Corner-kick analysis
TacticAI Lite
Which corner patterns are associated with a shot?
Explore observed deliveries and partial 360 configurations from two European Championships. A research tool inspired by TacticAI, with a reproducible Python pipeline behind every result.
- Dataset
- Loading reviewed findings…
Shot creation by observed pattern
One or more attacking shots within 15 seconds, in the same period and possession, before a restart or stoppage.
Dots show observed rates; lines show descriptive 95% Wilson intervals. The intervals assume independent corners and do not account for dependence within matches. These associations do not establish that changing delivery causes more shots.
Clusters group visible spatial features without using outcomes. IDs are descriptive, not inferred tactical routines. Camera coverage can drive similarity. The unassigned group is a coverage reference, not a configuration.
Exact counts and intervals
| Tournament | Pattern | Shot corners | Corners | Rate | 95% interval |
|---|
A modest model check
Can the visible configuration predict shot creation on different matches?
| Model | Brier ↓ | Log loss ↓ | ROC AUC ↑ |
|---|
The time window matters
The same possession and stoppage rules, measured over three explicit windows.
| Tournament | 10 seconds | 15 seconds | 20 seconds |
|---|
This table includes all recorded corners, including four directly shot corners. Delivery comparisons and the classifier exclude direct shots because their shot outcome is automatic.
What this system can support
Explicit definitions
Short means a recorded pass length of at most 15 standardized units. Swing uses the provider’s recorded technique. Unknown techniques remain unknown. Multiple shots count once for the binary outcome.
360 measures visible players, box counts, centroids, spread, nearest-defender distances and density within the camera-visible part of the box. Missing frames and unseen players are not zero defensive density.
Inspect, then interpret
This public site contains aggregate findings. The Python/Streamlit project adds local corner inspection, linked player frames, event timelines and configuration search after downloading directly from the provider.
360 is broadcast snapshots, not continuous tracking. It cannot establish runs, marking assignments or tactical intent. This is inspired by TacticAI, not a reproduction of its proprietary data or model.
Data: StatsBomb Open Data. Research inspiration: Wang et al., Nature Communications (2024).