Research Findings
Visual findings from validated ablation and temporal experiments. Associations are predictive — not causal claims.
Coach votes are a strong contemporaneous human-evaluation signal. Decomposing coach recognition into expected and residual components produced a smaller additional improvement. Do not interpret this as causality.
Historical out-of-time diagnostics on vote strata. Zero-vote dominance and mean prediction by actual vote illustrate how sparse Brownlow recognition is — and how the model responds across the vote ladder.
Historical rows
96,764
Zero-vote rate
93.3%
Champion OOT RMSE
0.355865
Mean prediction & MAE by actual Brownlow votes
Navy = mean prediction · Gold = MAE. Supported by project diagnostics only — no invented performance relationships.
Full historical training remained the strongest validated strategy. Trailing windows and excluding 2020 did not improve the benchmark.
| Strategy | OOT RMSE | Relative |
|---|---|---|
| expanding_allchampion | 0.355865 | |
| trailing_8 | 0.355922 | |
| exclude_2020 | 0.356953 | |
| trailing_6 | 0.356982 | |
| trailing_4 | 0.357914 |