Google Cloud Platform goes 8 for 8 in World Cup predictions
Wednesday, July 2, 2014
At Google I/O, we talked about how you can use Google Cloud Platform to analyze large datasets, build statistical models and use machine learning to predict the outcome of future events. And we demonstrated the technology by analyzing the World Cup.
As much as it broke the hearts of those of us rooting for Team USA, we got every game in the Round of 16 right. World Cup 2010 had Paul the Octopus. In 2014, we’re bringing Cloud Platform to the table.
If you don’t believe our record, you can watch Google Developer Advocate Felipe Hoffa (@felipehoffa) and BigQuery engineer Jordan Tigani (+JordanTigani) predict the outcomes at Google I/O on June 26. (Jump to minute 25 if you want to go right to the predictions).
Our model was built using touch-by-touch data from Opta covering multiple seasons of professional soccer leagues as well as the group rounds of the World Cup. Using data from the many leagues where World Cup players play, we were able to examine how behavior in previous games predicted performance in subsequent games. Our model also relied on two other data sources: (1) a power ranking that was built by BigQuery engineer Jordan Tigani and (2) a subjective measure of team support based on enthusiasm and the number of fans who had traveled to Brazil. This took the place of a simple “home team advantage” in the model.
We used Google Cloud Dataflow to ingest the data, Google BigQuery to build the derived features, iPython and Pandas to do the modeling, and Google Compute Engine to crunch the data.
Now, we know we’re not the only ones to get this round right. The Round of 16 didn’t feature any major upsets. However, we’re happy to show off how Cloud Platform can be used for doing machine learning and predictive analytics.
Other models use a poisson regression to calculate goal scores, but we found that a logistic regression to predict the winner rather than final score provided a more accurate prediction of future games.
And, now what you’re waiting for… We updated our model with the touch-by-touch data from the Round of 16 to predict the results of the quarterfinals. Here you go:
We’ll check in after the games to let you know how we’re doing. And, we’ll give you a deeper look at our methodology as well.
-Posted by Benjamin Bechtolsheim, Product Marketing Manager
As much as it broke the hearts of those of us rooting for Team USA, we got every game in the Round of 16 right. World Cup 2010 had Paul the Octopus. In 2014, we’re bringing Cloud Platform to the table.
If you don’t believe our record, you can watch Google Developer Advocate Felipe Hoffa (@felipehoffa) and BigQuery engineer Jordan Tigani (+JordanTigani) predict the outcomes at Google I/O on June 26. (Jump to minute 25 if you want to go right to the predictions).
Our model was built using touch-by-touch data from Opta covering multiple seasons of professional soccer leagues as well as the group rounds of the World Cup. Using data from the many leagues where World Cup players play, we were able to examine how behavior in previous games predicted performance in subsequent games. Our model also relied on two other data sources: (1) a power ranking that was built by BigQuery engineer Jordan Tigani and (2) a subjective measure of team support based on enthusiasm and the number of fans who had traveled to Brazil. This took the place of a simple “home team advantage” in the model.
We used Google Cloud Dataflow to ingest the data, Google BigQuery to build the derived features, iPython and Pandas to do the modeling, and Google Compute Engine to crunch the data.
Now, we know we’re not the only ones to get this round right. The Round of 16 didn’t feature any major upsets. However, we’re happy to show off how Cloud Platform can be used for doing machine learning and predictive analytics.
Other models use a poisson regression to calculate goal scores, but we found that a logistic regression to predict the winner rather than final score provided a more accurate prediction of future games.
And, now what you’re waiting for… We updated our model with the touch-by-touch data from the Round of 16 to predict the results of the quarterfinals. Here you go:
- Brazil vs. Colombia: Brazil (71%)
- France vs. Germany: France (69%)
- Netherlands vs. Costa Rica: Netherlands (68%)
- Argentina vs. Belgium: Argentina (81%)
We’ll check in after the games to let you know how we’re doing. And, we’ll give you a deeper look at our methodology as well.
-Posted by Benjamin Bechtolsheim, Product Marketing Manager