Shot value (experimental)

Warning

Read wa_setpieces.ml.shot_value’s module docstring in full before trusting this output for anything real. The five bundled models were trained elsewhere against a feature schema this package has to reconstruct from Opta qualifiers – some inputs are confidently derived, several are experimental best-effort defaults (documented in the module docstring), not verified ground truth.

Five pre-trained gradient-boosted models, bundled with the ml extra, score every shot in a match: P(on target), xG On Target (P(goal) given on target), Post-Shot xG (P(goal) from actual placement), a situational signal, and a 4-class outcome distribution. shot_value() runs all five and blends psxg and on_target_prob * xgot – two independent P(goal) estimates – into one shot_value column.

from pathlib import Path

from wa_setpieces import load_matches
from wa_setpieces.ml.shot_value import ShotValueModels, shot_value

try:
    _here = Path(__file__).resolve().parent
except NameError:
    _here = Path.cwd()
DATA = _here.parent / "tests" / "data" / "sample_match.json"

events = load_matches(DATA)

Load once, reuse across matches (the .pkl files are a few hundred KB each and slower to load than to run inference with):

models = ShotValueModels.load()
shots = shot_value(events, models)
shots[["eventId", "playerName", "is_goal", "set_piece_type", "on_target_prob", "xgot", "psxg", "shot_value"]]

	eventId	playerName	is_goal	set_piece_type	on_target_prob	xgot	psxg	shot_value
0	21	P. Velasco	False	None	0.316310	0.104364	0.000361	0.016686
1	22	G. Rostagno	False	None	0.280220	0.505882	0.000361	0.071060
2	88	S. Vasquez	False	None	0.394737	0.104364	0.018957	0.030077
3	161	J. Corozo	True	None	0.362346	0.104364	0.044335	0.041075
4	249	M. Estrada	False	None	0.406040	0.505882	0.375000	0.290204
5	281	A. Tobar	True	None	0.362346	0.156584	0.044335	0.050536
6	318	J. Pretell	False	None	0.355072	0.369231	0.000361	0.065732
7	337	D. Armas	False	None	0.277270	0.505882	0.000000	0.070133
8	361	R. Jaramillo	True	None	0.362346	0.104364	0.044335	0.041075
9	480	A. Herrera	False	None	0.323395	0.135593	0.000361	0.022105
10	504	P. Charpentier	False	None	0.401198	0.356061	0.000361	0.071606
11	483	A. Tobar	False	None	0.362346	0.264151	0.000361	0.048037
12	586	R. Jaramillo	False	None	0.401198	0.746269	0.000361	0.149881
13	594	N. Molina	False	None	0.362346	0.264151	0.000361	0.048037
14	599	P. Velasco	False	None	0.468034	0.554217	0.000000	0.129696
15	608	P. Charpentier	False	None	0.468034	0.264151	0.000361	0.061996
16	611	N. Molina	False	None	0.435115	0.264151	0.000361	0.057648
17	564	J. Corozo	False	None	0.362346	0.156584	0.000361	0.028549
18	641	A. Herrera	False	None	0.360161	0.302326	0.000361	0.054623
19	696	M. Bermúdez	False	None	0.468034	0.826590	0.001458	0.194165
20	721	M. Bermúdez	False	None	0.394737	0.264151	0.000361	0.052315
21	734	A. Herrera	False	None	0.360161	0.746269	0.000361	0.134569
22	766	V. Burgoa	False	None	0.277270	0.264151	0.000361	0.036801
23	674	G. Villamíl	False	None	0.323395	0.078261	0.018957	0.022133
24	678	K. Minda	False	None	0.362346	0.104364	0.000361	0.019088
25	855	S. Vasquez	False	None	0.468034	0.223214	0.000000	0.052236

Set-piece shots specifically, ranked by blended shot value:

set_piece_shots = shots[shots["set_piece_type"].notna()].sort_values("shot_value", ascending=False)
set_piece_shots[["playerName", "set_piece_type", "on_target_prob", "psxg", "shot_value"]]

	playerName	set_piece_type	on_target_prob	psxg	shot_value