Double pivot gets Football Whispers the win

The team behind Football Whispers submitted its entry to the DataIQ Talent Awards only hoping that the data sector would understand what it is doing – trawling through transfer chatter for nuggets of truth. The sector not only understood, …
kieren_beltrame

The team behind Football Whispers submitted its entry to the DataIQ Talent Awards only hoping that the data sector would understand what it is doing – trawling through transfer chatter for nuggets of truth. The sector not only understood, but rewarded and applauded the team – 28 data scientists, front- and back-end developers and content creators – for creating a player transfer prediction site.

Kieren Beltrame, Football WhispersChief technology officer Kieren Beltrame told DataIQ how the company takes publicly-available data from social media, news, blogs, forums and other sources to uncover what might be happening in the transfer market and then validates if it is true. 

Beltrame explained that transfer prediction is a two-step process. During discovery, it captures all the possible rumours and considers which is the most credible. If there are two rumours about a player moving to Arsenal, the one from a person who works at the club tweeting something heard at the pub will have a greater chance of being successful than the one from a person who lives in Canada.

For the validation part, Beltrame said the firm cross-validates with a number of factors – which it calls features – including the player’s age, nationality and the duration of their contract. Those features are put into a predictive model and it “spits out” a score with a confidence interval.

According to the CTO, the most challenging part was the technique of entity extraction – the process of understanding the context in which the information is provided. “When you see unstructured text like a tweet or a news article, things like names of companies or players or football clubs or brands are called entities and we need to extract those. But quite often people will use pseudonyms,” he said. A great example is Arsenal also being known as the Gunners. If some content refers to “gunners”, is it talking about something military or the football team?

Also, if just the surname of a player is mentioned, Football Whispers would again have to use entity extraction. Beltrame said there are times when six or seven players with a similar surname could be a potential match for a club. He said: “Trying to figure out which player specifically is being talked about or ensuring that the domain is football clubs and not military has been one of the bigger challenges we’ve been able to solve.”

Beltrame said the solution involved using supervised and unsupervised machine learning. He explained that supervised machine learning involves “a bunch of people” looking at articles to see whether they relate to transfer news or not or if they relate to a particular club or not. That creates a database of content that they know relates to a specific topic which they can then “train a model on.”

Unsupervised machine learning involves the team pitting models against each other as they are trying to get the machine firstly to learn whether something is about football or not and, secondly, whether it is about a transfer or not and then, subsequently, whether it’s about a specific club or player or not. “That process is constantly evolving, we are constantly looking at false positives for examples, so every time the model gets it wrong, it gets better next time and this process is continuing 24/7,” said Beltrame.

Vivion Cox, Football WhispersChief executive Vivion Cox said that Football Whispers also takes into account a “time decay element” and is able work out if a rumour will go viral. He said: “Much how Facebook looked at the way viral content grew, we would reverse engineer that and see rumours as they were starting to grow. There’s a certain inflexion point where you could start to predict that event is going to happen.”

He said that, with just $1,000 of marketing spend, Football Whispers itself went viral and amassed 100,000 new users within four weeks. Sports brands soon wanted to partner, with an approach from Sky Sports about using technology to power the broadcaster’s transfer segments. Thereafter, ESPN and Yahoo! also came on board as partners. Now, three million users access Football Whispers content on its site, while a further 10 to 12 million experience the content through its partners.

Cox explained that they felt creating a platform and allowing other media sites to plug into their data would be the best way to grow. “Football Whispers would get a far greater reach than a typical start-up trying to get everybody to one destination,” he said.

Football Whispers began life as a tech consultancy back in 2007, when Cox and some fellow Cranfield MBA graduates got the contract to extract Skype from eBay and build it as a stand-alone business. As they were preparing Skype for an IPO, Microsoft came in and bought it, so Cox had to change tack. “We thought we could try and build our own tech company and bring a little bit of Silicon Valley magic to Milton Keynes and a small office in London,” he said.

In 2011, they formed Klood, which built software and AI around working out the optimum time to release content on social media. However, when they saw that they were being overtaken by competitors with much deeper pockets, they decided to do a hard pivot and concentrate on big data, social data and digital data. “It was about making things more insightful,” Cox said of the Klood radar product.

At that point, Klood took funding from investors, including cardiologist Dr Mo Sacoor. A few months after this investment, Sacoor broached the idea of creating a football news site with Cox, who remembered him asking, “is it something you can do and is it something you’re interested in doing?” 

Cox’s reply was, “the answer to both is yes, but the reality is, it’s probably the most saturated space on earth and we’d need to think of a different angle.” That angle was to take everything they had learnt from social media and big data and apply it to the football news space.

Sacoor is also an alumnus and champion of the University of Sheffield. The team found out that the Organisations, Information and Knowledge Group (OAK) of the university is considered “a world leader in mining complex data entities from information networks, such as social media,” said Cox. He said they decided to partner with OAK to accelerate their go-to-market plans and incorporate the best breed of algorithms and technology in their transfer predictions.

However, Cox thinks that Football Whispers can branch out beyond algorithms. It already has a fanzine with professional writers covering transfers, tactics and the like. Of course, there is a technological twist and Football Whispers is employing technology to make story writing quicker and easier for its content creators. “Our next frontier is around robo-writing and auto-video and auto-graphic creation, so our football experts can be presented in a matter of seconds with all the content they need to produce stories ahead of anyone because the algorithms are surfacing which stories to focus on,” said Cox.

Other innovations they are working on are algorithms that process the data to create predictive line-ups and a player valuation algorithm to be launched in time for next season. Cox explained: “It’s the mathematical assessment of players, bridging the gap between the data geeks and the media companies that just rank a player out of ten.” He thinks that big brands may be very interested in the player valuation algorithm as they will be able to identify rising stars and engage them early in sponsorship deals or as brand ambassadors.

Fans of other sports may also benefit from an injection of data into their gossip as Cox said that they are now rolling out this type of thinking and technology to other sports, like baseball and basketball. 

Incidentally, the concept of assessing data about sports players to maximise the performance of their team originated in the US with baseball team Oakland Athletics. In the late 1990s and 2000s, general manager Billy Beane implemented sabermetrics – the empirical study of baseball – to obtain undervalued players and drive the team up the ranks of the major league teams. Data analysis of sports has landed in the UK also, for example with Matthew Benham, the majority shareholder of Danish club FC Midtjylland, who bought British club Brentford in 2015 and intends to use mathematical modelling to help recruit players.

For its efforts, the Football Whispers team also took home the gong for Outstanding Startup at the Sports Technology Awards in May. However, accolades aside, Beltrame said he gets an incredible sense of satisfaction from predicting that something will happen well in advance of the mainstream press that gives a fan hope and that then comes true.

“That’s really what we’re trying to get. We’re not 100% correct yet, but we’re getting better all the time,” he stated. It seems the talk around “the beautiful game” is being refined with the lots of baseless rumours and random speculation being sent off.

Related articles: DataIQ Talent Awards 2017 – Breathrough with Data: Football Whispers

Upcoming Events

No event found!
Print Friendly, PDF & Email