silly-kicks/xcross-attempt-position-only-v1

xCrossAttempt v1 (sc_extended_position_only) — Cross Propensity for Velocity-less Frames

Read this first. This repo serves the sc_extended_position_only (owner-tier, position-only) variant — the owner-tier xCross model with velocity features dropped, for scoring velocity-less snapshots such as StatsBomb-360 freeze-frames. It is NOT bundled with the wheel (restricted owner-tier training data — a licensing constraint). If you do not have owner-tier access, use the bundled position_only variant. Reachable ONLY via from_variant("sc_extended_position_only"); asking for sc_extended returns the faithful (velocity-bearing) model (ADR-070).

Model Description

XCrossAttemptModel (position-only) estimates P(the in-possession team attempts a cross within ~1 s of a tracking frame) using position-only features — the velocity-derived features are dropped (not NaN-filled), so the model scores frames that carry no per-player velocity at all.

It carries the STATE-anchored cross-propensity framing (Cao et al., arXiv:2505.11841) plus the isolatable GK-position confounder block, minus the velocity terms.

• 15 features (position_only), goal-relative coordinates via the shared _geometry helper
• Domain filter: alive-ball, wide-area; base rate 5.0% positive
• Fit on the owner-tier corpus (IDSSE + SkillCorner incl. owner-tier)

When to use this vs the bundled position_only

Variant	Corpus	Where it lives	Use it?
position_only (bundled)	public corpus	bundled in the wheel	Default for velocity-less frames without owner-tier access
sc_extended_position_only	IDSSE + SkillCorner incl. 98 owner-tier SkillCorner matches	this repo (HF-only)	A stronger velocity-less model, if you have owner-tier access and can accept a Hub download

Use this when your frames genuinely lack velocity. For velocity-bearing tracking, prefer the faithful sc_extended model (silly-kicks/xcross-attempt-v1).

Why this variant is HF-only

HF-only for licensing, not quality: trained on restricted owner-tier SkillCorner data that cannot be redistributed in the wheel (ADR-038). Only learned parameters are published — no raw provider tracking data. The Hub repo is the owner-tier archive (ADR-071). training_commit: 1ce63ef.

Held-out CV (5 folds, out-of-fold)

Metric	Value	Baseline
PR-AUC	0.1297 (± 0.0087)	base rate 0.0500
Brier	0.0454	base-rate Brier 0.0475
Log loss	0.1771	—

All four acceptance gates pass. Estimates are CV, not the shipped fit. Dropping velocity costs discrimination vs the faithful model (PR-AUC 0.189 → 0.130) — the trade is that this model runs at all on velocity-less frames.

TF-19 GK-substitution probe

The frozen GK-substitution probe (gk_substitution_probe in metrics.json; 200 frames; ADR-037's two-prong gate — ratio ≥ 2.0 × the nearest-defender control and an absolute floor ≥ 0.01):

Metric	Value
gk_median_abs_delta	0.01639
nearest_def_median_abs_delta	0.00664
ratio (gk / control)	2.47× — clears (≥ 2.0)
absolute floor	0.01639 ≥ 0.01 — clears
tf19_ready	true

This position-only variant clears both prongs of the frozen TF-19 gate — the only one of the four xShot/xCross re-fits to do so. The GK-block ablation is consistent (removing the GK block drops held-out PR-AUC by 0.0023). See ADR-037 for the gate definition.

Usage

from silly_kicks.tracking import XCrossAttemptModel

model = XCrossAttemptModel.from_variant("sc_extended_position_only")  # this repo, Hub download

Requires pip install silly-kicks[xcross] and silly-kicks >= 4.94.0 (the release that introduced this variant key — ADR-070). The >= 4.74.0 geometry floor (goal-relative-2) also applies; load() is fail-closed on the feature contract + chirality fingerprint (ADR-040). from_hub() takes no revision argument yet — treat the library version as the pin.

Integrity and load-time guards

load() is fail-closed on SHA256SUMS and the chirality fingerprint (ADR-040), plus a base_score guard for xgboost 3.x serialization. The position-only feature contract raises on a non-finite (NaN-filled) velocity feature — this model expects those columns absent, not imputed.

Limitations

• Not the bundled model (restricted corpus). Redistribution limit, not a performance one.
• Weaker than the faithful sc_extended by construction (velocity dropped) — use only when frames lack velocity.
• tf19_ready = true (see the TF-19 section) — this variant clears the frozen gate, but the verdict is a routing signal, not a construct-validity claim; validate before building a headline TF-19 consumer.
• Trained on an owner-tier corpus that is heavily one-club; club/style confounding is real and unquantified.
• Estimates are cross-validated, not a held-out test of the shipped fit.

References

See the NOTICE file in the silly-kicks repository for full bibliographic citations.

• Cao et al. "Framing Causal Questions in Sports Analytics: A Case Study of Crossing in Soccer." arXiv:2505.11841 (2025).
• Decisions: ADR-011 (trained-model lifecycle), ADR-015 (causal-validation port), ADR-038 (corpus + visibility), ADR-040 (chirality enforcement), ADR-063 (velocity-less lift), ADR-067 (position-only variants + velocity auto-select), ADR-070 (position-only Hub variant), ADR-071 (owner-tier archive).

Model Files

File	Purpose
model.json	XGBoost booster (pickle-free)
metadata.json	features, hyperparameters, chirality fingerprint, provenance
metrics.json	CV metrics, GK-substitution probe, ablation, permutation importance
SHA256SUMS	integrity manifest, verified by load()

More Information

https://github.com/karsten-s-nielsen/silly-kicks