BIS 2026
Brownlow Intelligence

Post-Event Evaluation

Frozen Forecast. Real Outcomes.

A post-event audit of 9,522 frozen player-match forecasts across all 207 matches of the 2026 AFL season.

0.3005

True Holdout RMSE

69.6%

3-Vote Winner Accuracy

75.7%

Top-3 Recall

0.968

Season Pearson

Expand a section below for detail. Collapsed by default for a compact overview.

Did the model generalise?

Holdout 0.3005 vs historical mean 0.3559

The frozen 2026 forecast achieved an RMSE of 0.3005 (match-bootstrap 95% CI: 0.2866–0.3145), compared with a mean rolling out-of-time RMSE of 0.3559 across the 2022–2025 development folds.

Forecast vs Reality

Predicted EV vs ACTUAL season votes for key contenders

  • Nick Daicos

    Nick Daicos

    Rank #1 → ACTUAL #1

    40.03

    Predicted

    47

    Actual

  • Marcus Bontempelli

    Marcus Bontempelli

    Rank #2 → ACTUAL #3

    29.45

    Predicted

    34

    Actual

  • Bailey Smith

    Bailey Smith

    Rank #3 → ACTUAL #2

    27.47

    Predicted

    36

    Actual

  • Patrick Cripps

    Patrick Cripps

    Rank #4 → ACTUAL #5

    26.47

    Predicted

    27

    Actual

  • Lachie Neale

    Lachie Neale

    Rank #5 → ACTUAL #16

    25.23

    Predicted

    20

    Actual

  • Will Ashcroft

    Will Ashcroft

    Rank #6 → ACTUAL #5

    25.00

    Predicted

    27

    Actual

  • Harry Sheezel

    Harry Sheezel

    Rank #9 → ACTUAL #5

    20.30

    Predicted

    27

    Actual

  • Max Gawn

    Max Gawn

    Rank #10 → ACTUAL #4

    19.76

    Predicted

    28

    Actual

  • Zak Butters

    Zak Butters

    Rank #12 → ACTUAL #8

    19.29

    Predicted

    26

    Actual

Match-level ranking quality

69.6% top-1 · 75.7% top-3 recall

69.6%

Model #1 received actual 3 votes

75.7%

Actual vote-getters in model Top 3

37.7%

Exact Top-3 player set

18.4%

Exact ordered 3–2–1

Where the model worked

RMSE, season correlation, Top-k overlap, champion ID

Holdout RMSE 0.3005 vs historical mean 0.3559. Season Pearson 0.968, Spearman 0.736. Top-5 / Top-10 overlap 80% / 80%. Predicted #1 Nick Daicos matched ACTUAL (47 votes).

Where the model missed

Largest under- and over-predictions (≥5 games)

Underpredictions

  • Bailey Smith

    Bailey Smith

    27.47 → 36 ACTUAL

    +8.53
  • Max Gawn

    Max Gawn

    19.76 → 28 ACTUAL

    +8.24
  • Clayton Oliver

    Clayton Oliver

    15.22 → 23 ACTUAL

    +7.78

Overpredictions

  • CD

    Caleb Daniel

    6.19 → 0 ACTUAL

    +6.19
  • Lachie Neale

    Lachie Neale

    25.23 → 20 ACTUAL

    +5.23
  • WM

    Wayne Milera

    5.08 → 0 ACTUAL

    +5.08

Forecast compression

3-vote mean EV 1.93 · pred SD 0.40 vs actual SD 0.54

1.93

Mean EV on actual 3-vote games

0.40

Prediction SD

0.54

Actual vote SD

The forecast distribution was narrower than the realised vote distribution — forecast compression / regression toward the mean on extreme outcomes. Descriptive, not causal.

Case study

Nick Daicos

40.03 EV → 47 ACTUAL · #1 → #1

Season forecast

40.03 EV

Actual

47

Predicted rank

#1

Actual rank

#1

Model integrity

SHA256 5be4cc338152… · 207 matches · no recalibration

Frozen forecast rows
9,522
Matches
207
Expected-vote total
1,242
Forecast SHA256
5be4cc338152…
Actual count
207 matches / 1242 votes
Positive-vote join
100%

No retraining. No post-event recalibration. 2026 actual votes were used only for post-event evaluation.

Labels reconstructed from AFL Tables — 2026 Brownlow Medal (retrieved 2026-09-24) — not labelled as an AFL official dump.

What comes next

Research directions · not a V3 model announcement

  • Tail calibration of extreme expected-vote mass
  • Modelling of extreme 3-vote performances
  • Positional under / overprediction patterns
  • Coach-versus-umpire disagreement structure
  • Uncertainty modelling around continuous forecasts