Compiled NHL datasets from fastRhockey, built from raw JSON in fastRhockey-nhl-raw.
graph LR;
A[fastRhockey-nhl-raw]-->B[fastRhockey-nhl-data];
B[fastRhockey-nhl-data]-->C1[nhl_pbp_full];
B[fastRhockey-nhl-data]-->C2[nhl_pbp_lite];
B[fastRhockey-nhl-data]-->C3[nhl_team_boxscores];
B[fastRhockey-nhl-data]-->C4[nhl_player_boxscores];
B[fastRhockey-nhl-data]-->C5[nhl_rosters];
B[fastRhockey-nhl-data]-->C6[nhl_schedules];
flowchart TB;
subgraph A[fastRhockey-nhl-raw];
direction TB;
A1[scripts/daily_nhl_scraper.sh]-->A2[R/scrape_nhl_raw.R];
end;
subgraph B[fastRhockey-nhl-data];
direction TB;
B1[scripts/daily_nhl_R_processor.sh]-->B2[R/nhl_data_creation.R];
end;
subgraph C[sportsdataverse Releases];
direction TB;
C1[nhl_pbp_full];
C2[nhl_pbp_lite];
C3[nhl_team_boxscores];
C4[nhl_player_boxscores];
C5[nhl_rosters];
C6[nhl_schedules];
end;
A-->B;
B-->C1;
B-->C2;
B-->C3;
B-->C4;
B-->C5;
B-->C6;
click C1 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_pbp_full" _blank;
click C2 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_pbp_lite" _blank;
click C3 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_team_boxscores" _blank;
click C4 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_player_boxscores" _blank;
click C5 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_rosters" _blank;
click C6 "https://github.com/sportsdataverse/sportsdataverse-data/releases/tag/nhl_schedules" _blank;
| Release tag | Content |
|---|---|
nhl_pbp_full |
NHL play-by-play data (full, includes shifts) |
nhl_pbp_lite |
NHL play-by-play data (lite, no shift CHANGE events) |
nhl_team_boxscores |
NHL team box scores |
nhl_player_boxscores |
NHL player box scores (skaters + goalies) |
nhl_rosters |
NHL rosters |
nhl_schedules |
NHL schedules |
fastRhockey-nhl-raw data repository (source: NHL API)
fastRhockey-nhl-data repository (source: NHL API)
fastRhockey-pwhl-raw data repository (source: HockeyTech API)
fastRhockey-pwhl-data repository (source: HockeyTech API)
fastRhockey-data legacy repository (archived; sources: NHL Stats API + PHF)
Part of the SportsDataverse.
This repository contains the expected goals (xG) model built on fastRhockey NHL play-by-play data. The model follows the same general approach as the hockeyR expected goals model (Morse 2022) but is retrained on a larger dataset spanning 16 NHL seasons (2010-11 through 2025-26) and uses an expanded set of five era groupings to capture rule changes over the longer time horizon.
Expected goals models attempt to quantify the probability of any given unblocked shot becoming a goal, based on contextual information about the shot: where it was taken, how it was taken, and what happened immediately prior. The resulting xG values are useful for evaluating player and team performance beyond raw goal counts.
The play-by-play data used to train this model comes from the
fastRhockey-nhl-data
repository, originally sourced via the
fastRhockey R package
(Gilani 2022). The data is stored as .rds files in
nhl/pbp/full/rds/.
suppressPackageStartupMessages({
library(dplyr)
library(tidyr)
library(stringr)
library(ggplot2)
library(readr)
library(purrr)
library(scales)
})
# Load all 16 seasons of PBP data
rds_dir <- "nhl/pbp/full/rds"
rds_files <- list.files(rds_dir, pattern = "\\.rds$", full.names = TRUE)
# Read all files into a list, then harmonise column types before binding.
# The old API (2010-2022) and new API (2023+) store some columns as different
# types (e.g. season: integer vs character, game_date: Date vs character).
# bind_rows handles integer↔double promotion automatically; we only need to
# fix truly incompatible pairs (character↔integer, Date↔character).
pbp_list <- map(rds_files, function(f) {
df <- readRDS(f)
# Fix character↔integer mismatches (old API stores as character, new as integer)
df$season <- as.character(df$season)
df$home_id <- as.character(df$home_id)
df$away_id <- as.character(df$away_id)
# Strip 'glue' subclass from description (new API) so bind_rows sees plain character
if ("description" %in% names(df)) df$description <- as.character(df$description)
# Fix Date↔character mismatch on game_date
if (is.character(df$game_date)) df$game_date <- as.Date(df$game_date)
# Coerce any remaining logical columns that clash with integer/numeric
# (empty_net, for example, is logical in some files and integer in others)
for (col in names(df)) {
if (is.logical(df[[col]])) df[[col]] <- as.integer(df[[col]])
}
df
})
pbp_all <- bind_rows(pbp_list)
rm(pbp_list) # free memory
gc() # reclaim ~10 GB from the list## used (Mb) gc trigger (Mb) max used (Mb)
## Ncells 12132176 648.0 19946810 1065.3 12403729 662.5
## Vcells 2110892063 16104.9 5647614053 43087.9 4826935529 36826.6
# --- Column harmonisation across seasons ---
# event_team (full name) and event_team_abbr exist in all seasons.
# home_name / away_name exist everywhere; create home_team / away_team aliases
# so that event_team == home_team comparisons work.
pbp_all <- pbp_all %>%
mutate(
home_team = home_name,
away_team = away_name
)
# Helper to pretty-print season codes ("20102011" -> "2010-2011")
format_season <- function(s) paste0(substr(s, 1, 4), "-", substr(s, 5, 8))
seasons <- sort(unique(pbp_all$season))
n_rows <- nrow(pbp_all)
n_seasons <- length(seasons)
cat(sprintf(
"Loaded %s play-by-play rows across %d seasons (%s through %s)\n",
format(n_rows, big.mark = ","), n_seasons,
format_season(min(seasons)), format_season(max(seasons))
))## Loaded 13,553,239 play-by-play rows across 16 seasons (2010-2011 through 2025-2026)
The NHL play-by-play data records the number of skaters on the ice for each team at the time of every event. As with most expected goals models, shots taken in different strength states need special handling.
# Define Fenwick events (unblocked shots only)
fenwick <- c("SHOT", "MISSED_SHOT", "GOAL")
# Investigate strength states of all Fenwick events
strength_summary <- pbp_all %>%
filter(event_type %in% fenwick & !is.na(strength_state)) %>%
count(strength_state, sort = TRUE) %>%
mutate(pct = round(100 * n / sum(n), 2))
knitr::kable(
head(strength_summary, 15),
col.names = c("Strength State", "Count", "Pct (%)"),
align = "lrr"
)| Strength State | Count | Pct (%) |
|---|---|---|
| 5v5 | 1275839 | 75.45 |
| 5v4 | 244117 | 14.44 |
| 4v5 | 54149 | 3.20 |
| 4v4 | 32300 | 1.91 |
| 6v5 | 27390 | 1.62 |
| 3v3 | 12748 | 0.75 |
| 0v0 | 11215 | 0.66 |
| 5v3 | 9766 | 0.58 |
| 5v6 | 9027 | 0.53 |
| 4v3 | 5482 | 0.32 |
| 6v4 | 4138 | 0.24 |
| 1v0 | 1576 | 0.09 |
| 3v4 | 818 | 0.05 |
| 4v6 | 717 | 0.04 |
| 3v5 | 589 | 0.03 |
The vast majority of Fenwick events occur at 5-on-5. For this model, a small number of events at unusual strength states (e.g. 6v3) and those with missing strength state information were removed from the training data. Penalty shots and shootout attempts were also excluded — shootouts are not assigned an xG value since they don’t count as official goals, and penalty shots were assigned a constant xG of ~0.32, reflecting the ~32% conversion rate across all penalty shot and shootout attempts in the dataset.
Events with missing critical features (shot location, shot type) were also removed, as were blocked shots. Blocked shots are excluded because the NHL records the location of the block, not the location of the shooter, making the distance and angle calculations unreliable.1
There are several excellent public xG models with readily available details:
-
Asmae Toumi and Dawson Sprigings created an even-strength model with rush chance and rebound identifiers. (Sprigings and Toumi 2015)
-
Matthew Barlowe’s public model covers all strengths and takes team strength as a variable. (Barlowe 2017)
-
Peter Tanner’s MoneyPuck.com model utilizes information from the events immediately prior to the shot. (Tanner 2016)
-
Josh and Luke Younggren of evolving-hockey.com made four separate models for different strength states. (Younggren and Younggren 2018)
All models agree that shot distance is the most important feature, with
shot angle close behind. The fastRhockey model uses features similar
to those listed above: distance and angle, rebound and rush chance
classifiers, information about the prior event, and a cross_ice_event
flag to indicate the goaltender had to move laterally across the crease.
Following hockeyR (Morse 2022) and in the spirit of the Younggren
models (Younggren and Younggren 2018), the fastRhockey xG model is
split into two separate models:
- 5-on-5 model — exclusively for 5-on-5 play
- Special teams model — for all other non-5v5 situations (power play, shorthanded, 4-on-4, 3-on-3, extra attacker, etc.)
The 4-on-4 and 3-on-3 situations are included in the special teams model rather than the 5v5 model because the increased open ice more closely resembles special teams play than the structured chaos of 5-on-5.
real_strengths <- c(
"5v5", "4v4", "3v3",
"5v4", "5v3", "4v3", "6v4", "6v5",
"4v5", "3v5", "3v4", "4v6", "5v6"
)
strength_goals <- pbp_all %>%
filter(strength_state %in% real_strengths & event_type %in% fenwick) %>%
# Remove empty-net shots for clean comparison
filter(str_detect(strength_state, "v6", negate = TRUE) &
strength_state != "6v3") %>%
group_by(strength_state) %>%
summarise(
sog = sum(event_type %in% c("SHOT", "GOAL")),
goals = sum(event_type == "GOAL"),
.groups = "drop"
) %>%
mutate(
sh_perc = round(100 * goals / sog, 2),
strength = case_when(
strength_state %in% c("5v5", "4v4", "3v3") ~ "Even Strength",
strength_state %in% c("3v4", "3v5", "4v5") ~ "Shorthanded",
strength_state %in% c("5v4", "5v3", "4v3", "6v4") ~ "Power Play",
TRUE ~ "Extra Attacker"
)
)
ggplot(strength_goals, aes(reorder(strength_state, -sh_perc), sh_perc)) +
geom_col(aes(fill = strength)) +
scale_fill_manual(values = c("#E69F00", "#999999", "#56B4E9", "#009E73")) +
theme_bw() +
guides(fill = guide_legend(ncol = 2)) +
theme(
legend.background = element_rect(color = "black"),
legend.position = c(.75, .75)
) +
labs(
x = NULL, y = "Shooting %", fill = NULL,
caption = "data from fastRhockey-nhl-data",
title = "Shooting % by game strength state",
subtitle = sprintf("%s through %s NHL seasons",
format_season(min(seasons)),
format_season(max(seasons)))
)
The special teams model includes two additional features to capture the
strength state: total_skaters_on (total skaters on ice, a proxy for
available space) and event_team_advantage (shooting team’s skater
count minus opposing team’s skater count). For example, a power play
shot at 5-on-3 has total_skaters_on = 8 and event_team_advantage =
2.
| Feature | Description |
|---|---|
shot_distance |
Distance from shooter to net (ft) |
shot_angle |
Angle between shooter and net (degrees) |
shot_type |
Wrist, Slap, Snap, Backhand, Tip-In, Deflected, Wrap-Around, Cradle |
rebound |
Binary: previous event was a Fenwick event within 2 seconds |
rush |
Binary: previous event was in NZ/DZ within 4 seconds |
last_event_type |
Type of the previous play-by-play event |
time_since_last |
Seconds elapsed since the previous event |
distance_from_last |
Distance from the previous event location (ft) |
cross_ice_event |
Binary: goalie had to move laterally across crease |
empty_net |
Binary: opposing net was empty |
last_x, last_y |
Coordinates of the previous event |
era_* |
Five era dummy variables (see below) |
| Feature | Description |
|---|---|
total_skaters_on |
Total skaters on ice for both teams |
event_team_advantage |
Shooting team skaters minus opposing team skaters |
Borrowing the era concept from nflfastR (Carl and Baldwin 2021) and
its expected points model, the fastRhockey model uses five era
groupings to account for changes in goal-scoring rates driven by NHL
rule changes:
| Era | Seasons | Rule Change |
|---|---|---|
era_2011_2013 |
2010-11 through 2012-13 | Pre-equipment reduction baseline |
era_2014_2018 |
2013-14 through 2017-18 | Goaltender leg pad size reduction |
era_2019_2021 |
2018-19 through 2020-21 | Goalie chest & arm protector reduction |
era_2022_2024 |
2021-22 through 2023-24 | Cross-checking penalty emphasis |
era_2025_on |
2024-25 and beyond | Latest rules and trends |
This is an expansion over the original hockeyR model’s four eras,
adding a fifth era to capture the most recent seasons as the dataset has
grown. Patrick Bacon’s expected goals model (Bacon 2020) also accounts
for rule changes, though Bacon trains on smaller rolling windows near
each season rather than using era dummy variables.
Once the final models were trained, feature importance plots were generated. As with every other model mentioned above, shot distance is the dominant feature.
For full implementation details, see R/build_xg_model.R in this
repository. The model was built with extreme gradient boosting using the
xgboost R package (Chen et al. 2022). Rachael Tatman provides a great
overview of how XGBoost models work
here.
(Tatman 2018)
The 16 seasons of data were randomly split into training and testing sets at the game level — all events from any given game are in the same set to prevent data leakage from sequential event features. A modified random grid search was used to tune hyperparameters with 5-fold cross validation.
The final model performance:
| Model | CV Log-loss | CV AUC |
|---|---|---|
| 5v5 | 0.2053 | 0.8322 |
| ST | 0.2567 | 0.8213 |
Both log-loss and AUC are between 0 and 1. A lower log-loss is better; a higher AUC is better. Gaurav Dembla provides great resources on interpreting log-loss (Dembla 2020a) and AUC (Dembla 2020b).
With the models finalized, they can be applied to every season in the dataset. To validate the model, we can look at how well expected goals track actual goals at the player and team levels.
# Load saved models and apply xG to the full dataset
model_meta <- readRDS("models/xg_model_meta.rds")
model_5v5 <- xgboost::xgb.load("models/xg_model_5v5.json")
model_st <- xgboost::xgb.load("models/xg_model_st.json")
# Shot type normalization — same mapping used in training
normalize_shot_type <- function(x) {
case_when(
x %in% c("Wrist Shot", "WRIST SHOT", "Wrist", "wrist") ~ "Wrist Shot",
x %in% c("Slap Shot", "SLAP SHOT", "Slap", "slap") ~ "Slap Shot",
x %in% c("Snap Shot", "SNAP SHOT", "Snap", "snap") ~ "Snap Shot",
x %in% c("Backhand", "BACKHAND", "backhand") ~ "Backhand",
x %in% c("Tip-In", "TIP-IN", "Tip In", "Tip", "tip-in") ~ "Tip-In",
x %in% c("Deflected", "DEFLECTED", "Deflection", "deflected") ~ "Deflected",
x %in% c("Wrap-Around", "WRAP-AROUND", "Wrap-around",
"Wraparound", "wrap-around") ~ "Wrap-around",
x %in% c("Cradle", "CRADLE", "cradle") ~ "Cradle",
x %in% c("Bat", "BAT", "bat", "Batted", "batted") ~ "Batted",
x %in% c("Between Legs", "BETWEEN LEGS", "between-legs") ~ "Between Legs",
x %in% c("Poke", "POKE", "poke") ~ "Poke",
TRUE ~ x
)
}
# Prepare features for xG prediction
pbp_xg <- pbp_all %>%
filter(event_type %in% fenwick & !is.na(strength_state)) %>%
filter(is.na(period_type) | period_type != "SHOOTOUT") %>%
filter(
!tolower(secondary_type) %in% c("penalty shot", "penalty-shot") |
is.na(secondary_type)
) %>%
group_by(game_id) %>%
arrange(event_id, .by_group = TRUE) %>%
mutate(
last_event_type = lag(event_type),
time_since_last = game_seconds - lag(game_seconds),
last_x = lag(x),
last_y = lag(y),
distance_from_last = round(sqrt(((y - last_y)^2) + ((x - last_x)^2)), 1),
event_zone = case_when(
x >= -25 & x <= 25 ~ "NZ",
(x_fixed < -25 & event_team == home_team) |
(x_fixed > 25 & event_team == away_team) ~ "DZ",
(x_fixed > 25 & event_team == home_team) |
(x_fixed < -25 & event_team == away_team) ~ "OZ"
),
last_event_zone = lag(event_zone),
rebound = ifelse(
last_event_type %in% fenwick & time_since_last <= 2, 1, 0
),
rush = ifelse(
last_event_zone %in% c("NZ", "DZ") & time_since_last <= 4, 1, 0
),
cross_ice_event = ifelse(
last_event_zone == "OZ" &
((lag(y) > 3 & y < -3) | (lag(y) < -3 & y > 3)) &
time_since_last <= 2, 1, 0
),
shot_type = normalize_shot_type(secondary_type),
empty_net = ifelse(is.na(empty_net) | empty_net == FALSE, 0, 1),
goal = ifelse(event_type == "GOAL", 1, 0),
event_team_skaters = ifelse(
event_team == home_team, home_skaters, away_skaters
),
opponent_team_skaters = ifelse(
event_team == home_team, away_skaters, home_skaters
),
total_skaters_on = event_team_skaters + opponent_team_skaters,
event_team_advantage = event_team_skaters - opponent_team_skaters
) %>%
ungroup() %>%
mutate(
# Era dummies (season format: "20102011", "20112012", etc.)
era_2011_2013 = ifelse(
season %in% c("20102011", "20112012", "20122013"), 1, 0
),
era_2014_2018 = ifelse(
season %in% c("20132014", "20142015", "20152016",
"20162017", "20172018"), 1, 0
),
era_2019_2021 = ifelse(
season %in% c("20182019", "20192020", "20202021"), 1, 0
),
era_2022_2024 = ifelse(
season %in% c("20212022", "20222023", "20232024"), 1, 0
),
era_2025_on = ifelse(
season %in% c("20242025", "20252026"), 1, 0
)
) %>%
# Remove rows with missing critical features
filter(
!is.na(shot_distance) & !is.na(shot_angle) &
!is.na(shot_type) & !is.na(last_event_type)
) %>%
# Keep only valid last_event_type values (same filter as training)
filter(last_event_type %in% c(
"FACEOFF", "GIVEAWAY", "TAKEAWAY", "BLOCKED_SHOT", "HIT",
"MISSED_SHOT", "SHOT", "STOP", "PENALTY", "GOAL"
)) %>%
# Drop rows with NA in any numeric feature (same as training na.omit)
filter(
!is.na(time_since_last) & !is.na(distance_from_last) &
!is.na(last_x) & !is.na(last_y) &
!is.na(rebound) & !is.na(rush) & !is.na(cross_ice_event)
)
# Get expected feature names from saved metadata
feats_5v5 <- model_meta$xg_feature_names_5v5
feats_st <- model_meta$xg_feature_names_st
# Valid strength states for ST model
st_strengths <- c(
"5v4", "5v3", "6v5", "6v4", "4v4", "4v3",
"3v3", "4v5", "3v5", "5v6", "4v6", "3v4"
)
# Split into 5v5 and ST datasets
data_5v5 <- pbp_xg %>% filter(strength_state == "5v5")
data_st <- pbp_xg %>% filter(strength_state %in% st_strengths)
# One-hot encode using pivot_wider + clean_names (same as training)
encode_and_predict <- function(df, model, feature_names) {
feat_df <- df %>%
mutate(row_id = row_number()) %>%
select(
row_id,
starts_with("era_"),
shot_distance, shot_angle, rebound, rush,
time_since_last, distance_from_last, cross_ice_event,
empty_net, last_x, last_y,
shot_type, last_event_type,
any_of(c("total_skaters_on", "event_team_advantage"))
) %>%
mutate(type_value = 1, last_value = 1) %>%
pivot_wider(
names_from = shot_type,
values_from = type_value,
values_fill = 0
) %>%
pivot_wider(
names_from = last_event_type,
values_from = last_value,
values_fill = 0,
names_prefix = "last_"
) %>%
clean_names() %>%
arrange(row_id) %>%
select(-row_id)
# Ensure every expected feature column exists (fill missing with 0)
for (f in feature_names) {
if (!f %in% names(feat_df)) feat_df[[f]] <- 0
}
mat <- as.matrix(feat_df[, feature_names])
dmat <- xgboost::xgb.DMatrix(mat)
predict(model, dmat)
}
data_5v5$xg <- encode_and_predict(data_5v5, model_5v5, feats_5v5)
data_st$xg <- encode_and_predict(data_st, model_st, feats_st)
# Combine back together
pbp <- bind_rows(data_5v5, data_st)
cat(sprintf(
"Applied xG to %s Fenwick events (%s 5v5 + %s ST)\n",
format(nrow(pbp), big.mark = ","),
format(nrow(data_5v5), big.mark = ","),
format(nrow(data_st), big.mark = ",")
))## Applied xG to 1,274,786 Fenwick events (967,500 5v5 + 307,286 ST)
The relationship between player-level expected goals and actual goals scored provides a strong validation of the model.
player_goals <- pbp %>%
filter(season_type %in% c("R", "REG") & period < 5) %>%
filter(event_type %in% fenwick) %>%
group_by(player = event_player_1_name, season) %>%
summarise(
fenwick = n(),
goals = sum(event_type == "GOAL", na.rm = TRUE),
xg = sum(xg, na.rm = TRUE),
.groups = "drop"
) %>%
mutate(gax = goals - xg) %>%
filter(goals > 0)
goal_mod <- lm(goals ~ xg, data = player_goals)
rsq_player <- summary(goal_mod)$r.sq %>% round(3)
ggplot(player_goals, aes(xg, goals)) +
geom_abline(slope = 1, intercept = 0, linetype = "dashed") +
geom_point(alpha = .2) +
geom_smooth(method = lm, color = "red", se = TRUE) +
annotate("text", x = 3, y = 22, label = paste("R\u00b2:", rsq_player),
color = "red", size = 5) +
scale_x_continuous(breaks = seq(0, 60, 10)) +
scale_y_continuous(breaks = seq(0, 60, 10)) +
theme_bw() +
labs(
x = "Expected Goals", y = "Observed Goals",
caption = "data from fastRhockey-nhl-data",
title = "Player season goal totals vs expectation",
subtitle = sprintf(
"%s through %s seasons | all situations",
format_season(min(seasons)), format_season(max(seasons))
)
)
Performing the same analysis at the team level yields a strong correlation as well.
team_goals <- pbp %>%
filter(season_type %in% c("R", "REG") & period < 5) %>%
filter(event_type %in% fenwick) %>%
group_by(team = event_team_abbr, season) %>%
summarise(
fenwick = n(),
goals = sum(event_type == "GOAL", na.rm = TRUE),
xg = sum(xg, na.rm = TRUE),
.groups = "drop"
) %>%
mutate(gax = goals - xg)
team_goal_mod <- lm(goals ~ xg, data = team_goals)
rsq_team <- summary(team_goal_mod)$r.sq %>% round(3)
ggplot(team_goals, aes(xg, goals)) +
geom_abline(slope = 1, intercept = 0, linetype = "dashed") +
geom_point(alpha = .4) +
geom_smooth(method = lm, color = "red", se = TRUE) +
annotate("text", x = min(team_goals$xg) + 10, y = max(team_goals$goals) - 10,
label = paste("R\u00b2:", rsq_team), color = "red", size = 5) +
theme_bw() +
labs(
x = "Expected Goals", y = "Observed Goals",
caption = "data from fastRhockey-nhl-data",
title = "Team season goal totals vs expectation",
subtitle = sprintf(
"%s through %s seasons | all situations",
format_season(min(seasons)), format_season(max(seasons))
)
)
Borrowing from nflfastR (Carl and Baldwin 2021) (and nflscrapR
(Yurko, Ventura, and Horowitz 2018)), a calibration plot shows how well
the model’s predicted probabilities match observed goal rates across
different probability bins.
xg_bin_plot <- pbp %>%
filter(!is.na(xg) & event_type %in% fenwick) %>%
filter(season_type %in% c("R", "REG") & period < 5) %>%
mutate(xg_bin = round(xg / 0.05) * 0.05) %>%
group_by(xg_bin) %>%
summarise(
fenwick = n(),
goals = sum(event_type == "GOAL"),
.groups = "drop"
) %>%
mutate(obs_goal_prob = goals / fenwick)
ggplot(xg_bin_plot, aes(xg_bin, obs_goal_prob)) +
geom_abline(slope = 1, intercept = 0, linetype = "dashed") +
geom_point(aes(size = fenwick)) +
scale_size_continuous(
breaks = c(10000, 20000, 40000, 60000, 80000),
labels = c("10k", "20k", "40k", "60k", "80k")
) +
geom_smooth() +
coord_equal() +
scale_x_continuous(limits = c(0, 1)) +
scale_y_continuous(limits = c(0, 1)) +
theme_bw() +
labs(
x = "Estimated goal probability", y = "Observed goal probability",
caption = "data from fastRhockey-nhl-data",
size = "Unblocked\nShots",
subtitle = sprintf(
"%s through %s seasons | all situations",
format_season(min(seasons)), format_season(max(seasons))
),
title = "xG Calibration"
)
Checking the total expected goal count against observed goals by season shows how well the model tracks actual scoring rates. A ratio near 1.0 indicates the model is well-calibrated overall.
xg_per_goal <- pbp %>%
filter(season_type %in% c("R", "REG") & period < 5) %>%
group_by(season) %>%
summarise(
goals = sum(event_type == "GOAL"),
xg = sum(xg, na.rm = TRUE),
.groups = "drop"
) %>%
mutate(
xg_per_g = round(xg / goals, 3),
xg = round(xg, 1),
season = format_season(season)
) %>%
select(season, goals, xg, xg_per_g)
knitr::kable(
xg_per_goal,
col.names = c("Season", "Goals", "xG", "xG per Goal"),
align = "lrrr"
)| Season | Goals | xG | xG per Goal |
|---|---|---|---|
| 2010-2011 | 6498 | 6703.3 | 1.032 |
| 2011-2012 | 6473 | 6612.9 | 1.022 |
| 2012-2013 | 3765 | 3804.7 | 1.011 |
| 2013-2014 | 6494 | 6689.7 | 1.030 |
| 2014-2015 | 6474 | 6767.5 | 1.045 |
| 2015-2016 | 6484 | 6796.7 | 1.048 |
| 2016-2017 | 6630 | 6855.7 | 1.034 |
| 2017-2018 | 7374 | 7429.1 | 1.007 |
| 2018-2019 | 7500 | 7544.0 | 1.006 |
| 2019-2020 | 6424 | 6476.9 | 1.008 |
| 2020-2021 | 4948 | 4970.8 | 1.005 |
| 2021-2022 | 8053 | 7870.8 | 0.977 |
| 2022-2023 | 8001 | 7831.1 | 0.979 |
| 2023-2024 | 4574 | 5891.6 | 1.288 |
| 2024-2025 | 7421 | 8439.0 | 1.137 |
| 2025-2026 | 3779 | 4487.4 | 1.187 |
overall_ratio <- round(sum(xg_per_goal$xg) / sum(xg_per_goal$goals), 3)
cat(sprintf(
"\nOverall: %.3f expected goals per observed goal across all seasons\n",
overall_ratio
))##
## Overall: 1.042 expected goals per observed goal across all seasons
Penalty shots and shootout attempts are handled separately from the two xgboost models. All penalty shot and shootout attempts across the dataset were assigned a constant xG value reflecting the observed conversion rate:
ps_data <- pbp_all %>%
filter(
event_type %in% fenwick &
(tolower(secondary_type) %in% c("penalty shot", "penalty-shot") |
period_type %in% "SHOOTOUT")
)
ps_goals <- sum(ps_data$event_type == "GOAL")
ps_total <- nrow(ps_data)
ps_rate <- round(ps_goals / ps_total, 4)
cat(sprintf(
"Penalty shot / shootout conversion: %d goals on %s attempts = %.2f%%\n",
ps_goals, format(ps_total, big.mark = ","), 100 * ps_rate
))## Penalty shot / shootout conversion: 3614 goals on 11,286 attempts = 32.02%
The complete training code is available in R/build_xg_model.R. The
trained models are saved in models/ as:
xg_model_5v5.json— XGBoost model for 5-on-5 playxg_model_st.json— XGBoost model for special teamsxg_model_meta.rds— Metadata (penalty shot constant, feature lists, training date)
Cross-validation results are saved in data/:
cv_results_5v5_final.rds— 5v5 CV evaluation logcv_results_st_final.rds— ST CV evaluation log
Bacon, Patrick. 2020. “A New Expected Goal Model That Is Better Than Corsi at Predicting Future Goals,” December. https://topdownhockey.medium.com/a-new-expected-goal-model-that-is-better-than-corsi-at-predicting-future-goals-ecfa44dc84e9.
Barlowe, Matthew. 2017. “NHL Expected Goals Model,” September. https://rstudio-pubs-static.s3.amazonaws.com/311470_f6e88d4842da46e9941cc6547405a051.html.
Carl, Sebastian, and Ben Baldwin. 2021. “nflfastR: Functions to Efficiently Access NFL Play by Play Data.”
Chen, Tianqi, Tong He, Michael Benesty, Vadim Khotilovich, Yuan Tang, Hyunsu Cho, Kailong Chen, et al. 2022. “Xgboost: Extreme Gradient Boosting.” https://CRAN.R-project.org/package=xgboost.
Dembla, Gaurav. 2020a. “Intuition Behind Log-Loss Score,” November. https://towardsdatascience.com/intuition-behind-log-loss-score-4e0c9979680a.
———. 2020b. “Intuition Behind ROC-AUC Score,” December. https://towardsdatascience.com/intuition-behind-roc-auc-score-1456439d1f30.
Gilani, Saiem. 2022. “fastRhockey: Functions to Access Premier Hockey Federation and National Hockey League Play-by-Play Data.” https://github.com/sportsdataverse/fastRhockey.
Morse, Dan. 2022. “hockeyR Expected Goals Model.” https://github.com/danmorse314/hockeyR-models.
Sprigings, Dawson, and Asmae Toumi. 2015. “Expected Goals Are a Better Predictor of Future Scoring Than Corsi, Goals,” October. https://hockey-graphs.com/2015/10/01/expected-goals-are-a-better-predictor-of-future-scoring-than-corsi-goals/.
Tanner, Peter. 2016. “Shot Prediction Expected Goals Model.” https://moneypuck.com/about.htm.
Tatman, Rachael. 2018. “Machine Learning with XGBoost (in r),” July. https://www.kaggle.com/code/rtatman/machine-learning-with-xgboost-in-r/notebook.
Younggren, Josh, and Luke Younggren. 2018. “A New Expected Goals Model for Predicting Goals in the NHL,” June. https://evolving-hockey.com/blog/a-new-expected-goals-model-for-predicting-goals-in-the-nhl/.
Yurko, Ronald, Samuel Ventura, and Maksim Horowitz. 2018. “nflWAR: A Reproducible Method for Offensive Player Evaluation in Football.” https://doi.org/10.48550/ARXIV.1802.00998.