Expected Threat in Ice Hockey

In soccer, expected threat (xT) approach combines expected goal models (xG) with game dynamics using a stochastic model for the next ball event. Moving the ball to another location by dribbling or passing is modeled as a Markov chain. If a shot is taken, the scoring probability is computed using a xG model. The model parameters representing transition probabilities are estimated from event data and the model is used to compute an xT value for each field location, i.e., the probability of the team scoring before its opponent when it has the ball at the given point.

The xT modeling approach can be transferred to ice hockey analytics with only minor adjustments that represent certain game specific factors. First, the movement, shooting, and turnovers are modeled in a similar manner as in soccer. However, the boards of the ice hockey rink usually keep the puck in play when a shot is attempted but it does not lead into a goal. Therefore, the dynamics of the unsuccessful shots rebounded by the offensive or defensive team have to be included in the model as transitions between the shot location and the spot where the puck is rebounded. Furthermore, in ice hockey penalties regularly lead into 5 vs. 4 power play situations that have to be modeled separately. In practice, this leads to building two additional xT models — one for the power play and another for the short-handed team — whose parameters are estimated simultaneously. (The authors are aware that also other power play situations exist, but they can be modeled in exactly same manner as long as there are enough data.)

Project co-authored with Matti Honkanen (see also kiekkotilasto.fi).