wa-setpieces

Set-piece metrics for football (soccer) matches from Opta / Stats Perform F24 event-feed JSON exports (natively) and StatsBomb open-data exports (via an adapter, see below): penalties, kick-offs, free kicks, corners, throw-ins and goal kicks.

Given a match file, this package tags every set-piece restart, aggregates attempts/success rates by team and player, tracks pass end locations for delivery maps, and links each set piece to the shot or goal it produced (via the provider's own assist-chain data). It also covers, for corners and free kicks specifically: second-phase detection, Expected Threat (xT), pitch zones/thirds/channels, possession retention, and a benchmarked team/player rating — with pitch plots built on mplsoccer.

Full documentation, with a runnable plot gallery: https://waltzinganalytics.readthedocs.io

Install

pip install wa-setpieces

Or install from source:

git clone https://github.com/marclamberts/waltzinganalytics.git
cd waltzinganalytics
pip install -e .

Quickstart

from wa_setpieces import load_events, set_piece_summary

match = load_events("match.json")
summary = set_piece_summary(match.events)
print(summary)

              contestantId set_piece_type  attempts  successful  success_rate  shots  goals
cxb4hqite921i...      corner         2           1         0.500      1      0
cxb4hqite921i...   free_kick        12           9         0.750      0      0
cxb4hqite921i...   goal_kick         8           5         0.625      0      0
cxb4hqite921i...    kick_off         1           1         1.000      0      0
cxb4hqite921i...    throw_in        20          16         0.800      1      0
...

The whole pipeline in one call

Everything below this section — team/player counts, delivery locations, second-phase detection, retention, added value, the report, and the rating — is one function chain most people want run together. run_workflow runs that whole chain for you and hands back every table at once, instead of wiring five function calls together yourself:

from wa_setpieces import run_workflow, XTModel

model = XTModel.fit(match.events)         # optional -- unlocks added value + player rating
result = run_workflow(match.events, "corner", model=model)

result.summary        # attempts, success rate, shots, goals
result.deliveries      # start/end coordinates for a delivery map
result.second_phases   # cleared / first-phase shot / second-phase shot, per corner
result.retention        # still in possession ~8s later?
result.added_value      # xT added + resulting shot quality + goals, per delivery
result.report           # all of the above, rolled up per team
result.team_rating       # 0-100 benchmark score per team
result.player_rating     # delivery score / finishing score per player

Reach for the individual functions below directly when you only need one piece, want different parameters per step, or are combining several matches. run_workflow computes nothing new — it's a convenience wrapper, not a shortcut that skips anything.

Second phases, xT, zones, retention, added value and outcomes

from wa_setpieces import (
    second_phases, second_phase_summary,   # corner/free-kick second-phase shots
    retention_detail, retention_rate,      # possession retained N seconds later
    add_thirds, add_channels, add_zone_grid,  # pitch location tagging
    XTModel, set_piece_delivery_xt, set_piece_xt_summary,  # Expected Threat
    set_piece_added_value, set_piece_value_summary,  # xT + shot quality + goals, blended
    corner_report, free_kick_report,       # all of the above, merged into one table per team
    delivery_outcomes, outcome_summary,    # per-delivery outcome category, for a shot map
)

second_phases(match.events, "corner")           # per-corner: cleared / first-phase shot / second-phase shot
second_phase_summary(match.events, "free_kick") # per-team roll-up

retention_rate(match.events, "corner")          # per-team: % of corners where the ball is retained ~8s later

tagged = add_thirds(match.events)               # defensive_third / middle_third / attacking_third
tagged = add_channels(tagged, n=5)              # wide / half-space / central

model = XTModel.fit(match.events)               # fit an xT grid (fit on many matches for real use!)
set_piece_xt_summary(match.events, "corner", model)  # total/average xT added per team

set_piece_added_value(match.events, "corner", model)  # per-delivery: xT added + resulting shot quality + goal
corner_report(match.events, model=model)              # attempts, success/retention/second-phase rate, added value -- one table

delivery_outcomes(match.events, "corner")  # per-delivery: short_corner / direct_shot / second_phase_shot /
                                            # aerial_duel (50/50) / cleared / first_touch_won / first_touch_lost

All of the above are derived heuristics, not raw Opta fields — see docs/source/advanced.rst (or the hosted docs) for the exact assumptions and tunable thresholds behind each one. That page also documents a real bug this uncovered and fixed: F24's eventId is only unique within one team's own event stream, not globally — every delivery/shot lookup in this package is scoped accordingly.

Shot value (experimental)

Five pre-trained gradient-boosted models, bundled with the package, score every shot in a match:

pip install -e ".[ml]"   # xgboost + scikit-learn + joblib

from wa_setpieces.ml.shot_value import ShotValueModels, shot_value

models = ShotValueModels.load()          # loads once; reuse across matches
shots = shot_value(match.events, models)
# eventId, playerName, is_goal, set_piece_type, on_target_prob, xgot, psxg,
# situational_prob, outcome_class_0..3, shot_value (blended)

Read wa_setpieces/shot_value.py's module docstring before trusting this for anything real. The five models were trained elsewhere against a feature schema this package has to reconstruct from Opta F24 qualifiers on each shot event; some inputs (shot geometry, set-piece origin, assist, left/right foot, goal-mouth placement) are confidently derived from already-tested logic elsewhere in this package, but several situational flags (big chance, one-on-one, fast break, scramble, header/volley) have no reliable qualifier signal in the two real matches this was checked against and default to False rather than a guessed-and-possibly-wrong qualifier ID — that gap is documented, not hidden, but it does mean predictions are degraded relative to the models' original training data.

Ratings

wa_setpieces.core.rating turns a report into a single 0-100 "how good" score, benchmarked (z-scored) against whoever else is in the table — always rate against a full season/competition, not one match; a two-row sample just tells you which of those two had the better match, not how good either team actually is.

from wa_setpieces.core.rating import team_rating, player_rating

team_rating(corner_report(season_events, model=model))
# ... success_rate, avg_added_value, retention_rate, plus a *_score column
# per metric and a composite `rating` (50 = this table's own average)

player_rating(season_events, "corner", model, min_deliveries=5, min_shots=3)
# delivery_score (taker quality) and finishing_score (shooter quality),
# merged -- a pure taker or pure finisher is rated on the component they
# have, not penalized for the one they don't

Plots

pip install -e ".[viz]"   # matplotlib + mplsoccer

from wa_setpieces.viz.plots import (
    plot_delivery_map,      # arrow map of deliveries, colored by outcome
    plot_zone_heatmap,      # where events happen, gridded onto the pitch
    plot_xt_grid,           # a fitted XTModel's grid, as a heatmap
    plot_second_phase,      # one corner/free-kick's phase sequence, numbered
    plot_team_comparison,   # grouped bars: both teams, every set-piece type
    plot_xt_added_bars,     # diverging bar chart of xT added per delivery
    plot_corner_sonar,      # polar plot of delivery angle + distance
    plot_match_timeline,    # every set piece on one shared match-minute axis
    plot_dashboard,         # one-figure report card combining several of the above
    plot_set_piece_radar,   # two-team radar over a corner_report/free_kick_report
    plot_set_piece_outcomes,  # shot map: every delivery, colored by outcome category
    plot_rating_benchmark,   # team/player rating vs. the sample-average baseline
    plot_routine_clusters,   # delivery map colored by cluster_routines' data-driven clusters
    plot_defensive_routine_bars,  # what a team concedes most, by routine type or zone
    plot_aerial_duel_win_rate,    # per-team aerial-duel win rate from aerial_duel_summary
)

plot_delivery_map(
    delivery_locations(match.events, "corner"), title="Corner deliveries",
    subtitle="20 June 2026 · Delivery map", footer="Data: Opta", dark=False,  # or dark=True (default)
)
plot_dashboard(match.events, team_id, set_piece_type="corner")  # the "hero" figure
plot_set_piece_radar(corner_report(match.events, model=model))  # team A vs. team B, one glance
plot_rating_benchmark(team_rating(corner_report(season_events, model=model)))

Every plotting function returns (fig, ax) (plot_dashboard returns just fig, being multi-panel) for further customization, and takes dark: bool = True -- the whole figure switches between a validated dark (navy) and light (white) palette with that one argument, see wa_setpieces.viz.theme.get_palette. Colors are assigned by the job they do — a validated categorical palette for team identity (team-vs-team charts use a fixed orange-then-blue pairing in both modes), a status pair for success/fail, gold for goals, single-hue sequential ramps for magnitude, and a diverging pair for signed quantities like xT added — not picked for looks; see wa_setpieces/viz/theme.py. subtitle (a muted line under the title) and footer (a small credit/source line, bottom-right) are optional on every plot. See the gallery for all fifteen plots (in both modes) with full source code.

Other data providers

Opta F24 is the native format (wa_setpieces.core.loader.load_events, handled directly, no adapter needed). wa_setpieces.providers converts other providers' feeds into that same internal frame, so every other module — filters, metrics, chains, phases, retention, xT, value, rating, viz — works unchanged regardless of source:

from wa_setpieces import load_statsbomb_events

events = load_statsbomb_events("statsbomb_events_export.json")
set_piece_summary(events)  # same functions, same DataFrame shape

Read wa_setpieces/providers/statsbomb.py's module docstring for exactly what is (and isn't) faithfully mapped — set-piece detection, the assist-chain shot link, retention, xT and rating are all faithful; one narrow edge case in second-phase timing is documented as an approximation.

Impect is not supported. It's a closed, proprietary feed with no public schema to build and verify an adapter against — contributing one needs a real sample export or an official schema reference to check the mapping against, the same way the StatsBomb adapter and the Opta constants in wa_setpieces/core/constants.py were verified against real exports.

Command line

wa-setpieces match.json
wa-setpieces match.json --csv summary.csv
wa-setpieces match.json --xt   # also fit + print xT for this match (illustrative on one match)

The command-line interface also exposes the complete workflow:

wa-setpieces summary match.json --output summary.json --format json
wa-setpieces train-xt season/*.json --output league-model.npz
wa-setpieces workflow match.json --type corner --model league-model.npz --output tables/
wa-setpieces report match.json --type corner --model league-model.npz --output report.html

Use --provider statsbomb with any new-style command for a StatsBomb events export. Outputs support CSV, JSON and Parquet where applicable.

Season analysis and defending

SeasonDataset makes multi-match aggregation safe by requiring a matchId boundary and running temporal heuristics within each match:

from wa_setpieces import SeasonDataset, defensive_set_piece_summary

season = SeasonDataset.from_sources(paths)
season.summary()                 # competition totals and per-match rates
season.rolling_summary(window=5) # rolling form
season.report("corner", model)   # match-level report rows

defensive_set_piece_summary(season.events)  # attempts/shots/goals conceded

validate_events documents and checks the provider-neutral event contract; event_capabilities reports which optional information an adapter supplies. first_contact_detail and first_contact_summary add player attribution, explicitly labelled event_sequence because event data cannot prove physical contact as tracking data can.

Full xT artifacts now persist all probability grids and training metadata:

model.save("league-model.npz")
loaded = XTModel.load("league-model.npz")
loaded.evaluate(held_out_events)  # shot count, goals and Brier score

Routine analysis

Every restart can also be described by how it was executed, not only its final outcome:

from wa_setpieces import restart_routines, routine_summary, all_routine_summaries

restart_routines(events, "corner")
# one row per corner: routine_type, delivery_technique (inswinger/outswinger),
# post_target (near_post/far_post/central), distance, progression, direction,
# side, start/end third, start/target channel, retention, shots and goals

routine_summary(events, "goal_kick")
# usage share and outcome rates for short_build / medium_build / long routines

all_routine_summaries(events)
# one combined tactical inventory covering all six restart types

The type-specific routine families are:

• Corners: short, central six-yard, penalty-area, recycled, deep/edge.
• Free kicks: direct shot, short, box delivery, progressive, recycled, lateral.
• Throw-ins: short, medium, long, with location and progression attached.
• Penalties: scored, saved, post, missed.
• Goal kicks: short build, medium build, long.
• Kick-offs: backward, lateral, short forward, direct long.

For a structured tactical analysis, use one call:

from wa_setpieces import analyze_routines

analysis = analyze_routines(events, "corner", min_taker_attempts=3)
analysis.detail          # every routine and its geometry/outcome
analysis.summary         # routine-family usage and efficiency
analysis.team_profiles   # diversity, predictability and preferred patterns
analysis.taker_profiles  # taker preferences and creation results
analysis.target_matrix   # routine family -> destination zone -> outcomes

detail also includes approximate distance in metres, delivery angle, verticality, a tactical destination zone, a hierarchical outcome category and a stable routine_key combining family, side and destination. Multi-match frames with matchId are automatically separated before temporal analysis.

Corner-specific extras

A few additions built specifically around corners, all thin layers over the pieces above rather than new inference:

from wa_setpieces import (
    cluster_routines, cluster_summary,          # data-driven routine clusters (k-means)
    defensive_routine_summary, defensive_zone_summary,  # what a team concedes, by routine/zone
    aerial_duel_summary,                          # per-team/player aerial-duel win rate
    corner_report_html,                           # ready-to-view HTML scouting report
    delivery_clip_windows,                        # clip in/out timestamps for video tools
)

clustered = cluster_routines(events, "corner", n_clusters=5)  # optional `ml` extra
cluster_summary(clustered)   # usage/outcome roll-up per cluster, instead of the fixed routine_type buckets

defensive_routine_summary(events, "corner")  # attempts/shots/goals CONCEDED, by opponent's routine_type
defensive_zone_summary(events, "corner")     # same, by destination_zone

team_win_rate, player_wins = aerial_duel_summary(events, "corner")  # who actually won each 50/50

html = corner_report_html(events, model=xt_model)  # write_html_report(path, ..., tables) also works directly

delivery_clip_windows(events, "corner", pre_seconds=5, post_seconds=15)
# clip_start_seconds / clip_end_seconds per delivery, in match-clock seconds

What counts as a set piece

These qualifier IDs are the standard Opta/Stats Perform F24 vocabulary and were cross-checked against a real match export (see tests/data/sample_match.json and tests/test_filters.py): tagged events line up with their expected pitch location (corner arc, touchline, centre spot, six-yard line).

Package layout

• wa_setpieces.core.loader — parse F24 JSON into a tidy pandas.DataFrame; load_events_multi stacks a whole season.
• wa_setpieces.core.constants — Opta typeId / qualifierId reference.
• wa_setpieces.core.filters — extract/tag each set-piece type.
• wa_setpieces.core.metrics — team/player counts, success rates, delivery locations.
• wa_setpieces.core.chains — link set pieces to the shots/goals they produced.
• wa_setpieces.core.zones — pitch thirds, channels and a configurable zone grid.
• wa_setpieces.core.phases — second-phase detection for corners/free kicks.
• wa_setpieces.core.retention — possession retention after any restart.
• wa_setpieces.core.xt — grid-based Expected Threat (xT), fit from data.
• wa_setpieces.core.value — set-piece added value: delivery xT + resulting shot quality + goals, blended.
• wa_setpieces.core.outcomes — per-delivery outcome classification (short corner, direct/second-phase shot, aerial duel, cleared, first/lost touch) for a shot-map scatter, plus aerial_duel_summary for who wins the 50/50s.
• wa_setpieces.core.routines — restart_routines's rule-based taxonomy (including delivery_technique/post_target for corners), plus cluster_routines/cluster_summary for a data-driven (k-means) alternative (optional ml extra).
• wa_setpieces.core.defending — defensive_set_piece_summary/defensive_rating, plus defensive_routine_summary/defensive_zone_summary for what a team concedes by routine type or destination zone.
• wa_setpieces.core.clips — delivery_clip_windows: clip in/out timestamps per delivery, for video-clipping tools.
• wa_setpieces.ml.shot_value — five bundled pre-trained models (on-target probability, xGOT, post-shot xG, situational quality, outcome class) for a richer per-shot value score (optional ml extra; experimental, read the module docstring).
• wa_setpieces.core.report — corner_report/free_kick_report: everything above, merged into one table per team.
• wa_setpieces.core.rating — benchmarked 0-100 team/player "how good" scores from a report (see Ratings below).
• wa_setpieces.core.workflow — run_workflow: the whole pipeline above, one function call (see "The whole pipeline in one call").
• wa_setpieces.reporting — corner_report_html/render_html_report/write_html_report: portable self-contained HTML reports.
• wa_setpieces.providers.statsbomb — convert a StatsBomb open-data export into the same internal frame Opta F24 produces.
• wa_setpieces.viz.plots — mplsoccer/matplotlib plots: delivery maps, heatmaps, sonar, timeline, dashboard, radar, rating benchmark (optional viz extra).
• wa_setpieces.viz.theme — the validated dark/light color palettes every plot draws from.
• wa_setpieces.convert.corners — batch-convert a directory of Opta F24 exports plus a match-list CSV into a flat corners table for tools that expect that schema (optional convert extra).
• wa_setpieces.cli — wa-setpieces command-line tool.

Development

pip install -e ".[dev]"
pytest

Releasing

Publishing to PyPI is automated via GitHub Actions trusted publishing (.github/workflows/publish.yml) — no API token stored anywhere. One-time setup (only a PyPI project owner can do this, since it requires logging into PyPI):

1. On PyPI: https://pypi.org/manage/account/publishing/ → add a pending trusted publisher with project name wa-setpieces, owner marclamberts, repository waltzinganalytics, workflow publish.yml, environment pypi.
2. From then on, publishing a GitHub Release (or pushing a v* tag) builds the sdist/wheel and uploads them automatically.

Docs

Docs are built with Sphinx (pydata-sphinx-theme + sphinx-gallery, the same stack mplsoccer's docs use) and hosted on Read the Docs (.readthedocs.yaml at the repo root). The gallery under examples_gallery/ is executed at build time, so its plots and DataFrame outputs are always current. To build locally:

pip install -e ".[docs]"
sphinx-build -b html docs/source docs/_build/html

License

MIT