𝑥 .
The abundance of data makes machine learning models attractive for estimating
win probabilities in esports. The technical implementation of a win probability model
is always game-dependent; Bayesian modeling might be suitable for one game [40]
and deep neural networks for another [22]. Win probability models are typically
evaluated based on their accuracy [9], the proportion of correctly predicted outcomes.
However, in addition to being accurate, a win probability should also be well-calibrated
[16], as discussed in Chapter 3. A well-calibrated model produces unbiased win
probability estimates that accurately reflect the true, unknown win probability (right
side of Equation 1). Calibration can be determined empirically using, e.g., expected
calibration error (ECE) [16] and reliability diagrams [31].
To evaluate the calibration of a win probability model, we need a data set of game
states and corresponding win probability estimates. For the computation of ECE and

22
reliability diagrams, these win probability estimates are distributed into 𝑀 bins. The
number of bins should be chosen so that each bin contains sufficiently many samples;
typically 𝑀 ∈ [5, 20] provides reliable results [9], but this depends on the number of
samples 𝑁. For each bin 𝐵𝑚 , we compute the mean win probability estimate 𝑤(𝐵𝑚 )
and the expected outcome 𝑦(𝐵𝑚 ), i.e., the proportion of wins in the bin. ECE is then
the mean absolute difference between 𝑦(𝐵𝑚 ) and 𝑤(𝐵𝑚 ), weighted by the number of
samples in each bin, i.e., the cardinality |𝐵𝑚 |;
𝑀
∑︁ |𝐵𝑚 |
ECE = |𝑦(𝐵𝑚 ) − 𝑤(𝐵𝑚 )|. (2)
𝑚=1
𝑁

A well-calibrated model has a low ECE; ECE = 0 indicates perfect calibration.

Reliability diagrams are generated by plotting the mean win probability estimate
𝑤(𝐵𝑚 ) against the expected outcome 𝑦(𝐵𝑚 ) for each bin 𝐵𝑚 . Examples of reliability
diagrams can be seen in Section 5.1.

4.2 Problem definition and challenges

Having formalized the concept of win probability, let us now elaborate the decision-
making problem at hand:
Given a game state 𝑥 ∗ ∈ X and a finite set of actions 𝐴 = {𝑎 1 , 𝑎 2 , . . . 𝑎 𝐿 } ⊆
A, which action 𝑎 ∗ ∈ 𝐴 should a player take to maximize their team’s
win probability, i.e., 𝑎 ∗ = arg max𝑎∈𝐴 Pr(𝑦 = 1 | 𝑥 ∗ , 𝑎)?
The state space X and action space A are dependent on the zero-sum game being
analyzed. The state space X can be continuous or discrete, but the action space A is
assumed to be discrete to rule out continuous gameplay actions like movement in LoL.
The size of the finite action set 𝐴 is denoted by 𝐿. Using examples from LoL, the
actions can be e.g. item purchases or events such as dragon kills and turret destructions.
Similarly, X can be a high-dimensional vector space representing information about
the in-game time, champion positions, destroyed towers, etc.
We are primarily interested in studying this problem from a strategic decision-
making perspective. Here, strategic refers to the premeditated or retrospective nature
of the analysis. We are not interested in automating player decision-making during
a game, but in supporting it before and after the game. There are several challenges
associated with the task:
1. Game state representation. In many contemporary esports, the game state is
only partially observable during play [4]. This might be the case even post-game due
to constraints of the game API [27]. Moreover, the state spaces of modern esports are
often high-dimensional and multimodal [4], which makes choosing the representation
X difficult. Simplified game state representations are warranted as long as they contain
sufficient information for reliable win probability estimates.
2. Complexity of actions. The complexity of actions can vary significantly
depending on the specific decision being analyzed. We are particularly interested
in complex actions that are difficult or impossible to model using the state-to-state

23
transitions typical of Markov decision processes [50]. Due to the complicated rules and
long time horizons of modern esports [4], the effects of actions can often be uncertain.
Some actions offer minimal immediate benefit, with their value appearing later at
unpredictable times. For example, Mejai’s Soulstealer, an item in LoL, provides only
minor benefits at the time of purchase. However, due to its Glory effect, it can quickly
transform into one of the most powerful items in the game.
3. Variance and external factors. In multiplayer games, the choices of one player
have a limited impact on the overall win probability of a team. External factors, such as
the performance of teammates and opponents, can obscure the true impact of individual
actions. The effects of individual actions are further diminished and obscured as
the number of players increases. Moreover, some esports (e.g. Hearthstone) even
incorporate randomness as a fundamental game mechanic. Thus, assessing the impact
of individual actions is often difficult.
4. Action selection bias. Esports games make it possible to collect massive data
sets of publicly played matches, providing an opportunity to evaluate actions through
aggregate statistics. However, this data is significantly affected by selection bias, since
the players are using their biased judgment to select which action to take. Some actions
are only taken by players to secure wins when they are already ahead or as a last resort
in desperate situations, inflating or deflating their win rates, respectively. Moreover,
esports communities often share popular strategies, which can result in the overuse
or misuse of some popularly recommended strategies, making the associated actions
seem worse than they truly are.
The next section addresses these challenges by introducing a contextualized method
of evaluating actions.

4.3 Contextualized evaluation of actions

The proposed method of evaluating actions requires a data set of actions with their
contexts extracted from historical match data. We denote this data set of actions with
𝐷 = {𝑑 (1) , 𝑑 (2) , . . . , 𝑑 (𝑁) }, where every data point 𝑑 (𝑖) = (𝑎 (𝑖) , 𝑥 (𝑖) , 𝑧 (𝑖) ) ∈ A × X × X
is a triple consisting of an action 𝑎, an initial state 𝑥, and a final state 𝑧. In the 𝑖-th
data point, the initial state 𝑥 (𝑖) is the game state in which action 𝑎 (𝑖) first took effect.
The final state 𝑧 (𝑖) is the last game state affected by 𝑎 (𝑖) . Depending on the game and
actions in consideration, determining 𝑧 can be difficult or impossible. The effects of
an action can also last until the end of a match. In these cases, 𝑧 should by default be
set to the terminal game state, i.e., the last known state of the match.
Continuing with examples from LoL, let us consider an action space consisting
of all purchasable items in LoL. In this example, the action data set 𝐷 is a list of
item purchases from matches collected through the LoL API [39]. A data point
𝑑 = (𝑎, 𝑥, 𝑧) ∈ 𝐷 describes an item purchase made by a player in one of the collected
matches. 𝑎 is the item purchased, 𝑥 is the game state at the time of purchase, and 𝑧 is
the game state when the item was sold, destroyed, or upgraded. If the item remains in
the player’s inventory until the end of the match, we use the terminal game state as 𝑧.
For every data point 𝑑 (𝑖) = (𝑎 (𝑖) , 𝑥 (𝑖) , 𝑧 (𝑖) ) ∈ 𝐷, we define two quantities:

24
 1. 𝑊 (𝑑 (𝑖) ), the initial win probability in 𝑑 (𝑖) ;
𝑊 (𝑑 (𝑖) ) = 𝑤(𝑥 (𝑖) ). (3)

2. Δ𝑊 (𝑑 (𝑖) ), the win probability added in 𝑑 (𝑖) ;

Δ𝑊 (𝑑 (𝑖) ) = 𝑤(𝑧 (𝑖) ) − 𝑤(𝑥 (𝑖) ). (4)

Here, 𝑤 : X → [0, 1] is a win probability estimate as defined in Section 4.1.

The initial win probability 𝑊 : A × X × X → [0, 1] and the win probability added
Δ𝑊 : A × X × X → [−1, 1] form the basis for our contextualized evaluation method.
𝑊 (𝑑 (𝑖) ) encapsulates the initial state 𝑥 (𝑖) in which action 𝑎 (𝑖) was taken into one win
probability. Δ𝑊 (𝑑 (𝑖) ) measures the change in win probability during the effective
window of 𝑎 (𝑖) . These quantities are illustrated in Figure 4.

1.0

z, effect of a ends
w(z)

∆W (d) = w(z) − w(x)

x, action a takes effect
W (d) = w(x)
w

effective window of a

0
0 T
t

Figure 4: A hypothetical match progression illustrating the initial win probability

𝑊 (𝑑) and the win probability added Δ𝑊 (𝑑). The horizontal axis represents the
in-game time 𝑡 ∈ [0, 𝑇], and the vertical axis the estimated win probability 𝑤.

The initial win probability 𝑊 (𝑑 (𝑖) ) reflects the context of an individual action
𝑎 (𝑖)but it does not contain information about the impact of 𝑎 (𝑖) . Moreover, the win
probability added Δ𝑊 (𝑑 (𝑖) ) is not a reliable estimate for the impact of 𝑎 (𝑖) ; Δ𝑊 (𝑑 (𝑖) )
is affected by variance and external factors, as discussed in Section 4.2. In order to
mitigate these adverse effects and generate meaningful insights, these quantities must
be averaged over many similar data points in 𝐷 = {𝑑 (1) , . . . , 𝑑 (𝑁) }. Thus, we define
two statistics for any action 𝑎 ∈ A and for any set of game states 𝑋 ⊆ X:
1. 𝑊 𝑋 (𝑎), the mean initial win probability of 𝑎 in 𝑋;
1 ∑︁
𝑊 𝑋 (𝑎) = 𝑊 (𝑑 (𝑖) ). (5)
|𝐼 𝑋 (𝑎)|
𝑖∈𝐼 𝑋 (𝑎)

25
 2. Δ𝑊 𝑋 (𝑎), the mean win probability added by 𝑎 in 𝑋;

1 ∑︁
Δ𝑊 𝑋 (𝑎) = Δ𝑊 (𝑑 (𝑖) ). (6)
|𝐼 𝑋 (𝑎)|
𝑖∈𝐼 𝑋 (𝑎)

Here, 𝐼 𝑋 (𝑎) = {𝑖 ∈ {1, . . . , 𝑁 } | 𝑎 (𝑖) = 𝑎 ∧ 𝑥 (𝑖) ∈ 𝑋 } is a set of all the indices 𝑖 for
data points 𝑑 (𝑖) ∈ 𝐷 where the action 𝑎 (𝑖) is 𝑎 and the initial state 𝑥 (𝑖) is an element of
the state set 𝑋. The cardinality |𝐼 𝑋 (𝑎)| is the associated sample size. While the action
space A is assumed to be discrete, the state space X can be continuous, necessitating
the use of the state set 𝑋 to group similar states. For the original decision-making
problem, we can define a state set 𝑋 ∗ consisting of states similar to 𝑥 ∗ . The definition of
𝑋 ∗ naturally depends on the state space representation X of the game in consideration.
The mean initial win probability 𝑊 𝑋 (𝑎) measures the systemic bias in data points
where action 𝑎 is selected in specific game states 𝑥 ∈ 𝑋. 𝑊 𝑋 (𝑎) > 0.5 indicates
that action 𝑎 is selected more often in winning situations. Conversely, 𝑊 𝑋 (𝑎) < 0.5
indicates that 𝑎 is selected more often in losing situations. The mean win probability
added Δ𝑊 𝑋 (𝑎) measures the average impact of taking action 𝑎 in the context of
𝑋. Δ𝑊 𝑋 (𝑎) > 0 indicates that taking action 𝑎 in game states 𝑥 ∈ 𝑋 is beneficial
on average and Δ𝑊 𝑋 (𝑎) < 0 the opposite. According to the law of large numbers,
Δ𝑊 𝑋 (𝑎) approaches the true win probability added by taking action 𝑎 in a game state
𝑥 ∈ 𝑋 as the associated sample size 𝐾 = |𝐼 𝑋 (𝑎)| approaches infinity and 𝑋 approaches
the singleton set {𝑥} i.e.,

lim lim Δ𝑊 𝑋 (𝑎) = Pr(𝑦 = 1 | 𝑥, 𝑎) − Pr(𝑦 = 1 | 𝑥). (7)

𝐾→∞ 𝑋→{𝑥}

Thus, given a sufficiently large sample size 𝐾 and a sufficiently specific (similar
to 𝑥 ∗ ) state set 𝑋 ∗ , we can approximate the original decision-making problem by
replacing the true win probability Pr(𝑦 = 1 | 𝑥 ∗ , 𝑎) with the mean win probability
added Δ𝑊 𝑋 ∗ (𝑎), i.e.,

max Pr(𝑦 = 1 | 𝑥 ∗ , 𝑎) (8)

𝑎∈𝐴
≡ max Pr(𝑦 = 1 | 𝑥 ∗ , 𝑎) − Pr(𝑦 = 1 | 𝑥 ∗ ) (9)
𝑎∈𝐴
≅ max Δ𝑊 𝑋 ∗ (𝑎). (10)
𝑎∈𝐴

Under these assumptions, given a game state 𝑥 ∗ ∈ X and a finite set of actions
𝐴 = {𝑎 1 , 𝑎 2 , . . . 𝑎 𝐿 } ⊆ A, the player should select the action with the highest mean
win probability added in states similar to 𝑥 ∗ , i.e., arg max𝑎∈𝐴 Δ𝑊 𝑋 ∗ (𝑎), to maximize
their team’s win probability.

26
5 Case study
In this chapter, we describe the data set and model used to estimate LoL win
probabilities. The performance of the model is evaluated based on accuracy, ECE, and
reliability diagrams. Then, we apply the contextualized evaluation method described
in Chapter 4 to find the best items in different situations for one champion, Shen.

5.1 In-game win probability model

Data and model description
In Section 3.2, we reviewed two notable papers addressing the task of in-game match
outcome prediction for LoL [22, 43]. We follow the approach of Kim et al. [22] and
train a deep neural network (DNN) to estimate the win probability given a game state.
In this approach, the win probability is estimated based on a single game state; the
previous game states are not considered, aligning with the independence assumption
of Markov chain modeling [26]. Thus, we do not account for any in-game momentum
effects, but previous research has shown that these effects are negligible, at least in
traditional sports [13, 19].
For this case study, we used the LoL API [39] to collect a data set of 400 000
Ranked matches played on Patch 14.15. We use 350 000 matches to train the DNN and
reserve 50 000 matches for testing. Because of the limitations of the LoL API [39], our
match data contains only one game state per minute by default. Additionally, we have
the game state at every important game event (kill, turret, dragon, etc.). The frequency
of these events increases as the game progresses. This results in a non-uniform
time distribution of game states, which is further skewed by LoL’s varying match
duration. The average match duration in our data set is roughly 25 minutes. Instead of
training the DNN using every state in every match, we divide matches into 5-minute
intervals and sample one game state uniformly per interval. In addition to removing
the non-uniformity caused by event timing, the uniform random sampling reduces
overfitting due to the correlation of successive game states [27, 46]. However, this
sampling method does not affect the skewness caused by the varying match duration;
the sample sizes are significantly smaller in the late game, i.e., 30 min and onwards.
The primary focus of this thesis is on the application of win probability estimation
and not its game-specific technical implementation. Thus, the precise DNN architecture
is not described here. Kim et al. [22] describe the implementation and calibration
of a similar in-game win probability model for LoL in detail. Nevertheless, Table 5
presents the most important explanatory variables (features) of the DNN. The model
relies primarily on in-game features, with Rank being the only pre-game feature.
Notably, we do not use items as explanatory variables due to their categorical nature.
The added complexity of 60 categorical variables (six item slots per champion) is
deemed imprudent despite the potential increase in predictive power. The exclusion of
items does not affect their analysis using win probability added, as the action space A
is distinct from the state space X, i.e., the explanatory variables of the model.

27
Table 5: Overview of the most important features of the win probability model.
Player-specific variables are repeated ten times, once for each player.

Variable Description
Position Current 2D map coordinates of each player.
Level Current champion level per player.
Gold Cumulative amount of gold acquired per player.
Damage Dealt Cumulative amount of damage dealt per player.
Damage Taken Cumulative amount of damage taken per player.
Kills Cumulative number of champion kills per player.
Assists Cumulative number of champion assists per player.
Deaths Cumulative number of deaths per player.
Voidgrubs Cumulative number of Voidgrubs killed per team.
Dragons Cumulative number of Dragons killed per team.
Barons Cumulative number of Baron Nashors killed per team.
Turrets Cumulative number of turrets destroyed per team.
Inhibitors Cumulative number of inhibitors destroyed per team.
Time Current in-game time.
Rank Average rank (skill level) of the players.

Model performance
The performance of the model is evaluated on the test set consisting of 50 000 matches
with 5 653 687 game states in total. On this set, the model achieves an aggregate
accuracy of 75.9% with an ECE of 0.90%. The ECE is computed with 𝑀 = 20
bins (see Equation 2). Table 6 presents these performance metrics alongside those
of similar in-game prediction models from the literature. The implemented model
surpasses the previous models in accuracy but has a slightly worse ECE than the
calibrated DNN by Kim et al. [22].

Table 6: In-game prediction model performance comparison.

Model Matches [#] Accuracy [%] ECE [%]

RNN [43] 7 621 63.9–83.5 -

Baseline DNN [22] 83 875 73.0 4.47
Calibrated DNN [22] 83 875 73.8 0.57
Presented DNN [This thesis] 350 000 75.9 0.90

28
 In order to investigate the effect of in-game time on the accuracy of the model, we
split the test states into 4-minute intervals and compute the accuracy in each interval
until 40 minutes, visualized in Figure 5. The accuracy starts at 55.8% and increases
with time, peaking at 84.8% in the [24, 28]-minute interval. The accuracy decreases
in the last three time intervals, likely due to the model having less training data in these
match stages. Alternatively, the dynamics of LoL might cause the game to become
less predictable in the late game; the Elder Dragon combined with long death timers
create comeback opportunities for the losing teams, possibly increasing the match
outcome variance.

1.0
Accuracy
Samples
0.9
Accuracy

0.8

0.7

0.6

0.5
[0, 4] [4, 8] [8, 12] [12, 16] [16, 20] [20, 24] [24, 28] [28, 32] [32, 36] [36, 40]
t (min)

Figure 5: A plot of the accuracy of the model in 4-minute in-game time intervals.
The histogram representing the number of samples in each time interval is normalized
so that the bar lengths add up to 1.

Figure 6 presents four reliability diagrams illustrating the calibration of the win
probability model. The test data is now divided into four time intervals, spanning
10 minutes each. The reliability diagrams are generated by plotting the mean win
probability estimate 𝑤 against the expected outcome 𝑦, i.e., the proportion of wins, for
each win probability bin. We use 𝑀 = 20 bins to align with the ECE computation.
The calibration declines slightly with time, but the model is overall well-calibrated.
The reliability diagrams indicate that the model provides reliable win probability
estimates, especially in the early game (𝑡 ∈ [0, 10] min), where the win probability
distribution is light-tailed. The histograms in Figure 6 show how the distribution of
win probabilities changes with time. The distributions are similar to those observed by
Choi et al. [9], who showed that LoL win probabilities can be modeled as symmetric
beta distributions with time-dependent parameters. As the game progresses, the tails
of the win probability distribution become heavier and the win probability estimates
become less reliable. This finding coincides with the previously postulated effects of
late-game outcome variance.

29
 Perfect Calibration Model Calibration Samples

t ∈ [0, 10] min, ECE = 0.0067 t ∈ [10, 20] min, ECE = 0.0119
1.0 1.0

0.8 0.8

y 0.6 y 0.6

0.4 0.4

0.2 0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
w w
t ∈ [20, 30] min, ECE = 0.0159 t ∈ [30, 40] min, ECE = 0.0149
1.0 1.0

0.8 0.8

y 0.6 y 0.6

0.4 0.4

0.2 0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
w w

Figure 6: Four reliability diagrams illustrating the calibration of the model. The titles
indicate the in-game time interval and the associated ECE value (𝑀 = 20).

5.2 Contextualized evaluation of items

Item system and problem definition
As mentioned in Chapter 2, items are a core gameplay system in LoL. Items can
be purchased using gold, increasing champion statistics and granting unique effects.
There are roughly two hundred purchasable items in LoL, divided into three main
categories: basic items, epic items, and legendary items. These categories are sorted in
increasing order of power and cost. Basic items are cheap and relatively weak, but they
can be upgraded into epic items, which can in turn be upgraded into legendary items.
Upgrading is the process of transforming items into more powerful ones using gold;
basic and epic items are components, i.e., intermediate parts, of legendary items. Most
legendary items have two to three required components. However, a player’s inventory
can only hold six items at a time, which makes it difficult to buy the components
for multiple legendary items simultaneously. Thus, players typically choose which
legendary item they want to obtain next and then purchase the components leading up
to it. Due to these reasons, we are primarily interested in analyzing the 107 legendary
items currently in the game.
Items generally provide the same benefits, no matter the champion in play. However,
every champion interacts with these benefits in different ways; an item effect can be
useless for one champion and extremely valuable for another. Thus, it is reasonable to
evaluate items in the context of a specific champion. The teammates and opponents
also affect the value of items. In the early stages of a game, it is common to build an
item to counter your lane opponent, the enemy champion that you are fighting most
often. In the following analysis, we focus on items specifically for Shen, a champion

30
primarily played in the top lane. This choice is motivated by the author’s extensive
experience playing Shen, allowing for a more thorough qualitative analysis of the
results.
The effects of items are not independent; some items have synergistic effects with
other items, while purchasing multiple items with similar effects leads to diminishing
returns. When a player is choosing their next item, they thus have to consider the items
already in their inventory. From a combinatorial perspective, while there are only 107
legendary items, the number of six-item permutations is 107!/101! ≈ 1.3 · 1012 . To
avoid this complexity, we will only consider the choice of the first legendary item in the
following analysis. Here is the final problem description with all its simplifications:

Given the opposing top lane champion, which legendary item should Shen
purchase first to maximize his team’s win probability?

Evaluation of Shen’s first legendary items

Let us begin by evaluating the first legendary items for Shen without the added
complexity of a lane opponent. We will consider the four most popular legendary
items for Shen: Sunfire Aegis, Titanic Hydra, Hollow Radiance, and Heartsteel. Using
the in-game win probability model, we compute the mean initial win probabilities 𝑊
and mean win probabilities added Δ𝑊 for these items. Here, 𝑊 and Δ𝑊 are shorthand
notation for 𝑊 𝑋 (𝑎) and Δ𝑊 𝑋 (𝑎), where 𝑎 ∈ 𝐴 = {Sunfire Aegis, Titanic Hydra,
Hollow Radiance, Heartsteel} and 𝑋 = {all game states where Shen purchases his first
legendary item}. Additionally, the win rate 𝑦 and the sample size 𝐾 = |𝐼 𝑋 (𝑎)| are
provided for context. These statistics are computed using a data set with 8 206 item
purchases matching our criteria. Table 7 presents the results of the initial evaluation
with no lane opponent information.

Table 7: Statistics of the most popular items for Shen in the top lane. The items are
sorted by the mean win probability added Δ𝑊. The mean initial win probability 𝑊,
the win rate 𝑦, and the sample size 𝐾 are provided for context.

Item 𝚫𝑾 [%] 𝑾 [%] 𝒚 [%] 𝑲 [#]

Titanic Hydra 1.46 51.41 51.94 2189

Heartsteel 0.66 53.66 54.15 1411
Sunfire Aegis -0.51 50.53 49.10 3004
Hollow Radiance -0.84 52.76 51.25 1602

Let us first consider the average impact of each item by looking at the mean win
probability added Δ𝑊. Titanic Hydra seems significantly better than the other items,
with a Δ𝑊 of 1.46%. Heartsteel comes in second with a Δ𝑊 of 0.66%. The mean
initial win probability 𝑊 of Heartsteel (53.66%) is significantly higher than that of

31
Titanic Hydra (51.41%), indicating that Heartsteel is purchased in more winning
situations than Titanic Hydra. This is to be expected as Heartsteel’s effect, Colossal
Consumption, is better the earlier it is purchased, making it an attractive item when
ahead. Greedily considering only the win rate 𝑦 of an item could lead a player to
believe that Heartsteel is the best alternative because it wins the most amount of
matches (54.15%). However, the win rate is biased; it fails to account for the context
in which Heartsteel is selected. The remaining two items, Sunfire Aegis and Hollow
Radiance, have negative Δ𝑊 values, -0.51% and -0.84%, respectively. It is possible
to argue for purchasing Sunfire Aegis in specific situations because it has the lowest
𝑊 and highest sample size 𝐾; it is the default option for Shen players and is thus
incorrectly being bought in every scenario, possibly deflating its Δ𝑊. On the other
hand, Hollow Radiance has a higher 𝑊 (52.76%) and a much lower Δ𝑊 than Titanic
Hydra. Thus, Hollow Radiance is dominated by Titanic Hydra.
Let us now increase the specificity by narrowing the state set 𝑋 based on the lane
opponent of Shen. Top lane champions can be divided into two categories based on
their primary damage type, which is either physical damage or magic damage. Two
state sets are defined accordingly: 𝑋phys and 𝑋magic . The total sample sizes for these
sets are |𝐼 𝑋phys | = 5994 and |𝐼 𝑋magic | = 2212; physical damage champions are more
common in the top lane. Tables 8 and 9 present the results of the evaluation based on
the opposing laner’s damage type.

Table 8: Statistics of the most popular items for Shen vs. physical damage top laners.
The bolded values differ noticeably from the general values of Table 7.

Item 𝚫𝑾 [%] 𝑾 [%] 𝒚 [%] 𝑲 [#]

Titanic Hydra 1.94 51.49 52.36 1696

Heartsteel 0.23 53.49 53.39 1107
Sunfire Aegis -0.68 50.34 48.69 2791
Hollow Radiance -0.71 53.01 51.25 400

Table 9: Statistics of the most popular items for Shen vs. magic damage top laners.

Item 𝚫𝑾 [%] 𝑾 [%] 𝒚 [%] 𝑲 [#]

Heartsteel 2.24 54.26 56.91 304

Sunfire Aegis 1.75 53.04 54.46 213
Titanic Hydra -0.21 51.16 50.51 493
Hollow Radiance -0.88 52.67 51.25 1202

32
 The item statistics versus physical damage are almost identical to the general item
statistics in Table 7. This is not unexpected, as 𝑋phys comprises 73% of all the data.
Nevertheless, three differences stand out in Table 8: the mean win probability added
Δ𝑊 of Titanic Hydra (1.46% → 1.94%) and Heartsteel (0.66% → 0.23%), and the
sample size 𝐾 of Hollow Radiance (1602 → 400). Based on Δ𝑊, Titanic Hydra is the
predominant legendary item for Shen versus physical damage champions. This result
is not surprising to most LoL players; Titanic Hydra covers Shen’s main weakness
(wave clear) and synergizes with Shen’s Q Ability, Twilight Assault. The margin by
which Titanic Hydra surpasses the alternatives is nonetheless noteworthy.
The relative sample size of Hollow Radiance decreased from 19.5% to 6.7% as we
specified the opponent to be a physical damage champion. This change is credited
to the magic resistance provided by Hollow Radiance, which reduces magic damage
taken by the purchaser. Magic resistance does not affect physical damage, making it
ineffective versus physical damage champions. For comparison, the relative sample
size of Hollow Radiance is 54.5% in Table 9, indicating its popularity versus magic
damage top laners. Despite its popularity, Hollow Radiance, with a low Δ𝑊 of -0.88%,
is outperformed by the alternatives. Moreover, Sunfire Aegis, an item that provides
physical resistance, is evidently better versus magic damage than physical damage.
This result is surprising and cannot be fully explained without more data. The high 𝑊
(53.04) and the low sample size (213) of Sunfire Aegis indicate that the item is only
purchased versus magic damage when Shen’s team is winning. In these scenarios, Shen
can itemize against the opposing team—most likely consisting of physical damage
champions—instead of his lane opponent, making Sunfire Aegis a viable alternative.
In Table 9, the order of Heartsteel and Titanic Hydra is reversed compared to
the previous tables; Heartsteel has the highest Δ𝑊 value yet (2.24%) while Titanic
Hydra falls into the negatives for the first time (-0.21%). Heartsteel is still systemically
bought when Shen’s team has a lead (𝑊 = 54.26%), but the high Δ𝑊 indicates
that Heartsteel succeeds in furthering that lead versus magic damage top laners.
Nevertheless, the results of Table 9 are tentative, and more data should be gathered to
provide recommendations for Shen’s itemization versus magic damage champions.

5.3 Summary of the results

To summarize the results of this LoL case study, our in-game win probability model
achieved 75.9% accuracy and 0.90% ECE on unseen test data, rivaling the performance
of similar models found in the literature. Applying the contextualized evaluation
method presented in Chapter 4, we studied the choice of the first legendary item for
Shen. Based on the mean win probability added, Titanic Hydra was deemed the best
item for Shen against physical damage lane opponents. Surprisingly, the most popular
item against magic damage lane opponents, Hollow Radiance, was discovered to also
be the worst alternative.

33
6 Conclusions
At the start of this thesis, we set out to explore the use of win probability estimation in
esports with the objective of applying it to strategy optimization. This focal point of
sports analytics remains relatively unexplored for esports, as discussed in Chapter 3.
In Chapter 4, we formalized win probability as a metric for strategic decision-making,
using mathematical notation appropriate for contemporary esports. We defined two
statistics for the contextualized evaluation of decision alternatives: the mean initial
win probability and the mean win probability added. These statistics were applied
to real esports data in Chapter 5, where we studied the choice of the first legendary
item for Shen, a champion in LoL, using a custom in-game win probability model that
rivaled the performance of similar models from the literature. This case study, while
niche, showed that win probability estimation can be used to generate novel insights
that can hopefully lead to better strategic decision-making in esports.
The abundance of data makes esports an attractive field for deep learning. In
addition to esports analytics, deep learning models have been used in game-playing
systems, such as OpenAI Five [4]. The OpenAI Five model famously beat the Dota
2 world champions in 2019, achieving superhuman performance [7] in a complex
five-on-five game. Interestingly, the OpenAI Five team completely ignored the
itemization problem studied in this thesis; the AI agents were designed to choose the
most popular items by default [4]. However, this does not mean that the choice of
items is irrelevant. Rather, it highlights the different objectives of esports analytics
and game-playing AI research; esports analytics focuses on evaluating and improving
human performance while game-playing AI systems are designed to solve problems
that have been historically difficult for machines. Nevertheless, the two fields share
similar technical challenges and will continue to grow intertwined.
Recently, the emphasis in sports analytics has shifted from descriptive (e.g.
performance evaluation) to prescriptive analytics (e.g. strategy optimization) [41].
This thesis is a step toward prescriptive analytics in esports, hopefully inspiring
others to analyze strategic decisions in their favorite games. The proposed win
probability statistics provide more insights than their biased predecessors, e.g., win
rate. Nevertheless, they suffer from a fundamental trade-off between specificity and
sample size. Future research could study the optimal choice of the state set 𝑋, balancing
the specificity and reliability of the statistics. Despite their value, athletes and coaches
should not follow statistics blindly; domain knowledge and intuition are required to
generate meaningful insights and make the best decisions.

34
References
[1] B. Alamar. Sports Analytics: A Guide for Coaches, Managers, and Other
Decision Makers. Columbia University Press, 2013.
[2] M. Asif and I. G. McHale. “In-play forecasting of win probability in one-day
international cricket: A dynamic logistic regression model”. International
Journal of Forecasting 32.1 (2016), pp. 34–43.
[3] M. Bar-Eli, O. H. Azar, I. Ritov, Y. Keidar-Levin, and G. Schein. “Action
bias among elite soccer goalkeepers: The case of penalty kicks”. Journal of
Economic Psychology 28.5 (2007), pp. 606–621.
[4] C. Berner et al. “Dota 2 with large scale deep reinforcement learning”. arXiv
preprint arXiv:1912.06680 (2019).
[5] K. U. Birant and D. Birant. “Multi-objective multi-instance learning: A new
approach to machine learning for esports”. Entropy 25.1 (2023), article 28.
[6] S. E. Buttrey, A. R. Washburn, and W. L. Price. “Estimating NHL scoring
rates”. Journal of Quantitative Analysis in Sports 7.3 (2011), pp. 1–18.
[7] M. Campbell, A. J. Hoane, and F.-h. Hsu. “Deep Blue”. Artificial Intelligence
134.1 (2002), pp. 57–83.
[8] P.-A. Chiappori, S. Levitt, and T. Groseclose. “Testing mixed-strategy equilibria
when players are heterogeneous: The case of penalty kicks in soccer”. American
Economic Review 92.4 (2002), pp. 1138–1151.
[9] E. Choi, J. Kim, and W. Lee. “Rethinking evaluation metric for probability
estimation models using esports data”. arXiv preprint arXiv:2309.06248 (2023).
[10] N. Clark, B. Macdonald, and I. Kloo. “A Bayesian adjusted plus-minus analysis
for the esport Dota 2”. Journal of Quantitative Analysis in Sports 16.4 (2020),
pp. 325–341.
[11] T. Decroos, V. Dzyuba, J. V. Haaren, and J. Davis. “Predicting soccer high-
lights from spatio-semporal match event streams”. Proceedings of the AAAI
Conference on Artificial Intelligence. Vol. 31. AAAI, 2017, pp. 1302–1308.
[12] T. D. Do, S. I. Wang, D. S. Yu, M. G. McMillian, and R. P. McMahan.
“Using machine learning to predict game outcomes based on player-champion
experience in League of Legends”. Proceedings of the 16th International
Conference on the Foundations of Digital Games. ACM, 2021, pp. 1–5.
[13] M. J. Fry and F. A. Shukairy. “Searching for momentum in the NFL”. Journal
of Quantitative Analysis in Sports 8.1 (2012), pp. 1–20.
[14] N. F. Ghazali, N. Sanat, and M. A. As’ari. “Esports analytics on PlayerUn-
known’s Battlegrounds player placement prediction using machine learning
approach”. International Journal of Human and Technology Interaction 5.1
(2021), pp. 17–28.

35
[15] T. Gilovich, R. Vallone, and A. Tversky. “The hot hand in basketball: On
the misperception of random sequences”. Cognitive Psychology 17.3 (1985),
pp. 295–314.
[16] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. “On calibration of modern
neural networks”. Proceedings of the 34th International Conference on Machine
Learning. Vol. 70. PMLR, 2017, pp. 1321–1330.
[17] V. J. Hodge, S. Devlin, N. Sephton, F. Block, P. I. Cowling, and A. Drachen.
“Win prediction in multiplayer esports: Live professional match prediction”.
IEEE Transactions on Games 13.4 (2021), pp. 368–379.
[18] D. Jeong and S. Youk. “Refining esports: A quantitative cartography of esports
literature”. Entertainment Computing 47 (2023), article 100597.
[19] A. W. Johnson, A. J. Stimpson, and T. K. Clark. “Turning the tide: Big plays
and psychological momentum in the NFL”. Proceedings of the 6th Annual MIT
Sloan Sports Analytics Conference. MIT Sloan, 2012.
[20] A. Jung. Machine Learning: The Basics. Springer Nature, 2022.
[21] C. H. Ke et al. “DOTA 2 match prediction through deep learning team fight
models”. 2022 IEEE Conference on Games. IEEE, 2022, pp. 96–103.
[22] D.-H. Kim, C. Lee, and K.-S. Chung. “A confidence-calibrated MOBA game
winner predictor”. 2020 IEEE Conference on Games. IEEE, 2020, pp. 622–625.
[23] League of Legends Wiki Community. League of Legends Wiki. 2024. u r l:
https://leagueoflegends.fandom.com/wiki/League _ of _ Legends _ Wiki (vis-
ited on 07/23/2024).
[24] G. R. Lindsey. “The progress of the score during a baseball game”. Journal of
the American Statistical Association 56.295 (1961), pp. 703–728.
[25] J. Liu, J. Huang, R. Chen, T. Liu, and L. Zhou. “A two-stage real-time
prediction method for multiplayer shooting e-sports”. Proceedings of the 20th
International Conference on Electronic Business. ICEB, 2020, pp. 9–18.
[26] D. Lock and D. Nettleton. “Using random forests to estimate win probability
before each play of an NFL game”. Journal of Quantitative Analysis in Sports
10.2 (2014), pp. 197–205.
[27] P. Z. Maymin. “Smart kills and worthless deaths: esports analytics for League
of Legends”. Journal of Quantitative Analysis in Sports 17.1 (2021), pp. 11–27.
[28] P. McFarlane. “Evaluating NBA end-of-game decision-making”. Journal of
Sports Analytics 5.1 (2019), pp. 17–22.
[29] T. W. Miller. Sports Analytics and Data Science: Winning the Game with
Methods and Models. FT Press, 2015.
[30] E. Morgulev, O. H. Azar, and R. Lidor. “Sports analytics and the big-data era”.
International Journal of Data Science and Analytics 5.4 (2018), pp. 213–222.

36
[31] A. Niculescu-Mizil and R. Caruana. “Predicting good probabilities with
supervised learning”. Proceedings of the 22nd international conference on
Machine learning. ACM, 2005, pp. 625–632.
[32] J. R. Norris. Markov Chains. Cambridge University Press, 1998.
[33] K. Pelechrinis. “iWinRNFL: A simple, interpretable & well-calibrated in-game
win probability model for NFL”. arXiv preprint arXiv:1704.00197 (2018).
[34] S. Pettigrew. “Assessing the offensive productivity of NHL players using
in-game win probabilities”. Proceedings of the 9th Annual MIT Sloan Sports
Analytics Conference. Vol. 2. MIT Sloan, 2015, article 8.
[35] J. Poropudas and T. Halme. “Dean Oliver’s four factors revisited”. arXiv
preprint arXiv:2305.13032 (2023).
[36] C. Reep and B. Benjamin. “Skill and chance in association football”. Journal
of the Royal Statistical Society, Series A 131.4 (1968), pp. 581–585.
[37] Riot Games. Dev Diary: Win Probability Powered by AWS at Worlds. 2023.
u r l: https://lolesports.com/en- GB/news/dev- diary- win- probability-
powered-by-aws-at-worlds (visited on 06/29/2024).

[38] Riot Games. League of Legends Support: Matchmaking and Autofill. 2023. url:
https : / / support - leagueoflegends . riotgames . com / hc / en - us / articles /
201752954-Matchmaking-and-Autofill (visited on 06/27/2024).
[39] Riot Games. Riot Developer Portal: League of Legends API. 2024. u r l:
https://developer.riotgames.com/docs/lol (visited on 06/27/2024).

[40] P. Robberechts, J. Van Haaren, and J. Davis. “A Bayesian approach to in-game

win probability in soccer”. Proceedings of the 27th ACM SIGKDD Conference
on Knowledge Discovery & Data Mining. ACM, 2021, pp. 3512–3521.
[41] V. Sarlis and C. Tjortjis. “Sports analytics — Evaluation of basketball players
and team performance”. Information Systems 93 (2020), article 101562.
[42] M. Schubert, A. Drachen, and T. Mahlmann. “Esports analytics through
encounter detection”. Proceedings of the 10th Annual MIT Sloan Sports
Analytics Conference. MIT Sloan, 2016.
[43] A. L. C. Silva, G. L. Pappa, and L. Chaimowicz. “Continuous outcome
prediction of League of Legends competitive matches using recurrent neural
networks”. Proceedings of SBGames 2018. SBC, 2018.
[44] R. M. Silva. “Sports analytics”. Ph.D. dissertation. Simon Fraser University,
2016.
[45] D. Silver et al. “Mastering chess and shogi by self-play with a general
reinforcement learning algorithm”. arXiv preprint arXiv:1712.01815 (2017).
[46] D. Silver et al. “Mastering the game of Go with deep neural networks and tree
search”. Nature 529.7587 (2016), pp. 484–489.
[47] D. Silver et al. “Mastering the game of Go without human knowledge”. Nature
550.7676 (2017), pp. 354–359.

37
[48] H. Stekler, D. Sendor, and R. Verlander. “Issues in sports forecasting”.
International Journal of Forecasting 26.3 (2010), pp. 606–621.
[49] H. S. Stern. “A Brownian motion model for the progress of sports scores”.
Journal of the American Statistical Association 89.427 (1994), pp. 1128–1134.
[50] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT
Press, 1998.
[51] E. T. S. Tan, K. Rogers, L. E. Nacke, A. Drachen, and A. Wade. “Communication
sequences indicate team cohesion: A mixed-methods study of ad hoc League
of Legends teams”. Proceedings of the ACM on Human-Computer Interaction.
Vol. 6. ACM, 2022, article 225.
[52] D. Tang, R. K.-w. Sum, M. Li, R. Ma, P. Chung, and R. W.-k. Ho. “What is
esports? A systematic scoping review and concept analysis of esports”. Heliyon
9.12 (2023), article e23248.
[53] T. M. Tango, M. G. Lichtman, and A. E. Dolphin. The Book: Playing the
Percentages in Baseball. Potomac Books, 2007.
[54] G. Tesauro. “TD-Gammon, a self-teaching backgammon program, achieves
master-level play”. Neural Computation 6.2 (1994), pp. 215–219.
[55] Valve Developer Community. Dota Bot Scripting. 2023. u r l: https :
/ / developer . valvesoftware . com / wiki / Dota _ Bot _ Scripting (visited on
06/29/2024).
[56] J. A. C. Vera, J. M. A. Terrón, and S. G. García. “Following the trail of esports:
The multidisciplinary boom of research on the competitive practice of video
games”. International Journal of Gaming and Computer-Mediated Simulations
10.4 (2018), pp. 42–61.
[57] O. Vinyals et al. “Grandmaster level in StarCraft II using multi-agent reinforce-
ment learning”. Nature 575.7782 (2019), pp. 350–354.
[58] M. Walker and J. Wooders. “Minimax play at Wimbledon”. American Economic
Review 91.5 (2001), pp. 1521–1538.
[59] A. Washburn. Two-Person Zero-Sum Games. Springer US, 2014.
[60] A. White and D. M. Romano. “Scalable psychological momentum forecasting
in esports”. arXiv preprint arXiv:2001.11274 (2020).
[61] P. R. Wurman et al. “Outracing champion Gran Turismo drivers with deep
reinforcement learning”. Nature 602.7896 (2022), pp. 223–228.

38