AI Canada

AI-Powered Hockey Analytics: A Game Changer

When the NBA’s Golden State Warriors decided to favour three-pointers over two-point shots in 2016, the winning strategy sent a shockwave through professional basketball. This was a “data-driven decision” based on higher score probability, explains Alex Martynov.

Analytics are all the rage in professional sports. The concept can be traced back to American statistician Bill James, who introduced his “Sabermetrics” method for in-game baseball analysis in the 1970s. When the NBA’s Golden State Warriors decided to favour three-pointers over two-point shots in 2016, the winning strategy sent a shockwave through professional basketball. This was a “data-driven decision” based on higher score probability, explains Alex Martynov.

Martynov is the 24-year old founder of ICEBERG, a Canadian startup using AI algorithms in sports analytics with a focus on ice hockey. Three years ago, Alex shared the idea of an AI sports analytics company with his investor father, who helped him kickstart the idea with $25,000. Not much, but it was enough for Alex to gather programmer friends in Toronto and Moscow and put together a working prototype.

ICEBERG installs a set of three FLIR thermal cameras around the rink before the start of each game. The video has lower resolution than an iPhone recording, but provides the constant full-ice view the company’s algorithms require, as broadcast feeds typically leave 50 percent or more of the ice surface and players out of frame.

Animated GIF-downsized_large (4).gif
Usually, there are no restrictions on filming hockey games for analytics use, and the process is taken care of with the help of each team’s video coach. Last year however ICEBERG was not allowed to film Czech Republic games at the World Championship, as video rights belonged to the International Ice Hockey Federation (IIHF).
Animated GIF-downsized_large (5)
There are still bugs to work out: ICEBERG’s computer vision algorithms struggled when team Kazakhstan wore gold numbers on white jerseys and team Switzerland wore dark jerseys with black numbers.

Artificial neural networks are trained to recognize all moving entities on the ice surface: 12 players grouped by jersey color, on-ice officials, and a small black puck that can reach speeds of 160 kph (100 mph). Computer vision algorithms previously trained on a dataset containing 10,000 variations of numbers from all angles can identify each player by jersey number.

By tracking player and puck coordinates 10 times per second, a sixty-minute game will generate one million data points. The algorithm matches individual player coordinates with those of the puck to record their passes, body checks, giveaways and takeaways, shots, and goals. Typically, about 7-9 percent of all shots result in goals, and variance here predicts higher or lower goal probability.

ICEBERG’s AI tracks a total of 500 different metrics which correspond to player and team behavior, and the company sometimes finds statistical nuances that are counterintuitive to hardcore fans or watchful coaches.

In a match between favorite Canada and underdog Switzerland, Iceberg’s AI found that Switzerland skated 1.7 more kilometers and were 5-10 centimeters closer to the puck in micro-episodes. The Swiss also had longer puck possession and generated 2.48 more expected goals (xG). Switzerland’s superior metrics should deliver a win seven times out of ten. But Canada won the game 3-0. Why?

viewfile.jpeg

viewfile2.jpeg

Martynov explains that “about 40 percent of all game outcomes is luck, but the other 60 percent can be predicted, which is what we are trying to do — predicting what isn’t random. Our clients can play five games and lose five times in a row, but data will show that they could’ve won every time. The coach will call our analyst, and we tell them, ‘calm down, it’s just the variation, you will get back to the mean, if you continue playing like this you will win five games in a row’.”

ICEBERG uses NVIDIA’s GPU and marketing expertise and Microsoft Azure’s cloud storage. The company also participates in NVIDIA’s Inception Program.

Portal subscription fee ranges from US$400 – $800 per game. If a team plays 60 games in a season that’s approximately US$30,000. Clients receive a report the morning after each game and can access detailed game numbers from the portal. ICEBERG also has on-call analysts to answer clients’ questions.

Finding clients can be a long process of convincing the coach, the manager, and the owner. ICEBERG’s deal with Austria’s Red Bull Salzburg required four months of negotiation. “There are coaches who are confused, asking ‘why do I need this?’” says Martynov. “We are not trying to replace the coach or the manager, but give teams an edge. It makes hockey more intellectual.”

There are also cases like Swedish teams Växjö Lakers and Färjestad BK, who signed contracts in five minutes. The competitive edge of data analytics is too good to be ignored.

“Currently, We have a market share of 5-7% of global professional hockey teams. But it’s not moving as fast as I would like,” says Martynov, “We want to go into the soccer market after this. If you get two percent of the soccer market, that’s approximately the same as the entire hockey market. We started in the niche market, but hockey is also a very complicated sport where players skate fast, collide often, change every minute, not to mention the puck is very small. Technically, it’s easy to downgrade from hockey into other sports.”


Journalist: Meghan Han | Editor: Michael Sarazen

0 comments on “AI-Powered Hockey Analytics: A Game Changer

Leave a Reply

Your email address will not be published.

%d bloggers like this: