Thursday 29 September 2016 11:34, UK
This month's Manchester derby was a thrilling contest between Jose Mourinho's United and Pep Guardiola's City. From City's first-half dominance to United's stubborn second-half response, the final outcome seemed impossible to predict.
Except, the data experts at Hewlett Packard Enterprise had done just that.
Using complex algorithms and equations to study patterns and trends across a range of statistics and previous results involving the teams and managers, they were able to correctly predict a 2-1 victory for the visitors at Old Trafford.
As well as the two Manchester clubs, they have also made an in-depth analysis of Antonio Conte and his Chelsea team and have come up with predicted scorelines for when the three managers meet this season - which you can see on their Making of a Manager page.
But how do you go about creating such a score predictor? We spoke to a couple of HPE experts to find out…
How does the predictor work?
Asaf Lahav, research and development manager - "The mechanism we have used is part of the Haven On Demand offering from HPE. That's a platform which provides a wide range of capabilities based on machine learning. With the right data set and the right formulas, the program can pick out the patterns you have within the data.
"It produces a model which can then be used to calculate results - in this instance, the probability of a football score."
How much data have you used to predict these scores?
Alex Gorohovski, software engineer and data scientist - "We've gone back over three seasons of results in England, and looked at seasons in Spain, Italy and Germany when Mourinho, Guardiola or Conte were working there. In total we've analysed 218 features. We've looked at goals for and against, shots on target, wins, losses, draws, yellow cards, red cards… "We then used the system to find the algorithm that generated the best model for predicting future results between the teams involved."
What are the challenges associated with predicting football scores?
Lahav - "We were surprised with how well the prediction worked out for the Manchester derby. And hopefully the following predictions will be just as successful. But that's not guaranteed because there are a huge number of parameters in team sports.
"In individual sports such as golf and tennis, the main issue is the skill set of each player and the history between them. It works differently in football, where each individual doesn't affect the outcome as much as they would in say golf or tennis. Different players, tactics and environments and so on all have an effect. That's why we tried to use as much data as we could to feed into the system."
Where else is the predictor used?
Lahav - "These predictors are used in a lot of areas outside of sport. For instance, they can be used to predict sales or user behaviour. Wherever you have tables of data which relate to an end result, and you would like to predict future circumstances, you can use a similar algorithm and method around that.
"We're using some algorithms which are generic but we've created a process around those which is rather unique, the way we're handling the data, preparing it then measuring it and optimising around it."
As with all prediction systems, the proof will be in the final outcomes. Keep an eye on HPE's predictions through the season and follow the results with Sky Sports.
Find out more about Hewlett Packard Enterprise data solutions.
