A Markov Chain Approach to Win Probability in Rugby Union

arXiv:2610.04813 · stat.AP · Submitted 2026-10-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "A Markov Chain Approach to Win Probability in Rugby Union".

Jane: Here's the gist of the paper. Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches The resulting model supports simulation and analysis of in-game trajectories, including win probability and expected point estimates from any given state.

Tom: Markov Chain Framework The authors present a Markov Chain framework that represents Rugby Union as a finite state space over score differential, remaining time, and pitch location This approach has two practical advantages for the setting.

Jane: First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts.

Tom: Second, and more importantly given our data constraints, the model is substantially more data-efficient.

Paper summary: Jane: So we’ve seen the high-level idea—the paper by Watts and Pipping-Gamón proposes using a Markov Chain framework to model rugby union as a finite state space defined by time, field position, and score difference. They claim this helps move beyond models that only predict final outcomes to actually simulate game evolution.

Tom: Right, the core thesis is that because of rugby union’s continuous play and variable possessions, standard approaches struggle with those complexities, so they build a state-space model that captures the context at any given possession.

Lu: The methodology involves defining the state as a tuple of time bucket, field position in meters from the try line, and score bucket based on the signed difference from whichever team has possession. This gives them a total of three hundred thirty-six possible states they work with.

Meng: I’m wondering about those buckets again; specifically how they handle the score grouping into seven buckets and the time division into six segments—does that granularity create too many variables to manage during training?

Lalam: They use those specific buckets to define their transition probabilities, which are estimated empirically from historical match data, meaning each entry in the transition matrix represents a specific observed frequency of moving between those states.

Jane: And they emphasize that because it’s generative, you get full trajectories rather than just isolated point estimates for win probability. This lets you see not just the win chance but the distribution over final margins and time to next score.

Tom: It’s a big shift from static prediction; now you can simulate forward from any state to recover a whole picture of what might happen in terms of momentum shifts during the game.

Lu: They also compare their method against existing work in American football, noting that while models like random forests and XGBoost are used, this Markov chain estimates transition probabilities from a larger pool of in-game state transitions.

Meng: So they’re saying that instead of learning how to predict the final outcome directly from match results, this method learns the probabilities of moving between states during actual play, which should lead to more stable estimates with less data.

Lalam: The paper suggests this is particularly valuable because it produces transition probability estimates that are considerably more stable given a one season sample than a model that must learn its structure solely from final match outcomes alone.

Jane: That stability is crucial for real-world application, and they show that their model performs similarly to other win probability models in terms of accuracy on the training set.

Tom: And they quantify that performance by showing a twenty one point four percent reduction in mean square error relative to the naive baseline model, which predicted a fifty percent win probability for each team.

Lu: The calibration analysis confirms that their Markov chain model is well-calibrated, with an R squared of zero point nine seven four on the training set and zero point nine one five out of sample <ref:2610.04813#pg3>.

Meng: I’m still focused on deployment costs; how much computational overhead does running this forward simulation for a single game state entail? If we have to run thousands of Monte Carlo simulations, that adds significant operational cost.

Lalam: The paper confirms they use Monte Carlo simulation to estimate win probabilities by running the Markov chain forward many times from a given state to generate the distribution of possible outcomes.

Conclusion: Jane: So, looking at the title and authors of this paper, A Markov Chain Approach to Win Probability in Rugby Union, what’s the big picture implication for people who don't read technical papers?

Tom: Essentially, it means we’ve moved past just trying to guess who wins based on final scores; now we have a way to understand the underlying mechanics of how a game unfolds moment by moment.

Lu: The authors are using this state-space modeling approach specifically because rugby union’s continuous nature makes standard models inadequate for capturing the sequence of events accurately.

Meng: For someone who only listens to the show, this means instead of just getting a final prediction, you get a dynamic view of how likely a certain outcome is based on where the game is right now.

Lalam: It changes things because it allows you to estimate win probabilities for any specific game state in real time, factoring in the stochastic nature of rugby and all those different pathways through which a match can go.

Tom: It’s about providing a tool that supports simulation and analysis of those in-game trajectories, including expected point estimates from any given state.

Jane: So it offers a way to analyze how games evolve between states, giving you insight into the evolution of momentum during the match, which is what this Markov Chain Approach to Win Probability in Rugby Union is all about.

University of Virginia · University of Pennsylvania

stat.AP

Submitted: 2026-10-03

Updated: 2026-11-14

Comments: 9 pages, 3 figures, 3 tables. Code & replication package: https://github.com/WhartonSABI/rugby-wp

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 73/100

The gist: Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches The resulting model supports simulation and analysis of…

Key concepts

State Space S
The state space represents every possible situation in the game, defined by a tuple of (time bucket, field position, score bucket). This decomposition allows the model to capture the complex context of Rugby Union at any given moment.
Time Bucket T
Time is divided into six segments based on seconds remaining in each half. These buckets help capture how different stages of the match—like the start, middle, or end of a half—influence future game dynamics.
First-Order Markov Chain
This framework assumes that the next state depends only on the current state and not on the entire history of previous states. This 'memoryless' property makes the model computationally efficient for predicting immediate transitions in Rugby Union.
Monte Carlo Simulation
This technique runs the Markov chain forward many times starting from a specific game state to generate a distribution of possible final outcomes. This is used to calculate win probability by counting how often each team wins across all simulations.

Terminology

Summary

Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches The resulting model supports simulation and analysis of in-game trajectories, including win probability and expected point estimates from any given state. p. 0

Where the paper says it

edu October 2026 Abstract Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches. We present a Markov Chain framework that represents Rugby Union as a finite state space over score differential, remaining time, and pitch location, with transition probabilities estimated from historical match data.

Markov Chain Framework

The authors present a Markov Chain framework that represents Rugby Union as a finite state space over score differential, remaining time, and pitch location This approach has two practical advantages for the setting. First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts. p. 0

Where the paper says it

This approach has two practical advantages for our setting. p. 0

Where the paper says it

We build a first order-markov chain built from a state space based on time, field position, and the current score. This approach has two practical advantages for our setting.

First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts, which supports richer post-hoc interpretation of trajectories.

p. 0
Where the paper says it

This approach has two practical advantages for our setting. First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts, which supports richer post-hoc interpretation of trajectories.

Second, and more importantly given our data constraints, the model is substantially more data-efficient. p. 0
Where the paper says it

First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts, which supports richer post-hoc interpretation of trajectories. Second, and more importantly given our data constraints, the model is substantially more data-efficient.

p. 0
Where the paper says it

First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts, which supports richer post-hoc interpretation of trajectories. Second, and more importantly given our data constraints, the model is substantially more data-efficient.

Defining the State Space

The state space S is defined as a tuple s = (t, f, g) representing the game context at a given possession The full state space decomposes as a Cartesian product S = T × F × G.

** Time bucket T:**

The time bucket T is divided into six segments based on seconds remaining

T = 6 buckets where start and end correspond to the first and last two minutes of each half, respectively, and middle covers the remainder.

** Field position F:**

Field position is expressed as the possession team’s distance in meters from the opposition’s try line

F is discretized into eight ordinal zones centered on meter lines such as 2.5, 13.5, 31, 45, 55, 69, 86.5, and 97.5.

** Score bucket G:**

The score margin is expressed from the perspective of the possession team as a signed difference δ = scoreposs − scoreopp

G is coarsened into seven ordinal buckets such as down 14p if δ ≤ −14, tied if δ = 0, and up 1 6 if 1 ≤ δ ≤ 6.

The total number of states is |S| = 336.

Transition Dynamics and Estimation

A first-order Markov chain is characterized by a transition probability matrix P, where each entry psn,sn+1 represents the probability of transitioning from state sn to state sn+1. Each row of P corresponds to a current game state sn ∈ S (defined by time bucket, field position, and score bucket), and each column represents a possible successor state sn+1. p. 3

Where the paper says it

4 Transition Matrix A first-order Markov chain is characterised by its transition probability matrix P, where each entry psn,sn+1 represents the probability of transitioning from state sn to state sn+1: P(Sn+1 = sn+1 | Sn = sn) = psn, sn+1 , sn, sn+1 ∈ S. In our rugby model, each row of P corresponds to a current game state sn ∈ S (defined by time bucket, field position, and score bucket), and each column represents a possible successor state sn+1.

All transition probabilities are estimated empirically from historical match data. p. 3
Where the paper says it

The columns represent transitions to different field-position states, possession changes (PC), and scoring events (S3, S5, S7 for 3, 5, and 7 points respectively). All transition probabilities are estimated empirically from historical match data, with each entry psn,sn+1 representing the observed frequency of transitioning from sn to sn+1 in the dataset.

Simulation and Performance

To estimate win probabilities, the authors employ Monte Carlo simulation, running the Markov chain forward from a given game state many times to generate a distribution of possible outcomes. Win probability is then calculated as the proportion of simulations in which each team wins P(Home Win State) = Number of simulations where Home wins / N. p. 4

Where the paper says it

2 Aggregating Simulations By running this process N times we obtain a distribution of final outcomes. Win probability is then calculated as the proportion of simulations in which each team wins: P(Home Win | State) = Number of simulations where Home wins N (3) Consider a match state in the middle of the second half where the home team holds a 21-14 lead with possession at the 45-meter line.

The Markov chain model achieves a 21. p. 5
Where the paper says it

7% Naive Baseline 0. 239 – The Markov chain model achieves a 21.

4% reduction in MSE relative to the naive baseline (predicting 50% win probability for each team). p. 5
Where the paper says it

239 – The Markov chain model achieves a 21. 4% reduction in MSE relative to the naive baseline (predicting 50% win probability for each team), performing competitively with machine learning alternatives such as XGBoost (23.

Calibration and Analysis

The calibration plots demonstrate that our Markov chain model is well-calibrated on the training set. We quantify this with the R2 of the binned points against the diagonal, which is 0. p. 6

Where the paper says it

The calibration plots demonstrate that our Markov chain model is well-calibrated on the training set. We quantify this with the R² of the binned points against the diagonal, which is 0.

974 in sample and 0. p. 6
Where the paper says it

We quantify this with the R² of the binned points against the diagonal, which is 0. 974 in sample and 0.

915 out of sample. p. 6
Where the paper says it

974 in sample and 0. 915 out of sample.

The distribution of final margins is much wider when more time remains in the game. p. 7
Where the paper says it

(a) Final score margin given down 3 points in the middle of the first half (b) Final score margin given down 3 points in the middle of the second half Figure 3 The distribution of final margins is much wider when more time remains in the game. Conversely, as time runs out, the current score difference increasingly dominates, becoming the most likely final margin.

This approach allows us to estimate win probabilities for any game state, accounting for the stochastic nature of Rugby and the multiple pathways through which a match can unfold. p. 4
Where the paper says it

Running 1000 simulations from this state confirms this intuition: Table 2: Aggregated Simulation Results Home Wins Away Wins Ties Home Win Prob Away Win Prob Tie Prob 832 129 39 0. 039 This approach allows us to estimate win probabilities for any game state, accounting for the stochastic nature of Rugby and the multiple pathways through which a match can unfold.

Future Work

With multiple seasons of data, researchers could stratify matrices by team strength or adjust scoring transitions more explicitly using offensive/defensive parameters for each team. We anticipate even better calibration once this has been accounted for. p. 7

Where the paper says it

With multiple seasons of data, researchers could stratify matrices by team strength or adjust scoring transitions more explicitly using offensive/defensive parameters for each team. We anticipate even better calibration once this has been accounted for.

The dataset was obtained from a previously published study (Martinez-Arastey et al. p. 1
Where the paper says it

Within Rugby Union itself, quantitative work has begun to adopt expected-points frameworks adapted from American football. Martinez-Arastey et al.

p. 7
Where the paper says it

We anticipate even better calibration once this has been accounted for. 1 Data The dataset was obtained from a previously published study (Martinez-Arastey et al.

, 2025). The model is reasonably well calibrated and performs similarly to other win probability models. p. 7
Where the paper says it

Simply, with more data the model would perform better. That being said, our model is reasonably well calibrated and performs similarly to other win probability models.

REFERENCES

Baldwin, B. (2021). Nfl win probability from scratch using xgboost in r. p. 8

Where the paper says it

References Baldwin, B. Nfl win probability from scratch using xgboost in r.

Open Source Football Blog

Bukiet, B., Harold, E. p. 8

Where the paper says it

Open Source Football Blog. , Harold, E.

R., and Palacios, J. p. 8
Where the paper says it

, Harold, E. , and Palacios, J.

L. (1997). A markov chain approach to baseball. p. 8
Where the paper says it

, and Palacios, J. A markov chain approach to baseball.

Oper. Res., 45:14–23

Goldner, K. (2012). A markov model of football: Using stochastic processes to model a football drive. p. 8

Where the paper says it

Goldner, K. A markov model of football: Using stochastic processes to model a football drive.

Journal of Quantitative Analysis in Sports, 8(1):1–18

Lock, D. and Nettleton, D. p. 1

Where the paper says it

This makes the transition-probability estimates considerably more stable given a one-season sample than a model that must learn its structure from final match outcomes alone. In American football, win probability models are now well established and inform coaching decisions and the evaluation of play-level outcomes (Lock and Nettleton, 2014; Yurko et al.

(2014). Using random forests to estimate win probability before each play of an nfl game. p. 8
Where the paper says it

and Nettleton, D. Using random forests to estimate win probability before each play of an nfl game.

Journal of Quantitative Analysis in Sports, 10(2):197–205

Martinez-Arastey, G., Datson, N. p. 8

Where the paper says it

Martinez-Arastey, G. , Datson, N.

, Smith, N. p. 8
Where the paper says it

, Datson, N. , Smith, N.

, and Robins, M. p. 8
Where the paper says it

, Smith, N. , and Robins, M.

(2025). Foundations of expected points in rugby union: A methodological approach. p. 1
Where the paper says it

(2025) lay methodological foundations for expected points in rugby union, with their phase-level data making it possible to construct the state space underlying our model. Watts and Pipping-Gamón (2026) applied an expected-points approach to a more specific in-game decision problem, evaluating whether kicking for goal or for touch is the better choice following a penalty.

Journal of Sports Analytics, 11:1–14

Newton, P. K. and Aslam, K. p. 0

Where the paper says it

1 Introduction Markov models have been applied extensively in various sports including American Football (Goldner, 2012), Baseball (Bukiet et al. , 1997), and Tennis (Newton and Aslam, 2009).

(2009). Monte Carlo tennis: A stochastic Markov chain model. p. 8
Where the paper says it

and Aslam, K. Monte Carlo tennis: A stochastic Markov chain model.

Journal of Quantitative Analysis in Sports, 5(3):1–44

Romer, D. (2006). Do firms maximize? p. 8

Where the paper says it

Journal of Quantitative Analysis in Sports, 5(3):1–44. Do firms maximize?

evidence from professional football. p. 8
Where the paper says it

Do firms maximize? evidence from professional football.

Journal of Political Economy, 114(2):340–365

Watts, K. and Pipping-Gamón, J. p. 0

Where the paper says it

A Markov Chain Approach to Modeling Rugby Union Kenny Watts University of Virginia xen5ek@virginia. edu Jonathan Pipping-Gamón University of Pennsylvania jpipping@wharton.

(2026). Kicking for goal or touch? p. 1
Where the paper says it

(2025) lay methodological foundations for expected points in rugby union, with their phase-level data making it possible to construct the state space underlying our model. Watts and Pipping-Gamón (2026) applied an expected-points approach to a more specific in-game decision problem, evaluating whether kicking for goal or for touch is the better choice following a penalty.

An expected points framework for penalty decisions in rugby union. p. 8
Where the paper says it

Kicking for goal or touch? An expected points framework for penalty decisions in rugby union.

Journal of Quantitative Analysis in Sports

Yurko, R., Ventura, S. p. 8

Where the paper says it

Journal of Quantitative Analysis in Sports. , Ventura, S.

, and Horowitz, M. p. 8
Where the paper says it

, Ventura, S. , and Horowitz, M.

(2019). nflWAR: a reproducible method for offensive player evaluation in football. p. 8
Where the paper says it

, and Horowitz, M. nflWAR: a reproducible method for offensive player evaluation in football.

Journal of Quantitative Analysis in Sports, 15(3):163–183

--- Page 1 ---

A Markov Chain Approach to Modeling Rugby Union Kenny Watts University of Virginia xen5ek@virginia.edu Jonathan Pipping-Gamón University of Pennsylvania jpipping@wharton. p. 0

Where the paper says it

A Markov Chain Approach to Modeling Rugby Union Kenny Watts University of Virginia xen5ek@virginia. edu Jonathan Pipping-Gamón University of Pennsylvania jpipping@wharton.

upenn.edu October 2026 Abstract Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches We present a Markov Chain framework that represents Rugby Union as a finite state space over score differential, remaining time, and pitch location The resulting model supports simulation and analysis of in-game trajectories, including win probability and expected point estimates from any given state. p. 0
Where the paper says it

edu October 2026 Abstract Rugby Union’s continuous play, variable-length possessions, and multiple scoring outcomes make it difficult to model with standard approaches. We present a Markov Chain framework that represents Rugby Union as a finite state space over score differential, remaining time, and pitch location, with transition probabilities estimated from historical match data.

--- Page 2 ---

Markov models have been applied extensively in various sports including American Football (Goldner, 2012), Baseball (Bukiet et al., 1997), and Tennis (Newton and Aslam, 2009) Rugby Union shares structural features with American Football: variable field position, alternating possessions, and scoring plays with differing point totals. p. 0

Where the paper says it

, 1997), and Tennis (Newton and Aslam, 2009). Rugby Union shares structural features with American Football: variable field position, alternating possessions, and scoring plays with differing point totals.

However, Rugby Union differs in that there aren’t cleanly defined plays or drives, with continuous possessions that can vary greatly in length. p. 0
Where the paper says it

Rugby Union shares structural features with American Football: variable field position, alternating possessions, and scoring plays with differing point totals. However, Rugby Union differs in that there aren’t cleanly defined plays or drives, with continuous possessions that can vary greatly in length.

This motivates a state-space model specifically built for Rugby’s unique structure. p. 0
Where the paper says it

However, Rugby Union differs in that there aren’t cleanly defined plays or drives, with continuous possessions that can vary greatly in length. This motivates a state-space model specifically built for Rugby’s unique structure.

We build a first order-markov chain built from a state space based on time, field position, and the current score. p. 0

Where the paper says it

This motivates a state-space model specifically built for Rugby’s unique structure. We build a first order-markov chain built from a state space based on time, field position, and the current score.

This approach has two practical advantages for our setting. First, because the chain is generative rather than purely predictive, it produces full match trajectories rather than isolated point estimates, meaning we can simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts. Second, and more importantly given our data constraints, the model is substantially more data-efficient.

In American football, win probability models are now well established and inform coaching decisions and the evaluation of play-level outcomes (Lock and Nettleton, 2014; Yurko et al. p. 1

Where the paper says it

This makes the transition-probability estimates considerably more stable given a one-season sample than a model that must learn its structure from final match outcomes alone. In American football, win probability models are now well established and inform coaching decisions and the evaluation of play-level outcomes (Lock and Nettleton, 2014; Yurko et al.

, 2019). Early implementations relied on random forests and generalized additive models (Lock and Nettleton, 2014; Yurko et al. p. 1
Where the paper says it

In American football, win probability models are now well established and inform coaching decisions and the evaluation of play-level outcomes (Lock and Nettleton, 2014; Yurko et al. Early implementations relied on random forests and generalized additive models (Lock and Nettleton, 2014; Yurko et al.

, 2019), while more recent work has moved toward gradient-boosted methods capable of capturing nonlinear interactions between game-state variables (Baldwin, 2021). p. 1
Where the paper says it

Early implementations relied on random forests and generalized additive models (Lock and Nettleton, 2014; Yurko et al. , 2019), while more recent work has moved toward gradient-boosted methods capable of capturing nonlinear interactions between game-state variables (Baldwin, 2021).

These models have also proven useful beyond in-game forecasting: game-state-based expected-value frameworks have been used to evaluate the rationality of coaching decisions, such as fourth-down strategy (Romer, 2006). p. 1
Where the paper says it

, 2019), while more recent work has moved toward gradient-boosted methods capable of capturing nonlinear interactions between game-state variables (Baldwin, 2021). These models have also proven useful beyond in-game forecasting: game-state-based expected-value frameworks have been used to evaluate the rationality of coaching decisions, such as fourth-down strategy (Romer, 2006).

Within Rugby Union itself, quantitative work has begun to adopt expected-points frameworks adapted from American football. p. 1

Where the paper says it

These models have also proven useful beyond in-game forecasting: game-state-based expected-value frameworks have been used to evaluate the rationality of coaching decisions, such as fourth-down strategy (Romer, 2006). Within Rugby Union itself, quantitative work has begun to adopt expected-points frameworks adapted from American football.

Martinez-Arastey et al. (2025) lay methodological foundations for expected points in rugby union, with their phase-level data making it possible to construct the state space underlying our model. p. 1
Where the paper says it

Martinez-Arastey et al. (2025) lay methodological foundations for expected points in rugby union, with their phase-level data making it possible to construct the state space underlying our model.

Watts and Pipping-Gamón (2026) applied an expected-points approach to a more specific in-game decision problem, evaluating whether kicking for goal or for touch is the better choice following a penalty. p. 1
Where the paper says it

(2025) lay methodological foundations for expected points in rugby union, with their phase-level data making it possible to construct the state space underlying our model. Watts and Pipping-Gamón (2026) applied an expected-points approach to a more specific in-game decision problem, evaluating whether kicking for goal or for touch is the better choice following a penalty.

Both works demonstrate the value of state-based, points-oriented modeling in rugby, but neither develops a full generative state-space model capable of simulating complete match trajectories or producing dynamic, in-game win probability estimates alongside expected points. p. 1
Where the paper says it

Watts and Pipping-Gamón (2026) applied an expected-points approach to a more specific in-game decision problem, evaluating whether kicking for goal or for touch is the better choice following a penalty. Both works demonstrate the value of state-based, points-oriented modeling in rugby, but neither develops a full generative state-space model capable of simulating complete match trajectories or producing dynamic, in-game win probability estimates alongside expected points.

--- Page 3 ---

The name “Markov chain” derives from the first-order Markov property, which states that the probability of transitioning to the next state depends only on the current state, not on the sequence of states that preceded it Let Sn ∈ S denote the state at time n, where S is the state space. Formally, for any timestep n, the next state Sn is conditionally independent of all previous states Sn−2, Sn−3,. p. 1

Where the paper says it

Let Sn ∈ S denote the state at time n, where S is the state space. Formally, for any timestep n, the next state Sn is conditionally independent of all previous states Sn−2, Sn−3, .

.., S1 given the current state Sn−1: P(Sn = snew Sn−1 = sn-1,. p. 1
Where the paper says it

Formally, for any timestep n, the next state Sn is conditionally independent of all previous states Sn−2, Sn−3, . , S1 given the current state Sn−1: P(Sn = snew | Sn−1 = sn−1, .

.., S1 = s1) = P(Sn = snew Sn-1 = sn-1) (1). p. 1
Where the paper says it

Formally, for any timestep n, the next state Sn is conditionally independent of all previous states Sn−2, Sn−3, . , S1 given the current state Sn−1: P(Sn = snew | Sn−1 = sn−1, .

This property is often described as “memoryless”: the system has no memory of how it arrived at its current state, only where it is now. p. 1
Where the paper says it

, S1 given the current state Sn−1: P(Sn = snew | Sn−1 = sn−1, . , S1 = s1) = P(Sn = snew | Sn−1 = sn−1) (1) This property is often described as “memoryless”: the system has no memory of how it arrived at its current state, only where it is now.

In the context of our model, this means we assume that predicting what happens next requires only the current game situation (score differential, remaining time, and pitch location) rather than the entire history of prior plays. p. 1
Where the paper says it

, S1 = s1) = P(Sn = snew | Sn−1 = sn−1) (1) This property is often described as “memoryless”: the system has no memory of how it arrived at its current state, only where it is now. In the context of our model, this means we assume that predicting what happens next requires only the current game situation (score differential, remaining time, and pitch location) rather than the entire history of prior plays.

While we admit that certain path-dependent effects may exist, this assumption is both computationally efficient and reasonable given the difficulty of reliably inferring how specific play sequences influence future outcomes. p. 1
Where the paper says it

In the context of our model, this means we assume that predicting what happens next requires only the current game situation (score differential, remaining time, and pitch location) rather than the entire history of prior plays. While we admit that certain path-dependent effects may exist, this assumption is both computationally efficient and reasonable given the difficulty of reliably inferring how specific play sequences influence future outcomes.

--- Page 4 ---

A state s ∈ S is a tuple s = (t, f, g) representing the game context at a given possession, where the full state space decomposes as a Cartesian product S = T × F × G.

We divide each half into three segments based on seconds remaining, yielding |T | = 6 buckets.

T = 6 buckets: 1h start, 1h middle, 2h start, 2h middle, and two end segments. p. 2

Where the paper says it

3 Defining States A state s ∈ S is a tuple s = (t, f, g) representing the game context at a given possession, where the full state space decomposes as a Cartesian product: S = T × F × G. Time bucket T We divide each half into three segments based on seconds remaining, yielding |T | = 6 buckets: T = {1h_start, 1h_middle, 1h_end, 2h_start, 2h_middle, 2h_end}, where start and end correspond to the first and last two minutes of each half, respectively, and middle covers the remainder.

Field position is expressed as the possession team’s distance in meters from the opposition’s try line

F is discretized into eight ordinal zones centered on meter lines such as 2.5, 13.5, 31, 45, 55, 69, 86.5, and 97.5.

The score margin is expressed from the perspective of the possession team as the signed difference δ = scoreposs − scoreopp

G is coarsened into seven ordinal buckets such as down 14p if δ ≤ −14, up 1 6 if 1 ≤ δ ≤ 6, and up 7 13 if 7 ≤ δ ≤ 13.

The total number of states is S = 6 × 8 × 7 = 336. p. 2

Where the paper says it

State encoding In practice, a state is encoded as a concatenated string key, for example 2h_end_22_up_1_6, denoting the second half’s final two minutes, field position at the 22 m line, with the possession team leading by one to six points. The total number of states is |S| = 6 × 8 × 7 = 336.

--- Page 5 ---

State encoding In practice, a state is encoded as a concatenated string key, for example 2h end 22 up 1 6, denoting the second half’s final two minutes, field position at the 22 m line, with the possession team leading by one to six points.

--- Page 6 ---

A first-order Markov chain is characterised by its transition probability matrix P, where each entry psn,sn+1 represents the probability of transitioning from state sn to state sn+1: P(Sn+1 = sn+1 Sn = sn) = psn, sn+1, sn, sn+1 ∈ S. Each row sums to unity ensuring that from any state sn the probabilities of all possible transitions sum to certainty. p. 3

Where the paper says it

The matrix has the following general structure: P =      ps1,s1 ps1,s2 · · · ps1, PC ps1, S3 ps1, S5 · · · ps2,s1 ps2,s2 · · · ps2, PC ps2, S3 ps2, S5 · · · . psn,s1 psn,s2 · · · psn, PC psn, S3 psn, S5 · · ·      (2) where each row sums to unity, X sn+1∈S psn, sn+1 = 1 ∀ sn ∈ S, ensuring that from any state sn the probabilities of all possible transitions sum to certainty.

--- Page 7 ---

where each row sums to unity, Xsn+1∈S psn,sn+1 = 1 ∀ sn ∈ S. The columns represent transitions to different field-position states, possession changes (PC), and scoring events (S3, S5, S7 for 3, 5, and 7 points respectively). p. 3

Where the paper says it

psn,s1 psn,s2 · · · psn, PC psn, S3 psn, S5 · · ·      (2) where each row sums to unity, X sn+1∈S psn, sn+1 = 1 ∀ sn ∈ S, ensuring that from any state sn the probabilities of all possible transitions sum to certainty. The columns represent transitions to different field-position states, possession changes (PC), and scoring events (S3, S5, S7 for 3, 5, and 7 points respectively).

All transition probabilities are estimated empirically from historical match data with each entry psn,sn+1 representing the observed frequency of transitioning from sn to sn+1 in the dataset. p. 3
Where the paper says it

The columns represent transitions to different field-position states, possession changes (PC), and scoring events (S3, S5, S7 for 3, 5, and 7 points respectively). All transition probabilities are estimated empirically from historical match data, with each entry psn,sn+1 representing the observed frequency of transitioning from sn to sn+1 in the dataset.

--- Page 8 ---

Table 1: Example Play-by-Play Sequence from a Single Simulation Play State Before Event Poss. Home Away State After 1 1h start 69 tied 1h middle 69 tied Home 0 0

2 Possession Change Away 0

3 Possession Change Home 0

4 Score Score 7 Away 0,7 up 7 13

5 Score Score 7 Away 1h middle 69 up 7 13

--- Page 9 ---

Table 2: Aggregated Simulation Results Home Wins Away Wins Ties Home Win Prob Away Win Prob Tie Prob 832 129 39 0.832 0.129 0.039

--- Page X ---

(This page is a placeholder for the actual content that would follow the structure of the paper)

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Development of generative match trajectory simulators allows AI systems to simulate forward from any given state to recover not just a win probability but a distribution over final margins, time-to-nextscore, and momentum shifts.

  2. Implementation of the Markov Chain framework enables AI systems to estimate win probabilities for any game state by running Monte Carlo simulations, providing a distribution of possible outcomes which supports richer post-hoc interpretation of trajectories.

  3. The model provides dynamic, in-game win probability estimates by conditioning on the current game situation (score differential, remaining time, and pitch location), enabling forecasting that accounts for the stochastic nature of Rugby.

  4. The system allows for analysis of final score margins given a certain state by answering questions like if my team is down by X points, what is the probability we win the game by Y points, using simulations to show how the distribution of final margins changes with time remaining.

Abstract

Win probability tells us how likely a team is to win, but says little about how the rest of the match might play out. We develop a Markov chain model for rugby union that follows changes in score, time, field position, and possession, with transition probabilities estimated from phase-level match data. The same model lets us estimate win probabilities and expected points, calculate distributions for the next score and final margin, and simulate the remainder of a match from its current state. We assess predictive accuracy and calibration on 27 held-out matches after fitting the model to 105 matches from the same season. The model achieves similar test error to logistic regression and XGBoost, showing that we can model the progression of play while remaining competitive with methods that predict final outcomes directly.

Related papers