The Washington Capitals hold a commanding three-games-to-one lead over the Vegas Golden Knights in the Stanley Cup Final, but one AI-powered hockey analytics platform has deemed Vegas to be a clear favorite in Thursday’s Game 5.
Iceberg harnesses computer vision and artificial intelligence to automate the collection and analysis of advanced hockey data—and Iceberg’s system tabs the Golden Knights as 55.8 percent favorites. The burgeoning sports analytics company, which is intent on starting an intellectual hockey revolution, has 50 clients across 15 leagues and four continents. The New York Islanders are the only publicly identified NHL team to use the service.
CEO Alex Martynov, who founded the company after working as a bank analyst, said one factor was that Vegas would be back on home ice and another was that the margin of Washington’s 6-2 win in Game 4 was deceptive.
Iceberg catalogs one million data points per game, including each player’s movement 10 times per second. It then uses an NVIDIA GPU to apply machine learning algorithms to slice entire hockey games into tiny segments of action. Drawing on a database of more than 10,000 games, Iceberg’s five-person data science team has used AI techniques to determine causal links between actions and goals and then to assign each play a value.
On-ice actions lead to a number of expected goals (xG) that a team would score each game. In Game 4, Washington was only expected to score 3.3 goals compared to 3.02 for Vegas, a margin that suggested the run of play was closer than the final tally.
Iceberg began to project winners before Game 2 of the Stanley Cup. Here’s how the company has fared in those so far.
Game 2: Iceberg favored the Capitals at 54.4 percent. The Capitals won 3-2.
Game 3: Iceberg favored the Golden Knights at 52.7 percent. The Capitals won 3-1.
Game 4: Iceberg favored the Capitals at 61.3 percent. The Capitals won 6-2.
Game 5: Iceberg favors the Golden Knights at 55.8 percent; game is Thursday night.
“If you think about the game, it can be broken into a chain of certain events,” Martynov said. “It’s a defender standing behind his goal, and he wants to start a rush. In this moment, there is a certain probability that, within five seconds from that particular moment, there would be a goal for that team.
“What we’ve done is analyzed those chain of events. We know he’s standing behind the goal line, so what decisions can he make at this moment. He can pass or skate with the puck and try to enter the offensive zone.”
That’s one basic example, but many of the metrics quickly become esoteric—that, of course, is the point. Some of Iceberg’s proprietary stats include danger save percentage (how often a goalie stops shots from the prime, ellipse-shaped scoring area in front of the net), controlled entries (how successfully a team retains possession of the puck while entering the offensive zone), reliability (how well a player can help protect a one-goal lead) and mental stability (how a goalie responds to allowing a goal). Most of this stems from conversations with the collaborating teams.
“They’re always challenging us—our clients, our coaches—‘Can you tell me if our goalie is stable and consistent enough? Or if he allows a first goal and gets nervous and allows three more in the next 10 minutes?’” Martynov said. “We’ve quantified that as well.”
Iceberg even computes a catch-call index that tries to account for all of a player’s contributions—baseball has tried doing the same with WPA (win probability added) and WAR (wins above replacement). Martynov hopes this index can help illuminate the contributions of unsung players who might be defensive stalwarts on the third line rather than among the scoring leaders.
“In hockey, specifically, it’s really hard to quantify value of a player that is not linked to offensive contributions,” he said. “That’s what we’ve tried to do: Take the defensive side of the game and try to quantify that.”
The system currently does not conduct its analysis in real-time. After each game, the footage gleaned from the company’s three cameras (one for each zone of the rink) is uploaded to the Microsoft Azure cloud. Martynov said his company is working to optimize the system, because that could be key to potential work with the NHL on its promised optical tracking system— which Iceberg will be vying for.
“Of course,” Martynov said. “That’s what we are working on right now.”