Attribution models explained using FIFA World Cup data

Image courtesy of Ben Sutherland
Image courtesy of Ben Sutherland

It started with a dilemma.

For four weeks in June and July, we (like most other people) got incredibly emotionally invested in the 2018 World Cup in Russia. We had sweepstakes, a special FIFA night, a lot of Three Lions on the office stereo and a ‘lot’ of English tears on semi-final day.

And, after our Analytics whizz Sam had incredible response from his (practically famous) attribution modelling talk, we got thinking….Could marketing learn a little from football?

What do marketing attribution channels tell us about the true value of footballers? And, what can football tell us about the effectiveness of certain attribution models?

Data Nerdery plus Football

A background to attribution modelling

Attribution modelling is all about measuring the various touchpoints involved in a conversion or sale. And it’s frequently the biggest bone of contention among marketers.

These days, it is common to completely attribute the success of a conversion to its last click because it’s included in all standard Google Analytics reports (last click conversion), but what about the other marketing factors involved in leading towards that click?

Last click attribution in Analytics tends to overstate lower-funnel techniques such as PPC or branded search, but downplays the effectiveness of content marketing, PR and most SEO.

This taps into ‘probably’ the most common question we’re asked by clients and prospects – how do I know which of my techniques are most effective? How can I tell my boss what’s working?

That’s where attribution modelling comes in. It uses a variety of data models to attribute value to all influencers and interactions (social/PPC/content marketing) to explain trends and behaviours. Most importantly, it gives a more rounded account of whether your overall marketing strategy is working – and where the true value of your activity is.

So, we analysed France’s World Cup win to see what it tells us about attribution. We wanted to show the most valuable French World Cup players using a range of attribution models. After all, football can help us to understand geopolitics, economic and social trends, it can tell us about the love between parent and child, and it can explain the often-misunderstood area of attribution modelling

 

The method behind the madness.

Combining our popular methodology and incessant need to watch the football, we devised (what we think is) a creative example of attribution modelling in an attempt to explain it in layman’s terms.

Our methodology begins with rivalry dispossession, whereby the first French touch is a representation of a first click and the goal scorer a representation of a last click. In the event of a corner, free kick or pen we counted each player that featured in the lead to that situation.

So, for example, when Kylian Mbappe scored France’s fourth and final goal of the World Cup Final, the goal looked like this.

Matuidi (ball winner) – Greizmann – Nzonzi – Pogba –  Hernandez – Matuidi – Pogba – Nzonzi – Pogba –Hernandez – Mbappe (goal scorer)

Are you with us?

 

We then used that methodology across four different attribution models, illustrated in the slides below:

Blog-Image_FIFA-Attribution-Models-Explained-02-Last-Click-InfographicBlog-Image_FIFA-Attribution-Models-Explained-04-First-Click-InfographicBlog-Image_FIFA-Attribution-Models-Explained-06-Linear-Model-InfographicBlog-Image_FIFA-Attribution-Models-Explained-12-Funnel-Based-InfographicBlog-Image_FIFA-Attribution-Models-Explained-17-Time-Decay-Infographic

Then we’d see how the different models rewarded different players for their roles in the goals. The French Football Association took home £38m for winning football’s biggest prize, so we used that sum as part of the attribution model to denote value.
Let’s take a look at how the different models explain attribution.
*For the purposes of the exercise we have swapped the currency to British pound without exchanging rates.

 

Last Click Attribution Model.

As mentioned above, most often attribution is entirely awarded to the last click. In football parlance, that would be the goalscorer.

This is how the French World Cup players, or rather Les Bleus goal scorers, would be paid had Didier Deschamps used the last click attribution model to influence salaries.

Last Click Results

You may have noticed that the formula divides the winning total with fourteen goals, but only 12 French goals are accounted for. This is due to two rival teams scoring own goals: Alex Behich for Australia and Croatia’s Mario Mandzukic.

France FIFA World Cup 2018 goal scorers

Using this model, only six of the twenty-three-man squad scored goals during the Russia World Cup games, meaning almost three quarters of the squad would go without a share of the £38 million and walk away with no prize money; even though they still played vital roles in the tournament. Likewise, Antoine Griezmann and Kylian Mbappe are receiving an unfair amount of money for their exploits – particularly when you consider that 3 of Griezmann’s goals came from penalties, and the fourth was a goalkeeping howler.

This is the perfect encapsulation of why last click attribution models are somewhat simplistic.

 

 

First Click Attribution Model.

The first-click attribution works similarly to the last-click but, as the name suggests, only the player who started the move which led to the goal is recognised.

last click model results

Understandably, this does suffer from the same flaws as last-click, but if marketers combine the two, they can start to get a more nuanced understanding of attribution. The funnel-based position, which we discuss further down the page, is a model which combines the two – with a little extra thrown in.

 

 

Linear Attribution Model.

The linear attribution model is the simplest of multi-touch models. It equally divides attribution between all influencers that are factored in before conversion. So for every goal, each player that touched the World Cup football was equal to the next regardless of at what stage they passed the ball.

linear attribution model example

In the France v Australia example above, Griezmann appears to earn the most with £900,000 as the goal scorer, but his position as goal scorer doesn’t have an effect on his salary- it’s due to the goal being his second touch of the ball in the move.

You can think of linear attribution models as the communist model, or like a participation certificate. It doesn’t matter if an influencer was a goal scorer, an assist to the goal or a quick 1-2 pass, they’re attributed equally.

Here’s the formula:

linear attribution model formula

Which divides attribution evenly per game. But what if Didier Deschamps wanted to award his players’ salaries equally over the entirety of the competition?

linear attribution model example number 2

Here, a “player’s worth” has been divided by total number of touches over the entire competition (or at least the touches since dispossession led to a goal) rather than by the number of touches per game.

Linear attribution model formula

So, unlike the first example whereby the number of touches per goal influenced a salary amount (fewer total touches from dispossession to goal would award higher salaries per touch than a longer play with more touches) this second example awards salaries evenly across the board.

Here are the overall results of each player that touched the ball in the play leading to a goal.

linear attribution model results

By attributing value to each player by touches overall, rather than by touches per game, there are fairer fees being awarded. Antoine Griezmann still leads his team-mates, but more French players are being attributed with a slice of the pie. However, if a player (i.e. Griezmann) has appeared in the play of a number of goals – some with a low number of touches leading to the goal – they are being devalued for a fairer share. But is that in itself fair?

What about the goal scorers? Surely their value should be worth more than a player whose touch of the ball was a quick, low-risk 1-2?

 

 

Funnel Based Attribution Model.

As the name suggests, a funnel based attribution model attributes value to influencers based on their position in the funnel. By default, the funnel based model attributes 40% value to the first and last clicks, and 20% to the remaining touchpoints.

In our model, we added an additional position for the assist and divided the percentages as 20% first kick, 40% last kick, 30% assist and 10% for remaining touches.

That’s because it’s harder to win the ball back from the opposing player, harder to play that key pass to assist a goal, and harder to score at the end of the move.

position based attribution model example

The second France v Australia goal ended with an own goal from Behich (ouch!) meaning no final kick attribution was awarded here, but Kante lands 20% attribution for dispossession and being the first French kick and Pogba earns a lovely 30% attribution for assisting the goal on top of his earlier touches. The four players with a touch share 10% between five touches, since Pogba got there twice.

position based attribution model formula

Here’s another example to get your head around the formula:

position based attribution model example two

Pogba wins the dispossession and claims 20% attribution value, Mbappe claims 40% as goal scorer and since Giroud’s pass was a deflection rather than an assist we have attributed his efforts as a touch – giving him 10%.

Using our position based attribution model, the players involved in the plays leading toward a goal would be paid the following salaries:

position based attribution model results

Remember, the total amount although shared from the impressive £38 million win will not total 38,000,000 since own goals were scored and there are some instances of zero assists per game.

 

 

Time Decay Attribution Model.

Another slightly more advanced multi-channel attribution model is time decay. Time decay attribution model attributes the least value to the first click and the most value to the final click; increasing the value per influencer as time goes on.

Now, whilst that sounds pretty simple, the calculation can be a little tricky – especially in a non-eCommerce situation. Let’s look at a time decay example to explain:

Blog-Image_FIFA-Attribution-Models-Explained-18-Time-Decay-Example

As above, every influencer behind the conversion – in this case the goal – has been attributed less than the influencer before it, making the first kick the least important and the final kick (appropriately in this instance being Benjamin Pavard’s wonderstrike) getting rewarded the most.

In a bid to create the formula to attribute time decay we need to find the half-life.

time decay attribution model formula

Half-life is the time required for a quantity, in this case the goal duration, to reduce to half its initial value. Typically, a time decay attribution model has a half-life value of 7 meaning an interaction that happened seven days prior to the conversion gets 0.5 the credit. In our case the half-life wasn’t so simple (since it’s based on time), for this reason we decided to keep the half-life the same for every game.

With the half-life determined, the time decay for each goal is calculated, normalised and used to find the value of individual touches. Eventually using the value of individual touches for each play, the player worth is calculated.

time decay attribution model advanced formula

Do not panic, I repeat, do not panic. Yes, the formula looks like something an A Level maths teacher would enjoy throwing at you as a curveball, but Analytics tools with attribution modelling integrated (like Google) will do the hard calculation work for you. In our case it wasn’t quite so simple since we were working with non-digital data that didn’t exist anywhere other than on our spreadsheet.

To see our data and workings-out you may access our spreadsheet here. It includes all models broken down on separate sheets.

time decay attribution model results

Time decay attribution models work well in situations whereby the product needs deliberation, for example a sofa. However, in our FIFA World Cup experiment to uncover which player should be paid what amount from the total winnings, we would opt for the position based model that attributes greater value to goal assists and goal scorers who ultimately place the ball where it needs to be.

 

So, what came of all this?

multi-channel attribution modelling

Our table above shows different attribution models give very different results, showing the importance of moving away from a sole last click model and creating custom attribution models that are specific to a business.

And it would be fair to say that no one custom model will work for all businesses, since the business model and product type will heavily influence the attribution model. What’s important is that attribution isn’t an afterthought – and more importantly still, it mustn’t be ignored. Brands need to pick a model (or ideally a series of models) that explain how users are reacting to their marketing activity, and which parts of a strategy are working well. That’s how great teams become world beaters.

 

We hope our explanation has been of some use to you, however, if our breakdown has blown your mind we can offer you two options.

  1. A consultancy session training yourself/your team about attribution modelling on an individual basis.
  2. We’ll build a custom model(s) specifically tailored to your business.

learn more

 

 

 

*Disclaimer: The author opted to avoid any England talk; it still hurts.

Photo credit: Main image courtesy of Ben Sutherland‘s amazing World Cup Final album under CC By 2.0)