ghost-gk-sweeper-position-only-v1

Ghost-GK Sweeper (position-only) v1 — Extended-Grid GK Positioning for Velocity-less Frames

The velocity-less companion to silly-kicks/ghost-gk-sweeper-v1: the same extended-grid (x_max = 52.5 m) high-sweeper model, but with the 5 velocity features dropped (21 features instead of 26) so it scores on StatsBomb-360 freeze-frames, which carry no per-player temporal history.

Part of the silly-kicks soccer analytics library (TF-60 rest-defense; ADR-067 velocity-keyed variants; ADR-083).

Why a separate model

The bundled Ghost-GK variants train on a fixed x ∈ [0, 30] m goal-relative box and hard-saturate at 30 m, so they cannot place an in-possession high sweeper (30–45 m). This sweeper_position_only variant lifts the ceiling to x_max = 52.5 m AND drops the velocity features so the model is valid on velocity-less freeze-frames (velocity features are dropped, not NaN-filled — the feature contract raises on a non-finite input rather than imputing). It is auto-selected at the serve seam for a sweeper request on declared velocity-less frames.

The frozen default/position_only/full variants are unchanged — additive; opt in via from_variant("sweeper") (which resolves to this variant on velocity-less frames).

Model description

• Point estimate only (ghost_gk_x/y, exact pickle-free boosted HGBR mean); predict_density is not supported on the extended grid and raises (ADR-083).
• 21 features (the 5 velocity features — ball_vx, ball_vy, ball_speed, defensive_line_speed, defending_centroid_vx — are dropped, not imputed).
• Grid: x ∈ [0, 52.5] m, y ∈ [18, 50] m, 0.5 m resolution.
• Parameters-only, pickle-free (npz + JSON + SHA-256); no per-sample or raw provider data (ADR-044).
• Hyperparameters: HistGradientBoostingRegressor, 500 trees, max depth 8, 5-fold CV.

Metrics

> 30 m coverage is IDSSE/Sportec-dominated (11.5 %); SkillCorner 0.24 %; Gradient Sports 0.0 % (see the data caveat).

Both-axes goal-relative re-fit (silly-kicks 4.111.0 / ADR-089; training_commit=4bda048). These weights were re-fit after the ghost-GK feature extractor was corrected to apply the full 180° goal-relative point reflection (both axes) — the previous convention flipped only x, leaving the signed-y features (the velocity-less variant keeps ball_y, ball_to_goal_angle, attacker_centroid_y) and the target gk_y in absolute frame coordinates, mislocating away-team keepers in y. The served ghost_gk_x/y change (chiefly for away-team keepers) and the feature-contract + chirality fingerprints change, so pre-4.111.0 artifacts fail closed on load. The aggregate MAE is essentially unchanged (the correction affects the away-team subset).

Training data

Same 179-match licensed corpus as the faithful sweeper (Sportec/DFL, SkillCorner, Gradient Sports WC2022); parameters-only, no raw tracking data redistributed; clean CI-green training commit.

⚠ Gradient Sports goalkeeper clamp. Gradient Sports clamps the goalkeeper's tracked position to a hard 27.5 m from goal (source-data limitation, verified on the raw provider data). GS contributes no high-sweeper signal and any GS goalkeeper-depth analysis is invalid past 27.5 m; silly-kicks flags it via validate_gk_position_clamp / GoalkeeperClampWarning. See docs/research/gs_keeper_clamp/.

Usage

import silly_kicks.tracking as tracking

# from_variant("sweeper") auto-selects THIS variant on declared velocity-less (freeze-frame) input.
model = tracking.GhostGkModel.from_variant("sweeper_position_only")
out = tracking.compute_ghost_gk(frames, model=model, home_team_id=1)  # ghost_gk_x / ghost_gk_y

Input frames must be LTR-normalized; coordinates are goal-relative. predict_density raises on the extended grid — use the mean/serve path.

Limitations

• League-average positioning, not a specific keeper's style; no shot-stopping ability.
• Velocity-less by design (21 features) — use the 26-feature ghost-gk-sweeper-v1 on velocity-bearing continuous tracking.
• No KDE density on the extended grid; Gradient Sports keeper positions unusable past 27.5 m.
• LTR normalization required; static per-frame estimate.

References

Le et al. 2017 (MIT Sloan); Dutta et al. 2024 (arXiv:2406.17220); Pospisil & Lee 2018 (arXiv:1804.05753). See the silly-kicks NOTICE for full citations.

More information

• License: MIT · Library: silly-kicks (v4.111.0+ for these both-axes weights) · GitHub · ADR-067 / ADR-083 / ADR-089.

F1b re-fit (4.128.0)

F1b float32-frame re-fit (silly-kicks 4.128.0 / ADR-106; training_commit=3ca609f). Re-fit on float32-stored frames (the storage rounding exceeds the trained-model feature-contract atol); the metrics table above reflects this re-fit, and the mirror is republished to match the wheel.