AFL Expected Possession Value

How we measure AFL expected possession value (EPV): a points value for where the ball is, and how each disposal and reception adds to a player’s game value.

What is AFL expected possession value?

Expected possession value, or EPV, puts a points value on where the ball is right now. Given the spot on the ground, the clock and how play got there, how many points is the team in possession on track to score, net of whatever the opponent scores first? A kick inside 50 in front of goal is worth a lot. A kick-out from the goal square is worth very little, or less than zero.

The change in that value from one action to the next is the value the action created or destroyed. Add those changes up for a player over a game and you have a measure of how much he moved the scoreboard, which is what we mean by a player’s value in a match.

How is the value of a position calculated?

We call the position value EP (expected points). It comes from a machine-learning model, XGBoost, which builds many small decision trees and combines their answers. The model reads the situation at each recorded action and predicts five outcomes for the rest of the quarter:

  1. the opponent scores the next goal
  2. the opponent scores the next behind
  3. the team in possession scores the next behind
  4. the team in possession scores the next goal
  5. nobody scores again this quarter

Those five probabilities become one number on the real AFL points scale:

EP = -6 x P(opponent goal) - P(opponent behind) + P(behind) + 6 x P(goal)

In practice EP runs from about -6 to +6. It is strongly negative deep in a team’s own defensive 50, strongly positive with the ball in range of its own goal, and close to zero around the centre. Centre bounces are set to zero, because a restart is a genuine coin flip.

The horizon is the quarter, not the match, because every quarter restarts with a centre bounce and we do not let the label carry across a quarter break. That also means an action in the last seconds of a quarter is pulled toward “no score”, however good the position looks.

The live model uses 14 inputs: where the ball is on the ground, how far into the quarter we are and how much playing time is left in the quarter and the match, whether the action is a shot, how many actions into the chain of possession it is, the type of play (kick, handball, reception) and the phase of play (hard ball, loose ball, set shot, handball received). It is trained on completed seasons from 2021 onward, with the current season held out as an informal check.

How does a player get credit?

Credit is shared between the two players on each action. For a disposal, the disposer gets a share of the change in EP and the receiver gets a share of the same change. On contested kicks the disposer’s share falls from half to a third, with the rest going to whoever wins the contest. Spoils, tackles, pressure acts, intercepts and hitouts earn value from fixed weights. The totals are then compared with the average player in the same role, so value is measured against what a midfielder or a key defender would normally produce.

The per-game value for a player is blended with a second measure, PSV (Player Stat Value). PSV applies the weights from our TORP statistics model to a single game’s box score. The per-game blend torp_value = 0.5 x EPV + 0.5 x PSV is the single-game sibling of the TORP rating.

What are net points?

Net points are the EPV figure the AFL pages show for each player in each game. They are tied to the real scoreboard: the players on a team add up to exactly what that team won or lost by, so in a 20-point game the winners total +20 and the losers total -20. Each player’s figure splits into three parts: value from their own disposals and other acts, value from winning the ball back (turnovers won, contested marks and spoils, ruck and stoppage work), and a share of a team pool, which also carries the small top-up needed to match the real margin. Because the totals are tied to the actual result, net points read in real points. A big game on a losing side can still show as a big positive number, because each player is credited for what they did and only the team total is pinned to the margin.

Across a season, the AFL pages also show TOPV (Total On-field Player Value): a player’s season EPV and PSV blended 50/50, the season-long counterpart of torp_value.

What is the difference between a rating and a value?

This trips people up, so we keep the naming strict. A rating answers “how good is this player?” It is smoothed over many games and predicts the next one. A value answers “what did he do in this game?” It bounces around from week to week.

  • EPR (Expected Possession Rating) is the rating, built by decaying and averaging per-game EPV credit over a player’s career. It is half of TORP.
  • EPV is the per-game or per-play value that EPR is built from.
  • PSR is the career-smoothed statistics rating, and PSV is the same model applied to one game.

See how TORP works for the rating side.

Where can I see it?

During a live game the position value is scored on our server from the live feed. The ratings that depend on it are calculated offline.

What are the limits?

  • EP only knows what the tracking data records: position, timing and the type of action. It cannot see defensive pressure, a player’s skill or whether a kick was well weighted.
  • The five outcomes are for the rest of the quarter. A goal at the 20-minute mark of a quarter that would have been scored anyway still earns the same value.
  • Per-game value is a noisy measure of one game. A great week does not make a great player.
  • The credit rules, such as the contested-kick split, are judgements we set by hand. They are not fitted.