Building and Understanding our Expected Goals (xG) Models

Scraping

We begin by assembling the dataset used to fit and evaluate our xG models. In earlier versions of this workflow, we had to manually loop over every game in the seasons of interest, fetch each game’s play-by-play, and then stitch them together as demonstrated in the legacy code here. With the new update 0.4.0, this process is much simpler: we can retrieve a full season’s worth of play-by-plays with a single function call, and then combine multiple seasons into one modeling dataset.

# Load data.
gc_pbps_20222023 <- nhlscraper::gc_pbps(20222023)
gc_pbps_20232024 <- nhlscraper::gc_pbps(20232024)
gc_pbps_20242025 <- nhlscraper::gc_pbps(20242025)

# Aggregate data.
common_cols <- Reduce(
  intersect,
  list(
    names(gc_pbps_20222023),
    names(gc_pbps_20232024),
    names(gc_pbps_20242025)
  )
)
gc_pbps_20222025 <- rbind(
  gc_pbps_20222023[common_cols], 
  gc_pbps_20232024[common_cols], 
  gc_pbps_20242025[common_cols]
)

Cleaning

Next, we prepare the play-by-play data for modeling by resolving a number of quirks and inconsistencies in the raw feed. For each event, we attach basic context such as whether the team is home or away, split the game ID into season, game type, and game number, and convert period/time information into continuous seconds elapsed in the game.

# Flag home/away.
gc_pbps_20222025_is_home_flagged         <- 
  nhlscraper::flag_is_home(gc_pbps_20222025)

# Strip game ID.
gc_pbps_20222025_game_id_stripped        <- 
  nhlscraper::strip_game_id(gc_pbps_20222025_is_home_flagged)

# Strip time and period.
gc_pbps_20222025_time_period_stripped    <- 
  nhlscraper::strip_time_period(gc_pbps_20222025_game_id_stripped)

We then derive the key hockey and game-state features that drive xG: strength situation (empty net status, skater counts, man-advantage differential, and strength state labels), rebound and rush indicators, and shot-volume measures such as goals, shots on goal, Fenwick (USAT), and Corsi (SAT).

# Strip situation code.
gc_pbps_20222025_situation_code_stripped <- 
  nhlscraper::strip_situation_code(gc_pbps_20222025_time_period_stripped)

# Flag rebound shot attempts.
gc_pbps_20222025_is_rebound_flagged      <- 
  nhlscraper::flag_is_rebound(gc_pbps_20222025_situation_code_stripped)

# Flag rush shot attempts.
gc_pbps_20222025_is_rush_flagged         <- 
  nhlscraper::flag_is_rush(gc_pbps_20222025_is_rebound_flagged)

# Count goals, SOG, Fenwick, and Corsi.
gc_pbps_20222025_goals_shots_counted     <- 
  nhlscraper::count_goals_shots(gc_pbps_20222025_is_rush_flagged)

Finally, we normalize coordinates so that all shots are taken toward the +x direction and compute the Euclidean distance and angle to the net, restricting the dataset to non-shootout, non-penalty-shot attempts.

# Normalize coordinates to +x.
gc_pbps_20222025_coordinates_normalized  <- 
  nhlscraper::normalize_coordinates(gc_pbps_20222025_goals_shots_counted)

# Calculate distance.
gc_pbps_20222025_distance_calculated     <- 
  nhlscraper::calculate_distance(gc_pbps_20222025_coordinates_normalized)

# Calculate angle.
gc_pbps_20222025_angle_calculated        <- 
  nhlscraper::calculate_angle(gc_pbps_20222025_distance_calculated)

# Keep only shots.
gc_shots_20222025 <- gc_pbps_20222025_angle_calculated[
  gc_pbps_20222025_angle_calculated$typeDescKey %in% 
    c('goal', 'shot-on-goal', 'missed-shot', 'blocked-shot'),
]

# Remove shootouts and penalty shots.
gc_shots_20222025_final <- gc_shots_20222025[
  !(gc_shots_20222025$situationCode %in% c('0101', '1010')),
]

# Indicate goal or not.
gc_shots_20222025_final$isGoal <- as.integer(
  gc_shots_20222025_final$typeDescKey == 'goal'
)

Baseline: xG_v1

The first model, xG_v1, is a baseline logistic regression for shot success. The response is isGoal (1 if the shot is a goal, 0 otherwise), and the predictors are distance (Euclidean distance from the shooter to the net), angle (shot angle relative to the center of the net), isEmptyNetAgainst (whether the opposing goalie has been pulled), and strengthState (game state at the time of the shot, such as even-strength, power-play, or penalty-kill).

# Build xG model version 1.
xG_v1 <- glm(
  isGoal ~
    distance +
    angle +
    isEmptyNetAgainst +
    strengthState,
  family = binomial,
  data   = gc_shots_20222025_final
)

# Summarize model 1.
summary(xG_v1)

On the log-odds scale, both distance and angle have negative coefficients: as you move farther from the net or shoot from a sharper angle, the probability of scoring decreases. The strong positive coefficient on isEmptyNetAgainstTRUE reflects how much easier it is to score into an empty net. Relative to even strength, both power-play and penalty-kill situations have positive effects, indicating higher conversion rates for shots taken in those states, conditional on a shot occurring.

Extended: xG_v2

The second model, xG_v2, extends the baseline specification by adding two play-context features: isRebound and isRush. The response is still isGoal, and we retain all the predictors from xG_v1 (distance, angle, isEmptyNetAgainst, and strengthState) while allowing the model to account for whether the shot is a rebound (fired shortly after a previous shot without a change of possession) or a rush chance (taken quickly following a transition up ice). The exact definitions are defined in the glossary and here.

# Build xG model version 2.
xG_v2 <- glm(
  isGoal ~
    distance +
    angle +
    isEmptyNetAgainst +
    strengthState +
    isRebound +
    isRush,
  family = binomial,
  data   = gc_shots_20222025_final
)

# Summarize model 2.
summary(xG_v2)

Compared to xG_v1, the coefficients for distance, angle, and the strength-related variables are broadly similar, but the model captures additional structure in how certain shot types perform. Rebound shots (isReboundTRUE) have a strongly positive and highly significant coefficient, reflecting the fact that rebounds, on average, are much more dangerous than non-rebound attempts once distance and angle are controlled for. Rush shots (isRushTRUE) have a slightly negative coefficient with only marginal statistical significance at conventional levels. This suggests that, conditional on location and other covariates, rush shots are not systematically more (or may even be slightly less) efficient than non-rush shots in this sample, even though they may intuitively feel more dangerous. The addition of isRebound and isRush reduces the residual deviance compared to xG_v1, indicating a modest but meaningful improvement in model fit while preserving the core spatial and game-state effects.

Contextual: xG_v3

The third model, xG_v3, builds on xG_v2 by adding a simple game-context variable: goalDifferential. As before, the response is isGoal, and we include all of the spatial and play-context predictors from the previous models (distance, angle, isEmptyNetAgainst, strengthState, isRebound, and isRush). The new term, goalDifferential, captures the score state from the shooting team’s perspective at the time of the shot (for example, leading vs. trailing).

# Build xG model version 3.
xG_v3 <- glm(
  isGoal ~
    distance +
    angle +
    isEmptyNetAgainst +
    strengthState +
    isRebound +
    isRush +
    goalDifferential,
  family = binomial,
  data   = gc_shots_20222025_final
)

# Summarize model 3.
summary(xG_v3)

Most of the core coefficients are very similar to those in xG_v2. Closer, more central shots (distance, angle) are still substantially more dangerous, empty-net attempts remain extremely likely to result in goals, and both penalty-kill and power-play states continue to show elevated finishing rates relative to five-on-five. Rebound shots retain a strong positive effect, while rush attempts again show a small negative coefficient with marginal statistical significance. The new goalDifferential term is positive and highly significant, indicating that, conditional on location and other covariates, shots taken when the shooting team is further ahead on the scoreboard convert at slightly higher rates than those taken when the game is tied or the team is trailing. This effect is modest in magnitude compared to the big spatial and empty-net effects, but it does capture additional structure in how game context influences finishing. The reduction in residual deviance relative to xG_v2 is incremental but consistent with an overall improvement in fit.

Shot Locations

The first pair of plots shows all shot attempts for one team, normalized so that they always attack to the right. Marker shape encodes the outcome (goal, shot on goal, missed, blocked) and color encodes the shot’s xG bin, from low-danger attempts in dark blue to the most dangerous chances in bright red.

# Plot shot locations for Game 7 Stanley Cup Finals 2025.
ig_game_shot_locations(
  game  = 2023030417, 
  model = 1, 
  team  = 'H'
)
ig_game_shot_locations(
  game  = 2023030417, 
  model = 1, 
  team  = 'A'
)

Looking at the home view, we can quickly pick out the team’s preferred shooting areas. In the example game, many of the Panthers’ attempts cluster around the slot and net-front, with several high-xG red markers just off the crease. That pattern is consistent with a team that frequently gets inside position, generates tips and rebounds, and is comfortable attacking through the middle of the ice. Switching to the away perspective, the Oilers’ shot map looks different. There are still dangerous chances around the crease, but we also see a higher volume of lower-xG shots from the outside such as point wristers, sharp-angle attempts, or quick shots off the rush that never quite reach the interior. This kind of map is a nice way to talk about “shot quality vs. shot volume”: a team may outshoot the opponent in raw attempts, but if most of those are blue markers from the perimeter, the expected goals will tell a more balanced story.

Cumulative xG

The second visualization shows cumulative xG over seconds elapsed in game for both teams at once. The x-axis runs from 0 to the end of regulation, with tick marks every 300 seconds (five minutes). The y-axis tracks the running sum of xG, so each step up corresponds to a new scoring chance; flat stretches indicate long periods without meaningful offense.

# Plot cumulative xG for Game 7 Stanley Cup Finals 2025.
ig_game_cumulative_expected_goals(
  game  = 2023030417, 
  model = 1
)

In our example, the red line (Florida) and blue line (Edmonton) track closely for much of the night. Early on, both teams climb in near lockstep, suggesting a fairly even trade of chances. Mid-game, Florida opens a small xG gap with a series of higher-quality looks, visible as a steeper red slope around the 1,000-1,200 second mark. Edmonton answers back later, narrowing the gap with their own sustained push. By the end of regulation the lines finish at similar heights, with Florida holding a modest xG edge. This is the classic “deserve-to-win-o-meter” view: rather than arguing from shots or goals alone, we can say that Florida generated slightly better chances overall, but the game was competitive throughout.