The Brownlow Model: Predicting an Umpire's Perspective

Tackling the problem of modelling the AFL Brownlow Medal is unique in sports prediction. 

The goal is not to build an algorithm to identify the best players in any given match, but rather, to identify the players the umpires officiating that game would think were the best three players, and in which order.

So, how does one predict an umpire’s perception?

DOWNLOAD STATS INSIDER'S FREE 2020 BROWNLOW REPORT

Stats Insider’s approach to answering this question was to first gather as much data as possible from each AFL game between 2014-2019 and train an ordinal logistic regression model on past Brownlow votes to find which variables were the ones most likely to catch the eye of the umpires, and thus correlate to Brownlow Medal votes for that game. That list of statistics is too long to detail here, as we used over 40 different in-game statistics for each player who ran out in every match over this 6-year period.

Using a regression model such as this is a well-worn path for people that have taken on the task of building a Brownlow Medal model, so we also identified other variables beyond in-game statistics to make our model more robust, including a more ‘human’ element with the addition of various media votes into the algorithm.

It is not enough to use just one source for this, as biases - intentional or not - may take place, so we tracked down several different outlets who supplied votes for each game over the past five AFL seasons, as well as using the AFL Coaches Association (AFLCA) votes. Adding these variables to our model produced much higher predictive value than just stats-based regression, and satisfied our goal of building a model that combined both raw data and a ‘wisdom of the crowds’ human approach.

For each player, in every match, our model gives us a probability of that player polling 1, 2 or 3 votes. We’ve then simulated the 2020 AFL season 10,000 times and assigned votes to each player based on these underlying probabilities. By aggregating these simulations, we were able to compute the probabilities and ‘fair odds’ for the different Brownlow Medal betting markets.

What do we mean by ‘fair odds?’

What do we mean by ‘fair odds?’

Fair odds are the price our probabilities indicate the player or market should be at, if the bookmakers were to generate a 100% market (paying out as much as people bet). Anything above this price would represent value, assuming the model is accurate.

To identify fair odds in this case, where there are opportunities for dead heats in some of the markets, we have assessed a ‘tie’ in one of our simulations as a fraction of a win, ie; if three players tie for the Brownlow Medal in 2020, we have awarded each player ⅓ of a Brownlow Medal.

Some interesting figures emerged from the model in relation to possible ties. Based on our 10,000 simulations of the 2020 Brownlow Medal, we have identified:

● 262 simulations with a 3-way dead heat;

● 83 simulations with a 4-way dead heat;

● 2 simulations with a 5-way dead heat.

The more common combinations of 2-way tie winners included:

● Neale / Boak  (188 simulations);

● Neale / Steele (160 simulations);

● Neale / Petracca (72 simulations);

● Boak / Steele (56 simulations)

While some of the more intriguing outright winner dead heats included:

● Boak / Neale / Petracca:Most common 3-way tie  (26 simulations);

● Boak / Neale / Steele /  Petracca: Most common 4-way tie  (5 simulations);

● Fyfe / Neale: Ex-teammates tie  (3 simulations);

● Petracca / Gawn: Melbourne double  (2 simulations);

● Bontempelli / Macrae: Western Bulldogs double (1 simulation)

While no predictive model is infallible - the Brownlow Medal is, in essence, a popularity contest - our 2019 model was very successful, correctly predicting eight out of the final Top 10. 

Two of the stronger plays from the 2019 Brownlow model were:

1. Nat Fyfe to win: while we didn’t pick the Fremantle captain as ’most likely’, our model was stronger on him than the bookmaker’s market. Average price for Fyfe to win the 2019 Brownlow was $6.50, while we had him a 22.9% shot - good for odds of $5.26 after adjusting for a tie (a 3.6% edge);

2. Patrick Cripps leading after Round 5: with the Carlton superstar midfielder on fire to start the 2019 season, our model had him polling near maximum votes in the first five weeks to lead the count after Round 5. The bookmakers had this market at $3.25 for Cripps, while we had him rated at 58.8% to be leading or $2.04 after adjusting for possible ties (an 18.24% edge).

We hope you enjoy our 2020 Brownlow Medal Report, and please remember if you are having a punt on Charlie to always gamble responsibly.

Did you enjoy this article? Leave a comment below, or join the conversation on the Stats Insider Twitter or Facebook page.