Ghost-Outfield v1 — League-Average Rearguard Positioning (Rest-Defense)
Predicts where a league-average outfield rearguard defender would position themselves, per
tracking frame, team, and lateral slot. The outfield sibling of
Ghost-GK v1: a point estimate (ghost_gr_x/y,
goal-relative) served as the exact pickle-free boosted HistGradientBoostingRegressor mean.
Part of the silly-kicks soccer analytics library
(TF-60 rest-defense; ADR-087). It is the league-average baseline the rest-defense outfield
counterfactual arm differences against ("how far does the actual rest-defense shape sit from average?").
What it models
For an in-possession team's rest defense (the rearguard it keeps while attacking, to blunt the
counter after a loss), the model predicts the league-average position of each of the deepest-n
rearguard defenders, ranked left-to-right by a lateral slot_index. It is possession-conditioned:
one model serves both regimes via a live team_in_possession feature — the ball-carrier's
rest-defense rearguard and the defending line facing an attack.
The feature vector is leakage-safe by construction: no input encodes the modeled team's own
rearguard coordinates (the prediction target). It is ball state + opponent counter-threat geometry +
game context + the lateral slot rank.
Variants
| Variant |
Repo |
Features |
Use |
default (this repo) |
silly-kicks/ghost-outfield-v1 |
20 (velocity-bearing) |
Continuous tracking (Sportec, SkillCorner, Gradient Sports) |
position_only |
silly-kicks/ghost-outfield-position-only-v1 |
16 (4 velocity features dropped) |
Velocity-less StatsBomb-360 freeze-frames |
The serve seam auto-selects the variant from the frames' declared velocity availability; a
velocity-less freeze-frame with no bundled position_only yields honest-NaN, never the (invalid)
velocity model.
Model description
- Point estimate only.
ghost_gr_x/y serve the exact boosted HGBR mean, reconstructed
pickle-free (baseline + Σ_trees leaf_value; no sklearn at inference). There is no KDE /
density read-out (a point estimate is what the counterfactual arm differences).
- Parameters-only, pickle-free: npz (two boosted x/y ensembles + baselines) + JSON metadata +
SHA-256 checksums. No per-sample training data, no raw provider tracking data (ADR-011/044).
- Fail-closed load: SHA-256 + a behavioural chirality fingerprint (a y-mirrored model is
rejected) + a feature-contract fingerprint. The chirality frame hash is pandas-major-invariant
(loads identically under pandas 2 and 3).
- FOV honest-NaN (StatsBomb-360): a served frame whose rearguard region is not sufficiently
observed is returned NaN (
ghost_outfield_source="fov_cropped"), never a fabricated ghost from the
deepest-n visible players.
- Hyperparameters:
HistGradientBoostingRegressor, 500 trees, max depth 8, 5-fold CV; trained at
1 frame/second (25 fps tracking is highly autocorrelated — 1 fps is ~25× fewer,
near-duplicate rows with no meaningful signal loss for a mean-positioning model).
Metrics
| Metric |
Value |
| Held-out CV euclidean MAE |
6.00 m (per-provider: Gradient Sports 6.14 / SkillCorner 5.91 / Sportec 6.33) |
| Per-possession CV MAE |
in-possession 6.97 m / out-of-possession 5.03 m |
| Per-slot CV MAE (slots 1–4) |
5.96 / 6.05 / 6.02 / 5.97 m |
| Training corpus |
179 games / 4170920 rows |
| Rearguard coherence (slot ordering) |
ordering_fraction = 1.0 (the independently-predicted slots order as a line) |
| Boosted-reconstruction parity vs sklearn |
exact (round-trip verified — safe to publish) |
Training data
179 matches / ~4.17M per-(frame, team, slot) rows (at 1 fps) of licensed professional tracking
(Sportec/DFL Bundesliga, SkillCorner, Gradient Sports FIFA World Cup 2022). Only the learned
parameters are published; no raw provider tracking data is redistributed (parameters-only
artifact). Trained from a clean, CI-green commit.
Usage
import silly_kicks.tracking as tracking
model = tracking.GhostOutfieldModel.from_variant("default")
# One row per (frame, team, slot) with goal-relative ghost_gr_x / ghost_gr_y + a provenance token.
served = tracking.serve_ghost_outfield_positions(frames, model=model, home_team_id=1)
Input frames must be LTR-normalized (home attacks right). All coordinates are goal-relative to the
modeled team's defended goal.
Limitations
- League-average positioning, not a specific player's or team's tactical style.
- Static per-frame estimate; inherits tracking-system noise.
- LTR normalization required.
- Unlike the goalkeeper model, this model is not affected by the Gradient Sports keeper clamp
(outfielders are tracked over the full pitch).
References
Le et al. 2017 (Data-Driven Ghosting, MIT Sloan). This model is the boosted-mean point-estimate
variant of the ghosting concept (no density estimation). See the silly-kicks NOTICE for the full
citation.
More information
TF-65 velocity-gap re-fit (4.134.0)
TF-65 velocity-gap re-fit (silly-kicks 4.134.0 / ADR-120; training_commit=a34e992). Re-fit on the dense-grid-corrected floor velocities (TF-65 removed fabricated through-gap velocity on broadcast providers — the prior F1b float32 re-fit at 4.128.0 / ADR-106 is superseded); the metrics table above reflects this re-fit, and the mirror is republished to match the wheel.
