BLOG: Expected assists in context

Article by Tom Worville

“Expected assists measures the likelihood that a pass will be a primary assist. The model is based on the finishing location of the pass, what type of pass it was and a variety of other factors. This model is not reliant on whether a shot was taken from this pass, so credits all passes, regardless of whether they result in a shot.”

This blog outlines the process behind this model, as well as explaining how it can be valuable in evaluating performance.


If Everton’s Sandro Ramirez crosses the ball into Wayne Rooney and Rooney mistimes the header, or heads the ball over the bar, Ramirez cannot control that, just in the same way he cannot control that Rooney may time his jump perfectly and send the ball into the net.

This is one reason why counting ‘assists’ in isolation is not an entirely representative way to assess creativity. The creator is often reliant on the finishing skill of the player receiving the pass.

We also know that not all key passes create the same quality of scoring opportunity, and this should be reflected in how we value them.

So how do we better credit players for their ability to create scoring opportunities?

The three goals shown below are all examples of how the creator can be credited based on the quality of chance they created.

Assigning credit to creators

Initially, it was simply taking the xG value from key passes. However this meant only situations were considered when the receiver got a shot away, and it also failed to acknowledge the occasions when the receiver carried the ball on to generate a better shooting position than the pass initially provided (consider Lionel Messi’s Copa Del Rey goal against Bilbao, shown later in this article).

The expected assists framework has now evolved to ensure credit is appropriately portioned and a more accurate reflection of each situation is captured.

Expected assists measures the likelihood that a pass will be a primary assist based on where the pass was received, what type of pass it was and a variety of other factors. This model is not reliant on whether a shot was taken from this pass, so is able to credit beyond the final pass.

By assigning credit to the creator for the quality of the chance created, we get a deeper level of insight beyond just the quantity of the chances made.

The above example is Andy Carroll’s memorable bicycle kick against Crystal Palace. Here, Manuel Lanzini (10) passes to Michail Antonio (30) who crosses to Andy Carroll (9) with the Englishman dispatching the ball into the top right hand corner.

Opta’s expected goal model gives Carroll a 5% chance of scoring based on the characteristics of the opportunity. Previously, Antonio would have been credited with 0.05 as that is the xG value of Carroll’s shot. However, in this updated model the subsequent shot is not relative to the pass, so Antonio is credited with an xA value of 0.034, despite the end point of the pass being the same location as the shot. This is because the model acknowledges that Carroll shooting was not a guarantee, as shown below.

–          A pass of those characteristics becomes an assist 3.4% of the time

–          A shot with those characteristics becomes a goal 5% of the time

This is Lionel Messi’s mazy solo run and goal against Athletic Bilbao from the 2015 Copa del Rey final. In this sequence, Dani Alves (22) passes the ball to Messi (10) who dribbles past several Bilbao players and scores.

The xG on Messi’s chance is 9%. After dancing past several Bilbao players, Messi generated himself a fairly high quality scoring opportunity, however this was of course not the case when he received Dani Alves’ pass.

If we were to adopt the ‘xG from Key Pass’ approach, Alves would be credited with 0.09, which we know is not an accurate reflection of the quality of the chance he ‘created’.

For this reason, every completed pass has an xA value, even if the next action is not a shot. The logic being that every pass can indeed become an assist – with Dani Alves’ pass being a perfect example of this. Indeed, Alves’ pass results in an assist less than 1% of the time.

A non-shot driven approach

As OptaPro’s recent work on possessions demonstrates, it is important to be able to analyse and credit events before the key pass and shot.

Therefore the model rewards the players who pass into dangerous areas without relying on their teammates to take a shot. They will still be credited when the receiver does not shoot because they have been dispossessed, or if they pass the ball on to a team-mate in an even better position.

Several of the names in this table have lower assist totals than their expected assists values, but the expected assist totals look encouraging.

David Silva is of particular note. His 10.1 xA value (against seven assists) is the perfect example of how all passes are acknolwedged as well as the passes that lead directly to shots or goals.

Taken from ProVision, the two heatmaps below show David Silva’s passes from the 2016/17 Premier League season, allowing us to further understand why the midfielder ranks so highly on this metric.

Marko Arnautovic is another player of note. With Stoke City, he played in a team that ranked only 13th in shots taken. Despite featuring for a low volume shot side, he still ranks seventh in this metric, demonstrating its impact in identifying players making dangerous passes, even when they don’t result in a shot.

In evaluating and quantifying creativity, expected assists is the next step in providing a more informed framework that appropriately credits all passes, rather than just those that directly result in a shot or goal.

Back to analysis