Expected Goals (xG), Uncertainty, and Bayesian Goalies

All xG model code can be found on GitHub.

Expected Goals (xG) Recipe

If you’re reading this, you’re likely familiar with the idea behind expected goals (xG), whether from soccer analytics, early work done by Alan RyderBrian MacDonald, or current models by DTMAboutHeart and Asmean, Corsica, Moneypuck, or things I’ve put up on Twitter. Each model attempts to create a probability of each shot being a goal (xG) given the shot’s attributes like shot location, strength, shot type, preceding events, shooter skill, etc. There are also private companies supplementing these features with additional data (most importantly pre-shot puck movement on non-rebound shots and some sort of traffic/sight-line metric) but this is not public or generated in the real-time so will not be discussed here.[1]

To assign a probability (between 0% and 100%) to each shot, most xG models likely use logistic regression – a workhorse in many industry response models. As you can imagine the critical aspect of an xG model, and any model, becomes feature generation – the practice of turning raw, unstructured data into useful explanatory variables. NHL play-by-play data requires plenty of preparation to properly train an xG model. I have made the following adjustments to date:

  • Adjust for recorded shot distance bias in each rink. This is done by using a cumulative density function for shots taken in games where the team is away and apply that density function to the home rink in case their home scorer is biased. For example (with totally made up numbers), when Boston is on the road their games see 10% of shots within 5 feet of the goal, 20% of shots within 10 feet of the goal, etc. We can adjust the shot distance in their home rink to be the same since the biases of 29 data-recorders should be less than a single Boston data-recorder. If at home in Boston, 10% of the shots were within 10 feet of the goal, we might suspect that the scorer in Boston is systematically recording shots further away from the net than other rinks. We assume games with that team result in similar event coordinates both home and away and we can transform the home distribution to match the away distribution. Below demonstrates how distributions can differ between home and away games, highlighting the probable bias Boston and NY Rangers scorer that season and was adjusted for. Note we also don’t necessarily want to transform by an average, since the bias is not necessarily uniform throughout the spectrum of shot distances.
home rink bias

No Place Like Home

  • Figure out what events lead up to the shot, what zone they took place in, and the time lapsed between these events and the eventual shot while ensuring stoppages in play are caught.
  • Limit to just shots on goal. Misses include information, but like shot distance contain scorer bias. Some scorers are more likely to record a missed shot than others. Unlike shots where we have a recorded event, and it’s just biased, adjusting for misses would require ‘inventing’ occurrences in order to adjust biases in certain rinks, which seems dangerous. It’s best to ignore misses for now, particularly because the majority of my analysis focuses on goalies. Splitting the difference between misses caused by the goalie (perhaps through excellent positioning and reputation for not giving up pucks through the body) and those caused by recorder bias seems like a very difficult task. Shots on goal test the goalie directly hence will be the focus for now.
  • Clean goalie and player names. Annoying but necessary – both James and Jimmy Howard make appearances in the data, and they are the same guy.
  • Determine the strength of each team (powerplay for or against or if the goaltender is pulled for an extra attacker). There is a tradeoff here. The coefficients for the interaction of states (i.e. 5v4, 6v5, 4v3 model separately) pick up interesting interactions, but should significant instability from season to season. For example, 3v3 went from a penalty-box filled improbability to a common occurrence to finish overtime games. Alternatively, shooter strength and goalie strength can be model separately, this is more stable but less interesting.
  • Determine the goaltender and shooter handedness and position from look-up tables.
  • Determine which end of the ice and what coordinates (positive or negative) the home team is based, using recordings in any given period and rink-adjusting coordinates accordingly.
  • Calculate shot distance and shot angle. Determine what side of the ice the shot is from, whether or not it is the shooters off-wing based on handedness.
  • Tag shots as rushes or rebound, and if a rebound how far the puck travelled and the angular velocity of the puck from shot 1 to shot 2.
  • Calculate ‘shooting talent’ – a regressed version of shooting percentage using the Kuder-Richardson Formula 21, employed the same way as in DTMAboutHeart and Asmean‘s xG model.

All of this is to say there is a lot going on under the hood, the results are reliant on the data being recorded, processed, adjusted, and calculated properly. Importantly, the cleaning and adjustments to the data will never be complete, only issues that haven’t been discovered or adjusted for yet. There is no perfect xG model, nor is it possible to create one from the publicly available data, so it is important to concede that there will be some errors, but the goal is to prevent systemic errors that might bias the model. But these models do add useful information regular shot attempt models cannot, creating results that are more robust and useful as we will see.

Current xG Model

The current xG model does not use all developed features. Some didn’t contain enough unique information, perhaps over-shadowed by other explanatory variables. Some might have been generated on sparse or inconsistent data. Hopefully, current features can be improved or new features created.

While the xG model will continue to be optimized to better maximize out of sample performance, the discussion below captures a snapshot of the model. All cleanly recorded shots from 2007 to present are included, randomly split into 10 folds. Each of the 10 folds were then used a testing dataset (checking to see if the model correctly predicted a goal or not by comparing it to actual goals) while the other 9 corresponding folders were used to train the model. In this way, all reported performance metrics consist of comparing model predictions on the unseen data in the testing dataset to what actually happened. This is known as k-fold cross-validation and is fairly common practice in data science.

When we rank-order the predicted xG from highest to lowest probability we can compare the share of goals that occur to shots ordered randomly. This gives us a gains chart, a graphic representation of the how well the model is at finding actual goals relative to selecting shots randomly. We can also calculate the Area Under the Curve (AUC), where 1 is a perfect model and 0.5 is a random model. Think of the random model in this case as shot attempt measurement, treating all shots as equally likely to be a goal. The xG model has an AUC of about 0.75, which is good, and safely in between perfect and random. The most dangerous 25% of shots as selected by the model make up about 60% of actual goals. While there’s irreducible error and model limitations, in practice it is an improvement over unweighted shot attempts and accumulates meaningful sample size quicker than goals for and against.

gains chart

Gains, better than random

Hockey is also a zero-sum game. Goals (and expected goals) only matter relative to league average. Original iterations of the expected goal model built on a decade of data show that goals were becoming dearer compared to what was expected. Perhaps goaltenders were getting better, or league data-scorers were recording events to make things look harder than they were, or defensive structures were impacting the latent factors in the model or some combination of these explanations.

Without the means to properly separate these effects, each season receives it own weights for each factor. John McCool had originally discussed season-to-season instability of xG coefficients. Certainly this model contains some coefficient instability, particularly in the shot type variables. But overall these magnitudes adjust to equate each seasons xG to actual goals. Predicting a 2017-18 goal would require additional analysis and smartly weighting past models.

Coefficient Stability

Less volatile than goalies?

xG in Action

Every shot has a chance of going in, ranging from next to zero to close to certainty.  Each shot in the sample is there because the shooter believed there was some sort of benefit to shooting, rather than passing or dumping the puck, so we don’t see a bunch of shots from the far end of the rink, for example. xG then assigns a probability to each shot of being a goal, based on the explanatory variables generated from the NHL data – shot distance, shot angle, is the shot a rebound?, listed above.

Modeling each season separately, total season xG will be very close to actual goals. This also grades goaltenders on a curve against other goaltenders each season. If you are stopping 92% of shots, but others are stopping 93% of shots (assuming the same quality of shots) then you are on average costing your team a goal every 100 shots. This works out to about 7 points in the standings assuming a 2100 shot season workload and that an extra 3 goals against will cost a team 1 point in the standings. Using xG to measure goaltending performance makes sense because it puts each goalie on equal footing as far as what is expected, based on the information that is available.

We can normalize the number of goals prevented by the number of shots against to create a metric, Quality Rules Everything Around Me (QREAM), Expected Goals – Actual Goals per 100 Shots. Splitting each goalie season into random halves allows us to look at the correlation between the two halves. A metric that captures 100% skill would have a correlation of 1. If a goaltender prevented 1 goal every 100 shots, we would expect to see that hold up in each random split. A completely useless metric would have an intra-season correlation of 0, picking numbers out of a hat would re-create that result. With that frame of reference, intra-season correlations for QREAM are about 0.4 compared to about 0.3 for raw save percentage. Pucks bounce so we would never expect to see a correlation of 1, so this lift is considered to be useful and significant.[2]

intra-season correlations

Goalies doing the splits

Crudely, each goal prevented is worth about 1/3 of a point in the standings. Implying how many goals a goalie prevents compared to average allows us to compute how many points a goalie might create for or cost their team. However, a more sophisticated analysis might compare goal support the goalie receives to the expected goals faced (a bucketed version of that analysis can be found here). Using a win probability model the impact the goalie had on win or losing can be framed as actual wins versus expected.

Uncertainty

xG’s also are important because they begin to frame the uncertainty that goes along with goals, chance, and performance. What does the probability of a goal represent? Think of an expected goal as a coin weighted to represent the chance that shot is a goal. Historically, a shot from the blueline might end up a goal only 5% of the time. After 100 shots (or coin flips) will there be exactly 5 goals? Maybe, but maybe not. Same with a rebound from in tight to the net that has a probability of a goal equal to 50%. After 10 shots, we might not see 5 goals scored, like ‘expected.’ 5 goals is the most likely outcome, but anywhere from 0 to 10 is possible on only 10 shots (or coin flips).

We can see how actual goals and expected goals might deviate in small sample sizes, from game to game and even season to season. Luckily, we can use programs like R, Python, or Excel to simulate coin flips or expected goals. A goalie might face 1,000 shots in a season, giving up 90 goals. With historical data, each of those shots can be assigned a probability of a being a goal. If the average probability of a goal is 10%, we expect the goalie to give up 100 goals. But using xG, there are other possible outcomes. Simulating 1 season based on expected goals might result in 105 goals against. Another simulation might be 88 goals against. We can simulate these same shots 1,000 or 10,000 times to get a distribution of outcomes based on expected goals and compare it to the actual goals.

In our example, the goalie possibly prevented 10 goals on 1,000 shots (100 xGA – 90 actual GA). But they also may have prevented 20 or prevented 0. With expected goals and simulations, we can begin to visualize this uncertainty. As the sample size increases, the uncertainty decreases but never evaporates. Goaltending is a simple position, but the range of outcomes, particularly in small samples, can vary due to random chance regardless of performance. Results can vary due to performance (of the goalie, teammates, or opposition) as well, and since we only have one season that actually exists, separating the two is painful. Embracing the variance is helpful and expected goals help create that framework.

It is important to acknowledge that results do not necessarily reflect talent or future or past results. So it is important to incorporate uncertainty into how we think about measuring performance. Expected goal models and simulations can help.

simulated seasons

Hackey statistics

Bayesian Analysis

Luckily, Bayesian analysis can also deal with weighting uncertainty and evidence. First, we set a prior –probability distribution of expected outcomes. Brian MacDonald used mean Even Strength Save Percentage as prior, the distribution of ESSV% of NHL goalies. We can do the same thing with Expected Save Percentage (shots – xG / shots), create a unique prior distribution of outcome for each goalie season depending on the quality of shots faced and the sample size we’ll like to see. Once the prior is set, evidence (saves in our case) is layered on to the prior creating a posterior outcome.

Imagine a goalie facing 100 shots to start their career and, remarkably, making 100 saves. They face 8 total xG against, so we can set the Prior Expected Save% as a distribution centered around 92%. The current evidence at this point is 100 saves on 100 shots, and Bayesian Analysis will combine this information to create a Posterior distribution.

Goaltending is a binary job (save/goal) so we can use a beta distribution to create a distribution of the goaltenders expected (prior) and actual (evidence) save percentage between 0 and 1, like a baseball players batting average will fall between 0 and 1. We also have to set the strength of the prior – how robust the prior is to the new evidence coming in (the shots and saves of the goalie in question). A weak prior would concede to evidence quickly, a hot streak to start a season or career may lead the model to think this goalie may be a Hart candidate or future Hall-of-Famer! A strong prior would assume every goalie is average and require prolonged over or under achieving to convince the model otherwise. Possibly fair, but not revealing any useful information until it has been common knowledge for a while.

bayesian goalie

Priors plus Evidence

More research is required, but I have set the default prior strength of equivalent to 1,000 shots. Teams give up about 2,500 shots a season, so a 1A/1B type goalie would exceed this threshold in most seasons. In my goalie compare app, the prior can be adjusted up or down as a matter of taste or curiosity. Research topics would investigate what prior shot count minimizes season to season performance variability.

Every time a reported result actives your small sample size spidey senses, remember Bayesian analysis is thoroughly unimpressed, dutifully collecting evidence, once shot at a time.

 Conclusion

Perfect is often the enemy of the good. Expected goal models fail to completely capture the complex networks and inputs that create goals, but they do improve on current results-based metrics such as shot attempts by a considerable amount.  Their outputs can be conceptualized by fans and players alike, everybody understands a breakaway has a better chance of being a goal than a point shot.

The math behind the model is less accessible, but people, particularly the young, are becoming more comfortable with prediction algorithms in their daily life, from Spotify generating playlists to Amazon recommender systems. Coaches, players, and fans on some level understand not all grade A chances will result in a goal. So while out-chancing the other team in the short term is no guarantee of victory, doing it over the long term is a recipe for success. Removing some the noise that goals contain and the conceptual flaws of raw shot attempts helps the smooth short-term disconnect between performance and results.

My current case study using expected goals is to measure goaltending performance since it’s the simplest position – we don’t need to try to split credit between linemates. Looking at xGA – GA per shot captures more goalie specific skill than save percentage and lends itself to outlining the uncertainty those results contain. Expected goals also allow us to create an informed prior that can be used in a Bayesian hierarchical model. This can quantify the interaction between evidence, sample size, and uncertainty.

Further research topics include predicting goalie season performance using expected goals and posterior predictive distributions.

____________________________________________

[1]Without private data or comprehensive tracking data technology analysts are only able to observe outcomes of plays – most importantly goals and shots – but not really what created those results. A great analogy came from football (soccer) analyst Marek Kwiatkowski:

Almost the entire conceptual arsenal that we use today to describe and study football consists of on-the-ball event types, that is to say it maps directly to raw data. We speak of “tackles” and “aerial duels” and “big chances” without pausing to consider whether they are the appropriate unit of analysis. I believe that they are not. That is not to say that the events are not real; but they are merely side effects of a complex and fluid process that is football, and in isolation carry little information about its true nature. To focus on them then is to watch the train passing by looking at the sparks it sets off on the rails.

Armed with only ‘outcome data’ rather than comprehensive ‘inputs data’ analyst most models will be best served with a logistic regression. Logistic regression often bests complex models, often generalizing better than machine learning procedures. However, it will become important to lean on machine learning models as reliable ‘input’ data becomes available in order to capture the deep networks of effects that lead to goal creation and prevention. Right now we only capture snapshots, thus logistic regression should perform fine in most cases.

[2] Most people readily acknowledge some share of results in hockey are luck. Is the number closer to 60% (given the repeatable skill in my model is about 40%), or can it be reduced to 0% because my model is quite weak? The current model can be improved with more diligent feature generation and adding key features like pre-shot puck movement and some sort of traffic metric. This is interesting because traditionally logistic regression models see diminishing marginal returns from adding more variables, so while I am missing 2 big factors in predicting goals, the intra-seasonal correlation might only go from 40% to 50%. However, deep learning networks that can capture deeper interactions between variables might see an overweight benefit from these additional ‘input’ variables (possibly capturing deeper networks of effects), pushing the correlation and skill capture much higher. I have not attempted to predict goals using deep learning methods to date.

Leave a Reply

Your email address will not be published. Required fields are marked *