Datasets:

luxury-lakehouse
/

obso-trained-grids

origin_zone int64 0 399 ⌀	target_zone int64 0 399 ⌀	probability float64 0 1 ⌀	competition_id stringclasses 9 values
0	0	0.5	7
0	105	0.5	7
4	79	1	7
6	6	0.666667	7
6	33	0.333333	7
7	7	1	7
8	7	0.5	7
8	35	0.5	7
9	9	0.666667	7
9	28	0.333333	7
10	10	1	7
11	0	0.0625	7
11	11	0.6875	7
11	25	0.0625	7
11	35	0.0625	7
11	36	0.0625	7
11	230	0.0625	7
12	12	0.666667	7
12	60	0.333333	7
13	13	0.714286	7
13	81	0.142857	7
13	106	0.142857	7
14	14	0.333333	7
14	16	0.333333	7
14	197	0.333333	7
16	16	1	7
17	17	0.666667	7
17	47	0.166667	7
17	49	0.166667	7
19	19	0.5	7
19	47	0.5	7
20	20	0.5	7
20	22	0.5	7
22	22	0.75	7
22	48	0.25	7
23	23	1	7
25	25	0.666667	7
25	26	0.166667	7
25	50	0.166667	7
26	25	0.142857	7
26	26	0.571429	7
26	51	0.142857	7
26	105	0.142857	7
27	27	0.75	7
27	229	0.25	7
28	28	0.333333	7
28	29	0.333333	7
28	53	0.333333	7
29	157	1	7
30	30	0.5	7
30	100	0.25	7
30	203	0.25	7
32	9	0.166667	7
32	32	0.5	7
32	51	0.166667	7
32	128	0.166667	7
33	33	0.714286	7
33	51	0.142857	7
33	77	0.142857	7
34	32	0.2	7
34	34	0.4	7
34	50	0.1	7
34	59	0.1	7
34	71	0.1	7
34	93	0.1	7
35	35	0.571429	7
35	61	0.142857	7
35	109	0.142857	7
35	258	0.142857	7
36	31	0.0625	7
36	36	0.3125	7
36	37	0.0625	7
36	41	0.0625	7
36	50	0.0625	7
36	56	0.0625	7
36	58	0.0625	7
36	80	0.0625	7
36	104	0.0625	7
36	116	0.0625	7
36	121	0.0625	7
36	198	0.0625	7
37	37	0.428571	7
37	82	0.142857	7
37	104	0.142857	7
37	123	0.142857	7
37	142	0.142857	7
38	0	0.083333	7
38	38	0.5	7
38	39	0.083333	7
38	89	0.083333	7
38	91	0.083333	7
38	171	0.083333	7
38	246	0.083333	7
39	14	0.083333	7
39	38	0.083333	7
39	39	0.416667	7
39	49	0.083333	7
39	59	0.083333	7
39	64	0.083333	7
39	155	0.083333	7

End of preview. Expand in Data Studio

OBSO Trained Grids — Reachability, EPV, and Completion Matrices

Pre-trained grid artifacts for Off-Ball Scoring Opportunity (OBSO) computation: ball reachability surfaces, expected possession value (EPV) grids, and pass completion probability matrices. These are the static lookup tables that power real-time OBSO evaluation — derived from observed passing, shooting, and transition patterns across ~4,900 open-data matches.

Part of the (Right! Luxury!) Lakehouse soccer analytics platform.

Quick Start

from huggingface_hub import hf_hub_download
import pandas as pd

repo = "luxury-lakehouse/obso-trained-grids"

# Load the global reachability grid (100x64)
reach_path = hf_hub_download(repo, "data/reachability_grid_global.parquet", repo_type="dataset")
reach_df = pd.read_parquet(reach_path)
reach_matrix = reach_df.pivot(index="zone_y", columns="zone_x", values="reachability")
print(f"Reachability grid shape: {reach_matrix.shape}")  # (100, 64)

# Load the global EPV grid (50x32)
epv_path = hf_hub_download(repo, "data/epv_grid_global.parquet", repo_type="dataset")
epv_df = pd.read_parquet(epv_path)
epv_matrix = epv_df.pivot(index="zone_y", columns="zone_x", values="epv_value")
print(f"EPV grid shape: {epv_matrix.shape}")  # (50, 32)

Explore interactively: HF Space demo

What Is This Dataset?

The OBSO model evaluates off-ball scoring opportunities by combining three components:

Reachability — the probability that a ball played to a given zone can be controlled by the receiving team, based on observed reception patterns.
Expected Possession Value (EPV) — the probability that a possession in a given zone will result in a goal, estimated via value iteration over transition and shot frequencies.
Pass Completion Probability — the likelihood that a pass from one zone to another is completed, estimated from observed pass outcomes.

These grids are pre-computed from event data and serve as static inputs to the real-time OBSO pipeline, which combines them with dynamic pitch control surfaces.

Data Fields

The dataset contains three types of grid artifacts at different resolutions, plus per-competition variants.

Reachability Grid (`reachability_grid_global.parquet`)

Column	Type	Description
`zone_y`	`int`	Grid row index (0–99)
`zone_x`	`int`	Grid column index (0–63)
`reachability`	`float`	Ball reachability probability (0–1)

Grid resolution: 100×64 (105m ÷ 100 = 1.05m per row, 68m ÷ 64 = 1.0625m per column).

EPV Grid (`epv_grid_global.parquet`)

Column	Type	Description
`zone_y`	`int`	Grid row index (0–49)
`zone_x`	`int`	Grid column index (0–31)
`epv_value`	`float`	Expected possession value (0–1, higher = closer to goal)

Grid resolution: 50×32 (105m ÷ 50 = 2.1m per row, 68m ÷ 32 = 2.125m per column).

Completion Matrix (`completion_matrix_global.parquet`)

Column	Type	Description
`origin_zone`	`int`	Flat index of the origin zone (0–399)
`target_zone`	`int`	Flat index of the target zone (0–399)
`probability`	`float`	Pass completion probability (row-normalized, 0–1)

Zone grid: 25×16 = 400 zones (105m ÷ 25 = 4.2m per row, 68m ÷ 16 = 4.25m per column).

Per-Competition Variants

Per-competition files (*_all.parquet) include an additional competition_id column. These allow competition-specific OBSO computation where sufficient data exists.

Coordinate System

All grids map to the SPADL 105×68 meters pitch coordinate space. Grid resolution varies by artifact type (see tables above). The origin (0, 0) is at the bottom-left corner of the attacking team's half.

Data Sources

Source	Matches	License
StatsBomb Open Data	~3,000	CC-BY 4.0
Wyscout Public Dataset	~1,900	CC-BY-NC 4.0

All event data is converted to SPADL format before grid estimation.

Companion Resources

Resource	Description
OBSO/PAUSA Values	Per-pass OBSO and PAUSA scores computed using these grids
Expected Threat (xT) Grids	Markov chain xT grids from the same source data
SPADL/VAEP Action Values	Per-action VAEP scores from the same source events

Limitations

Open data only: Grids are trained on publicly available StatsBomb and Wyscout data. Commercial datasets with denser event coverage may yield different surfaces.
Competition-agnostic global grids: The global variants pool all competitions. League-specific tactical patterns (e.g., high-press vs. low-block) are averaged away.
Static estimates: Grids are computed from full-season aggregates. They do not adapt to in-game state (score, time, fatigue, personnel).
Resolution trade-offs: Each grid uses a different resolution optimized for its purpose. Interpolation is required when combining grids of different resolutions.
No goalkeeper modeling: Reachability and EPV grids do not distinguish goalkeeper positioning from outfield player patterns.

Citation

If you use this dataset, please cite the underlying models:

@misc{singh2018expected,
  title={Introducing Expected Threat (xT)},
  author={Singh, Karun},
  year={2018},
  url={https://karun.in/blog/expected-threat.html}
}

@inproceedings{spearman2018beyond,
  title={Beyond Expected Goals},
  author={Spearman, William},
  booktitle={MIT Sloan Sports Analytics Conference},
  year={2018}
}

@inproceedings{fernandez2018wide,
  title={Wide Open Spaces: A statistical technique for measuring space creation in professional soccer},
  author={Fernandez, Javier and Bornn, Luke},
  booktitle={MIT Sloan Sports Analytics Conference},
  year={2018}
}

@inproceedings{lee2026pausa,
  title={Valuing La Pausa: Quantifying the Timing and Quality of Soccer Passes Using Off-Ball Scoring Opportunities},
  author={Lee, Minho and Jo, Hyunsung and Hong, Seungwon and Bauer, Pascal and Ko, Sangkuk},
  booktitle={MIT Sloan Sports Analytics Conference},
  year={2026}
}