AFL Player Ratings: How TORP Works

How our AFL player ratings work: TORP blends EPR, a possession-value rating, with PSR, a stats-based rating, into one number for every AFL player.

What is TORP?

TORP (Total On-field Rating of Players) is our headline AFL player rating. It answers one question: how good is this player right now? It is an average of two ratings that look at a player in different ways, EPR and PSR, weighted 50/50. The scale is points of value above a league baseline, so a rating above zero means we expect the player to help his team more than a typical player in his role.

TORP is a rating, not a score for one game. A rating is a smoothed estimate of how a player will perform next. For what a player did in a single match, see how we value individual games.

What goes into TORP?

Two parts, each covering the other’s blind spot.

EPR (Expected Possession Rating) is built from what actually happened on the field. Its input is each player’s net points in each game: the change in expected points on every action in the play-by-play data, shared out between the players involved, with each team’s total tied to its real margin. EPR rates three parts of that value separately and adds them: the player’s own acts, the ball he wins back, and his share of the team pool.

PSR (Player Stat Rating) is a regression, a model that finds which statistics best predict a result. We take each player’s recent box-score profile (goals, disposals, tackles, marks and dozens of other statistics), add up the profiles of the 22 players in a team, and fit a model that predicts the match margin. The fitted weights are then read back onto each individual player. PSR splits into an offensive part (OSR) and a defensive part (DSR) that always add up to the whole.

EPR trusts the observed value of a player’s actions. PSR trusts the shape of his statistical profile. A player who is strong on one view and ordinary on the other lands near the middle rather than being over- or under-rated by either.

How is it calculated?

Step 1: split each game’s net points. For every player in every game we take his net points and split them three ways: his own acts (his disposals and what he received, less what his turnovers gave away), the ball he won back (turnovers, contests and stoppages won), and his share of the team pool. Totals are put on a per-80-minute basis and then centred within each player’s role, so a ruckman is compared with ruckmen and a key defender with key defenders.

Step 2: weight recent games more. Older games count for less, and the rate of decay differs by part. The decay setting is 630 days for own acts, 273 for the ball won back and 523 for the team pool. Weighting is by minutes on the ground, so a game where a player was on for ten minutes counts for far less than a full game.

Step 3: pull toward average when evidence is thin. A player with three games of data should not be rated as if three games told us everything. We shrink each part toward a prior rate. The noisier a part is from game to game, the more evidence it takes to move: the prior is worth about 24 games for own acts, 14 for the ball won back and 11 for the team pool. A player with a long record is barely moved and a debutant stays close to the starting value. Shrinking pulls every rating toward the prior, so the finished EPR is multiplied by 2.4 to put it back in margin points. That changes no ranking.

Step 4: fit PSR and blend. PSR comes from an elastic-net regression (a regression that drops the statistics it does not need), fitted on matches from seasons before 2025 and tested on 2025 onward. The final rating is TORP = 0.5 x EPR + 0.5 x PSR. Historical ratings use the PSR as it stood at the time, not today’s. One weight looks backwards: centre clearances count against a player, in PSR and in its defensive part, DSR. That is a measured effect, not a mistake. Once the rest of a player’s statistics are known, more centre clearances go with worse results, so we keep the sign the fit found rather than force it positive.

Where can I see it?

Ratings are calculated offline, not live during games, and published as a table the site reads.

What are the limits?

  • TORP is a prediction of a player’s next-game impact, not a verdict on his last game.
  • EPR only sees what the tracking data records. Off-ball work that never shows up as an action is invisible to it.
  • PSR only sees the statistics we count.
  • Players with few games are pulled toward average by design, so a breakout star takes time to register.
  • The rules that share each action’s value between players are set by hand, not fitted. They are judgements, and we say so. The decay settings were tuned by an optimiser against match margins, but for an earlier version of EPR, before it moved to net points.

How accurate is it?

We test the match forecasts built on TORP by replaying past seasons, refitting each round only on games already played. Across 387 matches in 2025 and 2026 (measured August 2026), the forecast margin was off by 25.6 points on average and the forecast picked the winner in 286 of them, 74%. Its Brier score, the average squared gap between the stated win chance and what happened, was 0.179. Lower is better, and always guessing 50% scores 0.25. That test ran before EPR moved to net points in September 2026, so it measures the earlier version of EPR. We have not re-run it since.