Who Will Win It?

An In-game Win Probability Model for Football

Pieter Robberechts1 , Jan Van Haaren2 , and Jesse Davis1

1
Dept of Computer Science, KU Leuven, Belgium
{first.lastname}@cs.kuleuven.be
2
SciSports, Amersfoort, The Netherlands
[email protected]

Abstract. In-game win probability is a statistical metric that provides

a sports team’s likelihood of winning at any given point in a game, based
on the performance of historical teams in the same situation. In baseball,
basketball and American football, these models serve as a tool to enhance
the fan experience, evaluate in-game decision making and measure the
risk-reward balance for coaching decisions. In contrast, they have received
less attention in association football, because its low-scoring nature makes
it far more challenging to analyze. In this paper, we build an in-game
win probability model for football. Specifically, we first show that port-
ing existing approaches from other sports does not yield good in-game
win probability estimates. Second, we introduce our Bayesian statistical
model that utilizes a set of eight variables to predict the running win,
tie and loss probabilities for the home team. We train our model using
event data from the last four seasons of the major European football com-
petitions. Our results indicate that our model provides well-calibrated
probabilities. Finally, we elaborate on two use cases for our win proba-
bility metric: enhancing the fan experience and evaluating performance
in crucial situations.

1 Introduction
In-game win probability models provide the likelihood that a particular team will
win a game based upon a specific game state (i.e., score, time remaining, . . . ).
Those models have become increasingly popular in a variety of sports over the last
decade. Nowadays, in-game win probability is widely used in baseball [12], basket-
ball [1,8] and American football [3,16,13]. It has a number of relevant use-cases
within these sports’ ecosystems. First, the win probability added (WPA) metric
computes the change in win probability between two consecutive game states.
It allows one to rate a player’s contribution to his team’s performance [17,11],
measure the risk-reward balance of coaching decisions [13,15], or evaluate in-
game decision making [14]. Second, win probability graphs can improve the fan
experience by telling the story of a game.3 For example, they can help identify
3
ESPN includes win probability graphs in its match reports for basketball (e.g.,
http://espn.com/nba/game?gameId=401071795) and American football (e.g., http:
//espn.com/nfl/game?gameId=401030972)
2 P. Robberechts et al.

exciting or influential moments in the game [19], which may be useful for broad-
casters looking for game highlights or to evaluate performance in crucial game
situations [18,17]. Third, they are relevant to in-game betting scenarios. Here,
gamblers have the option to continue to bet once an event has started, and adapt
their bets depending on how the event is progressing. This became a popular
betting service in many countries, and is estimated to account for over one-third
of online betting gross gambling yield in Britain [5].
While well established in these American sports, in-game win probability is
a relatively new concept in association football. It first emerged during the 2018
World Cup when both FiveThirtyEight and Google published such predictions.
The lack of attention in win probability in association football can probably be
attributed to its low-scoring nature and high probability of ties, which makes
the construction of a good in-game win probability model significantly harder in
comparison to the aforementioned sports. Unfortunately, FiveThirtyEight and
Google do not provide any details about how they tackled those challenges.
We present a machine learning approach for making minute-by-minute win
probability estimates for association football. By comparing with state-of-the-art
win probability estimation techniques in other sports, we introduce the unique
challenges that come with modelling these probabilities for football. In particular,
it involves challenges such as capturing the current game state, dealing with
stoppage time, the frequent occurrence of ties and changes in momentum. To
address these challenges, we introduce a Bayesian model that models the future
number of goals that each team will score as a temporal stochastic process. We
evaluate our model on event stream data from the four most recent seasons of
the major European football leagues.

2 Task and Challenges

The task of a win probability model is to predict the probability distribution

over the possible match outcomes given the current game state. In football, this
corresponds to predicting the probability of a win, a draw and a loss. While
the same task has been solved for the major American sports, football has some
unique distinguishing properties that impact developing a win probability model.
We identify four such issues.
1. Describing the game state. In-game win probability models are based
on features that capture the estimated pre-game strength difference between the
two teams, as well as features that describe the in-game situation. The former are
well researched in pre-game prediction models [9,7]. The latter are unique to in-
game prediction models. For each sport, they include at least the time remaining
and the score differential, but the remaining features that describe the in-game
situation differ widely between sports. American football models typically use
features such as the current down, distance to the goal line and number of
remaining timeouts [16,13], while basketball models incorporate possession and
lineup encodings [1,8]. Since there are no publicly available models for association
football, it is unclear which features should be used to describe the game state.
 Who Will Win It? 3

2. Dealing with stoppage time. In most sports, one always knows exactly
how much time is left in the game, but this is not the case for football. Football
games rarely last precisely 90 minutes. Each half is 45 minutes long, but the
referee can supplement those allotted periods to compensate for stoppages during
the game. There are general recommendations and best practices that allow fans
to project broadly the amount of time added at the end of a half, but no one can
ever be quite certain.
3. The frequent occurrence of ties. Another unique property of football
is the frequent occurrence of ties. Football is a low-scoring game, hence the two
teams are often separated by less than two goals. In this setting the win-draw-loss
outcome provides essentially zero information. At each moment in time, a win
or loss could be converted to a tie, and a tie could be converted to a win or a
loss for one of both teams.
4. Changes in momentum. Additionally, the fact that goals are scarce
in football (typically less than three goals per game) means that when they do
occur, they often change the subsequent ebb and flow of the game in terms of
how space opens up and who dominates the ball – and where they do it. The
existing win probability models are very unresponsive to such shifts in the tenor
of a game.

3 A Win Probability Model for Football

In this section, we outline our approach for constructing a win probability model
for association football. First, we discuss how to describe the game state. Second,
we introduce our four win probability models. The first three models are inspired
by the existing models in other sports. The fourth model is our novel Bayesian
approach, which addresses the challenges described in the previous section.

3.1 Describing the game state

To deal with the variable duration of games due to stoppage time, we split each
game into T = 100 time frames, each corresponding to a percentage of the game.
Halftime always corresponds to T = 50. For unfinished games, one can estimate
the time that will be added at the end using a random forest model based on
the current game time, the number of substitutes, the number of yellow and red
cards, the number of goals scored, the time lost due to injuries and whether the
game is tense (goal difference less than or equal to one). Next, we describe the
game state in each of these frames for each team separately using the following
variables:
1. Base features
– Game Time: Percentage of the total game time completed.
– Score Differential: The current score differential.
2. Team strength features
– Rating Differential: The difference in Elo ratings [9] between both teams, which
represents the prior estimated difference in strength with the opponent.
4 P. Robberechts et al.

3. Contextual features
– Team Goals: The number of goals scored so far.
– Yellows: The total number of yellow cards received.
– Reds: The difference with the opposing team in number of red cards received.
– Attacking Passes: A rolling average of the number of successfully completed
attacking passes (a forward pass ending in the final third of the field) during
the previous 10 time frames.
– Duel Strength: A rolling average of the percentage of duels won in the previous
10 time frames.
The challenge here is to design a good set of contextual features. The addition
of each variable increases the size of the state space exponentially and makes
learning a well-calibrated model significantly harder. On the other hand, they
should accurately capture the likelihood of each team to win the game. The five
contextual features that we propose are capable of doing this: the number of goals
scored so far gives an indication of whether a team was able to score in the past
(and is therefore probably capable of doing it again); a difference in red cards
represents a goal-scoring advantage [4]; a weaker team that is forced to defend can
be expected to commit more fouls and incur more yellow cards; the percentage
of successful attacking passes captures a team’s success in creating goal scoring
opportunities; and the percentage of duels won captures how effective teams are
at regaining possession. Besides these five contextual features, we experimented
with a large set of additional features.

3.2 Applying existing win probability models to football

At first, association football seems not very different from basketball and Amer-
ican football. In all these sports, two teams try to score as much as possible
while preventing the other team from scoring. After a fixed amount of time, the
team that scored the most wins. Therefore, it seems straightforward that a model
similar to the ones used in basketball and American football could be applied to
association football too. We consider three such models:
1. Logistic regression model [16,3] (LR). This is a basic multi-class logistic
regression model that calculates the probability of the win, tie and loss
outcomes given the current state of the game:
T
ew xt
P (Y = o|xt ) = , (1)
1 + ewT xt
where Y is the dependent random variable of our model representing whether
the game ends in a win, tie or loss for the home team, xt is the vector with
the game state features, while the coefficient vector w includes the weights
for each independent variable and is estimated using historic match data.
2. Multiple logistic regression classifiers [2,20] (mLR). This model re-
moves the remaining time from the game state vector and trains a separate
logistic classifier per time frame. As such, this model can deal with non-linear
effects of the time remaining on the win probability.
3. Random forest model [13] (RF). Third, a random forest model can deal
with non-linear interactions between all game state variables.
 Who Will Win It? 5

3.3 Our model

Most existing win probability models use a machine learning model that directly
estimates the probability of the home team winning. Instead, we model the
number of future goals that a team will score and then map that back to the
win-draw-loss probability. Specifically, given the game state at time t, we model
the probability distribution over the number of goals each team will score between
time t + 1 and the end of the match. This task can be formalized as:

Given: A game state (xt,home , xt,away ) at time t.

Do: Estimate probabilities
– P (y>t,home = g | xt,home ) that the home team will score g ∈ N
more goals before the end of the game
– P (y>t,away = g | xt,away ) that the away team will score g ∈ N
more goals before the end of the game
such that we can predict the most likely final scoreline (yhome , yaway )
for each time frame in the game as (y<t,home + y>t,home , y<t,away +
y>t,away ).
This formulation has two important advantages. First, the goal difference
contains a lot of information and the distribution over possible goal differences
provides a natural measure of prediction uncertainty [8]. By estimating the
likelihood of each possible path to a win-draw-loss outcome, our model can
capture the uncertainty of the win-draw-loss outcome in close games. Second, by
modelling the number of future goals instead of the total score at the end of the
game, our model can better cope with these changes in momentum that often
happen after scoring a goal.
We model the expected number of goals that the home (y>t,home ) and away
(y>t,away ) team will score after time t, as independent Binomial distributions:
y>t,home ∼ B(T − t, θt,home ),
(2)
y>t,away ∼ B(T − t, θt,away ),
where the θ parameters represent each team’s estimated scoring intensity in the
tth time frame. These scoring intensities are estimated from the current game
state xt,i . However, the importance of these game state features varies over time.
At the start of the game, the prior estimated strengths of each team are most
informative, while near the end the features that reflect the in-game performance
become more important. Moreover, this variation is not linear, for example,
because of a game’s final sprint. Therefore, we model these scoring intensity
parameters as a temporal stochastic process. In contrast to a multiple regression
approach (i.e., a separate model for each time frame), the stochastic process
view allows sharing information and performing coherent inference between time
frames. As such, our model can make accurate predictions for events that occur
rarely (e.g., a red card in the first minute of the game). More formally, we model
the scoring intensities as:
6 P. Robberechts et al.

αt ∼ N (αt−1 , 2)
θt,home = invlogit(αt ∗ xt,home + β + Ha)
β ∼ N (0, 10) (3)
θt,away = invlogit(αt ∗ xt,away + β)
Ha ∼ N (0, 10)
where αt are the regression coefficients, β is the intercept and Ha models the
home advantage.
Our model was trained using PYMC3’s Auto-Differentiation Variational In-
ference (ADVI) algorithm [10]. To deal with the large amounts of data, we also
take advantage of PYMC3’s mini-batch feature for ADVI.

4 Experiments
The goal of our experimental evaluation is to: (1) explore the prediction accuracy
and compare with the various models we introduced in the previous section and
(2) evaluate the importance of each feature.

4.1 Dataset
The time remaining and score differential could be obtained from match reports,
but the contextual features that describe the in-game situation require more
detailed data. Therefore, our analysis relies on Wyscout event stream data from
the English Premier League, Spanish LaLiga, German Bundesliga, Italian Serie A,
French Ligue 1, Dutch Eredivisie, and Belgian First Division A. For each league,
we used the 2014/15, 2015/16 and 2016/17 seasons to train our models. This
training set consists of 5967 games (some games in the 2014/15 and 2015/16
season were ignored due to missing events). The 2017/18 season was set aside as
a test set containing 2227 games. Due to the home advantage, the distribution
between wins, ties and losses is unbalanced. In the full dataset, 45.23% of the
games end in a win for the home team, 29.75% end in a tie and 25.01% end in a
win for the away team.
To assess the pre-game strength of each team, we scraped Elo ratings from
http://clubelo.com. In the case of association football, the single rating difference
between two teams is a highly significant predictor of match outcomes [9].

4.2 Model evaluation

To set the model parameters, we 3-fold cross validated on the 2014/15, 2015/16
and 2016/17 seasons with respect to the Ranked Probability Score (RPS) [6]:
2 i i
1X X X
RP St = ( pt,j − ej ) 2 , (4)
2 i=1 j=1 j=1

where pt = [P (Y = win | xt ), P (Y = tie | xt ), P (Y = loss | xt )] are the

estimated probabilities at a time frame t and e encodes the final outcome of the
 Who Will Win It? 7

game as a win (e = [1, 1, 1]), a tie (e = [0, 1, 1]) or a loss (e = [0, 0, 1]). This
metric reflects that an away win is in a sense closer to a draw than a home win.
That means that a higher probability predicted for a draw is considered better
than a higher probability for home win if the actual result is an away win.

Win Tie Loss

Actual Actual
probability LR probability LR

Predicted Predicted
probability probability

Actual Actual
probability RF probability Our proposed model

Predicted Predicted
probability probability

Fig. 1: Probability calibration curves and histograms of the predicted probabilities.

Only the Bayesian classifier has well calibrated win, draw and loss probabilities.

Because our model uses strength parameters which are time dependent, we
evaluate the quality of the estimated win probabilities on the external 2017/18
season. This evaluation is a challenging task. For example, when a team is given
an 8% probability of winning at a given state of the game, this essentially means
that if the game was played from that state onwards a hundred times, the team
8 P. Robberechts et al.

is expected to win approximately eight of them. This cannot be assessed for

a single game, since each game is played only once. Therefore, we collect all
games for which our model predicts a 8% home win probability and then look at
whether about 8% of those actually resulted in a home win. This is reflected in the
probability calibration curves in Figure 1. Only our proposed Bayesian classifier
has a good probability calibration curve. Among the three other models, the
RF classifier performs well, but the predictions for the probability of ties break
down in late game situations. It never predicts a high tie probability. Similarly,
the LR and mLR models struggle to accurately predict the probability of ties.
Additionally, their win and loss probabilities are also not well calibrated.

Besides the probability calibration, we also look at how the accuracy and
RPS of our predictions on the test set evolve as the game progresses (Figure 2).
To measure accuracy, we take the most likely outcome at each time frame and
compare this with the actual outcome at the end of the game. Both the RPS
and accuracy of all in-game win probability models improve when the game
progresses, as they gain more information about the final outcome. Yet, only the
Bayesian model is able to make consistently correct predictions at the end of each
game. For the first few time frames of each game, the models’ performance is
similar to a pre-game logistic regression model that uses the Elo rating difference
as a single feature. Furthermore, the Bayesian model clearly outperforms the LR,
mLR and RF models.

Accuracy Accuracy RPS

RPS
100% RF 0.25
Bayesian model
LR

90% Pre-game
mLR 0.20

RF
80%
0.15

70%
0.10

60%
LR
0.05
mLR
50% Pre-game
Bayesian model
0.00
0 20 40 60 80 100 0 20 40 60 80 100
Time frame Time frame

Fig. 2: All models’ performance improves as the game progresses, but only our
Bayesian model makes consistently correct predictions at the end of each game.
Early in the game, the performance of all models is similar to an Elo-based
pre-game win probability model.
 Who Will Win It? 9

4.3 Feature importance

In addition to the predictive accuracy of our win probability estimates, it is

interesting to observe how these estimates are affected by the different features.
To this end, we inspect the simulated traces of the weight vector α for each
feature. In the probabilistic framework, these traces form a marginal distribution
on the feature weights for each time frame. Figure 3 shows the mean and variance
of these distributions. Primarily of note is that winning more duels has a negative
effect on the win probability, which is not what one would intuitively expect. Yet,
this is not a novel insight.4 Furthermore, we notice that a higher Elo rating than
the opponent, previously scored goals, yellow cards for the opponent and more
successful attacking passes all have a positive impact on the scoring rate. On
the other hand, receiving red cards decreases a team’s scoring rate. Finally, the
effect of goals, yellows and attacking passes increases as the game progresses.
Red cards and duel strength have a bigger impact on the scoring rate in the first
half. For the pre-game Elo rating differential, mainly the uncertainty about the
effect on the scoring rate increases during the game.

Weights Rating Differential Team Goals

Yellows Reds

Attacking Passes Duel Strength

Time

Fig. 3: Estimated mean weight and variance for each feature per time frame.

4
https://fivethirtyeight.com/features/what-analytics-can-teach-us-about-the-
beautiful-game/
10 P. Robberechts et al.

5 Use Cases
In-game win-probability models have a number of interesting use cases. In this
section, we first show how win probability can be used as a story stat to enhance
fan engagement. Second, we discuss how win probability models can be used as a
tool to quantify the performance of players in the crucial moments of a game. We
illustrate this with an Added Goal Value (AGV) metric, which improves upon
standard goal scoring statistics by accounting for the value each goal adds to the
team’s probability of winning the game.

5.1 Fan Engagement

Win probability is a great “story stat” – meaning that it provides historical con-
text to specific in-game situations and illustrates how a game unfolded. Figure 4
illustrates how the metric works for Belgium’s illustrious comeback against Japan
at the 2018 World Cup. As can be seen, the story of the game can be told right
from this win probability graph. It shows how Japan managed to take the lead,
shortly after a scoreless first half in which neither team could really threaten
the opponent. The opening goal did not faze the Belgians. Their win probability
increased as Eden Hazard hit a post. Nevertheless, Japan scored a second on a
counter-attack. In the next 15 minutes, Belgium hit a lull, which further increased
Japan’s win probability. Right when a Belgium win seemed improbable, Belgium
got the bit of luck they probably deserved when Vertonghen’s header looped
over the Japanese keeper. This shifted the momentum of the game in Belgium’s
favour and five minutes later Belgium were level, before snatching the win with
a stunning counter attack in the last second of the game. With this comeback,
Belgium created a little bit of history by becoming the first team to come from
two goals down to win a World Cup knockout match since 1970.

World Cup ‘18 - Round of 16

3 : 2 0:1 0:2 1:2 2:2 3:2

100%

80%
Belgium wins
60%

40%

20% Draw

Japan wins
15 30 HT 60 75 90 min

Fig. 4: Win probability graph for the 2018 World Cup game between Belgium
and Japan.

Undoubtedly fans implicitly considered these win probabilities too as the game
unfolded. Where football fans and commentators have to rely on their intuition
and limited experience, win probability stats can deliver a more objective view
 Who Will Win It? 11

on these probabilities. Therefore, win probability could be of interest to fans as

they watch a game in progress or afterwards, to put the (un)likeliness of certain
game situations into a historical context. For example, one could wonder whether
Belgium’s comeback was truly that exceptional. After all, only 145 World Cup
knockout games were played since that one game in 1970 of which very few (if
any) had a similar scenario, making this statistic not very valuable. According
to our model, Belgium had a win probability of about only 8% right before their
first goal. This indicates that it was indeed an exceptional performance, although
perhaps not as exceptional as the “once in 50 years” statistic suggests.

5.2 Quantifying Performance under Mental Pressure

“Clutch” performance, or performance in crucial situations is a recurring concept
in many sports – including football. Discussions about which players are the most
clutch are popular among fans5 and teams define the ability to perform under
pressure as a crucial asset.6 However, such judgements are often the product
of short-term memory among fans and analysts. Perhaps the most interesting
application of win probability is its ability to identify these crucial situations.
By calculating the difference in win probability between the current situation
and the win probability that would result after a goal, one can identify these
specific situations where the impact of scoring or conceding a goal would be much
greater than in a typical situation [18]. It is a reasonable assumption that these
situations correspond to the crucial moments of the game.
To illustrate this idea, we show how win probability can be used to identify
clutch goal scorers. The number of goals scored is the most important statistic
for offensive players in football. Yet, not all goals have the same value. A winning
goal in stoppage time is clearly more valuable than another goal when the lead
is already unbridgeable. By using the change in win probability7 when a goal
is scored, we can evaluate how much a player’s goal contributions impact their
team’s chance of winning the game. This leads to the Added Goal Value metric
below, similar to Pettigrew’s added goal value for ice hockey [17].
PKi
k=1 3 ∗ ∆P (win|xtk ) + ∆P (tie|xtk )
AGVp90i = ∗ 90 (5)
Mi
where Ki is the number of goals scored by player i, Mi is the number of minutes
played by that same player and tk is the time at which a goal k is scored.
This formula calculates the total added value that occurred from each of
player i’s goals, averaged over the number of games played. Since both a win
and a draw can be an advantageous outcome in football, we compute the added
5
https://www.thetimes.co.uk/article/weight-of-argentinas-collapse-is-too-much-for-
even-messi-to-shoulder-wl7xbzd83
6
https://api.sporza.be/permalink/web/articles/1538741367341
7
We remove the pre-game strength from our win probability model for this analysis.
Otherwise, games of teams such as PSG that dominate their league would all start
with an already high win probability, reducing a goal’s impact on the win probability.
12 P. Robberechts et al.

value as the sum of the change in win probability multiplied by three and the
change in draw probability. The result can be interpreted as the average boost
in expected league points that a team receives each game from a player’s goals.
Figure 5 displays the relationship between AGVp90 and goals per game for
the most productive Bundesliga, Ligue 1, Premier League, LaLiga and Serie A
players who have played at least the equivalent of 20 games and scored at least
10 goals in the 2016/2017 and 2017/2018 seasons. The diagonal line denotes the
average AGVp90 for a player with a similar offensive productivity. The players
with the highest AGVp90 are Lionel Messi, Cavani, Balotelli, Kane and Giroud.
Also, players such as Neymar, Lewandowski, Lukaku, Mbappé and Mertens have
a relatively low added value per goal; while players such as Austin, Balottelli,
Dybala, Gameiro and Giroud add more value per goal than the average player.

AGV per 90 min

Lionel Messi

0.6

Edinson Cavani
Mario Balotelli Harry Kane
Olivier Giroud

Anthony Modeste Cristiano Ronaldo

0.5 Charlie Austin Sergio Agüero
Luis Suárez Robert Lewandowski
Aubemayang
Ciro Immobile
Gianluca Lapadula
Mohamed Salah
Luis Fernando
0.4

Dries Mertens

Kylian Mbappé

0.3
Neymar
Romelu Lukaku

0.4 0.6 0.8 1.0

Goals per 90 min

Fig. 5: The relation between goals scored per 90 minutes and AGVp90 for the
most productive Bundesliga, Ligue 1, Premier League, LaLiga and Serie A players
in the 2016/2017 and 2017/2018 seasons.

6 Conclusions
This paper introduced a Bayesian in-game win probability model for football. Our
model uses eight features for each team and models the future number of goals
that a team will score as a temporal stochastic process. Our evaluations indicate
that the predictions made by this model are well calibrated and outperform the
typical modelling approaches that are used in other sports.

Acknowledgements
Pieter Robberechts is supported by the EU Interreg VA project Nano4Sports. Jesse
Davis is partially supported by the EU Interreg VA project Nano4Sports and the KU
Leuven Research Fund (C14/17/07, C32/17/036).
 Who Will Win It? 13

References
1. Beuoy, M.: Updated NBA win probability calculator. https://
www.inpredictable.com/2015/02/updated-nba-win-probability-calculator.html
(2015), [Online; Accessed: 2019-06-12]
2. Burke, B.: Modeling win probability for a college basketball game.
http://wagesofwins.com/2009/03/05/modeling-win-probability-for-a-college-
basketball-game-a-guest-post-from-brian-burke/ (2009)
3. Burke, B.: (WPA)explained. http://archive.advancedfootballanalytics.com/2010/
01/win-probability-added-wpa-explained.html (2010)
4. Červenỳ, J., van Ours, J.C., van Tuijl, M.A.: Effects of a red card on goal-scoring
in world cup football matches. Empirical Economics 55(2), 883–903 (2018)
5. Commission, G.: In-play (in-running) betting: Position paper. https://
gamblingcommission.gov.uk/pdf/In-running-betting-position-paper.pdf (2016)
6. Constantinou, A.C., Fenton, N.E.: Solving the problem of inadequate scoring rules
for assessing probabilistic football forecast models. JQAS 8(1) (2012)
7. Constantinou, A.C., Fenton, N.E.: Determining the level of ability of football teams
by dynamic ratings based on the relative discrepancies in scores between adversaries.
Journal of Quantitative Analysis in Sports 9(1), 37–50 (2013)
8. Ganguly, S., Frank, N.: The problem with win probability. In: Proc. of the 12th
MIT Sloan Sports Analytics Conf. (2018)
9. Hvattum, L.M., Arntzen, H.: Using Elo ratings for match result prediction in asso-
ciation football. Int. J. Forecast 26(3), 460–470 (2010)
10. Kucukelbir, A., Tran, D., Ranganath, R., Gelman, A., Blei, D.M.: Automatic dif-
ferentiation variational inference. JMLR 18(1), 430–474 (2017)
11. Lindholm, S.: Using WPA to measure pitcher effectiveness. https:
//www.beyondtheboxscore.com/2014/5/19/5723968/chicago-white-sox-chris-
sale-wpa-major-league-baseball-pitcher-effectiveness (2014)
12. Lindsey, G.R.: The progress of the score during a baseball game. J. Am. Stat. Assoc
56(295), 703–728 (1961)
13. Lock, D., Nettleton, D.: Using random forests to estimate win probability before
each play of an NFL game. JQAS 10(2) (2014)
14. McFarlane, P.: Evaluating NBA end-of-game decision-making. J. Am. Stat. Assoc
5(1), 17–22 (2019)
15. Morris, B.: When To Go For 2, For Real. https://fivethirtyeight.com/features/when-
to-go-for-2-for-real/ (2017)
16. Pelechrinis, K.: iWinRNFL: A Simple, Interpretable & Well-Calibrated In-Game
Win Probability Model for NFL. arXiv:1704.00197 [stat] (2017)
17. Pettigrew, S.: Assessing the offensive productivity of NHL players using in-game
win probabilities. In: Proc. of the 9th MIT Sloan Sports Analytics Conf. (2015)
18. Robberechts, P., Bransen, L., Van Haaren, J., Davis, J.: Choke or Shine? Quantifying
Soccer Players’ Abilities to Perform Under Mental Pressure. In: Proc. of the 13th
MIT Sloan Sports Analytics Conf. (2019)
19. Schneider, T.: What Real-Time Gambling Data Reveals About Sports: Introducing
Gambletron 2000. http://www.gambletron2000.com/about (2014)
20. Torvik, B.: How I built a (crappy) basketball win probability model. http:
//adamcwisports.blogspot.com/2017/07/how-i-built-crappy-basketball-win.html
(2017)