Given high-resolution information on the position of a Rocket League replay frame in play that occurs in between goals, how accurately can we identify its "latent goal-probability state"; that is, the underlying probabilities of one or neither team scoring within some period of time that describe the current "momentum" of the game.
Training Data: Rocket League replays collected from the following seasons and parsed with the AnalyzeRL Boxcars library:
- 2021-22
- 2022-23
- 2024
- 2025
Testing Data: Rocket League replays collected from the following seasons and parsed with the AnalyzeRL Boxcars library:
- 2026
Outputs from replay parsing yield data in 30 frames per second, or 30hz. For each replay in these seasons, the full replay frames dataset is first downsampled into 10hz (hinged on the frames directly before a goal is scored) and then is reduced into the following "valid" training frames:
- Frame that occurs during an "active" (mid-play) frame state
- Frame occuring in a game-state in which neither team is winning by more than three goals
- Frame occuring in the game-state in which there is at most three players on the field for a single team
Model Metrics, Out-of-Sample Test Data:
- Log Loss (Collapsed, Three-Way): 0.402510
- Brier Score: 0.206863
- Macro ROC-AUC: 0.773212
- Macro PR-AUC: 0.577936
