Expected points, usually shortened to xP, gives a rugby match situation an average scoring value. A figure of 1.7 xP does not mean that 1.7 points will appear, nor does it predict the final score. It combines several possible scoring outcomes, weighted by how likely each is in comparable situations. The number can describe the quality of a possession under stated conditions. It cannot know which pass will spill, whether a kick will drift wide, or how one defender will react.
A 2025 rugby union methodology study recorded 35,199 phases across 132 Premiership matches. Its variables included the team in possession, field location, play type, score difference, time remaining, and the next scoring outcome. Four machine-learning methods were tested, yet the best model fell below the study’s baseline for practical use, which gives us some sense of how difficult it can be to get these predictions right. That said, they are useful to viewers everywhere, so let’s figure out a bit more about them!
- Probability Has a Timestamp
A probability is tied to an information set: everything known at the moment it is calculated. In rugby, that may include some of the variables mentioned above, possibly alongside other elements. Together, those conditions form the game state. A kick to touch, a turnover, or the next phase creates a new state with different possible outcomes, but it does not alter what was known beforehand. Whenever information arrives in stages, each new reveal narrows the remaining possibilities without turning the earlier estimate into an error.
We can see this in lots of other areas too. Poker, for instance, exposes probability one card at a time. Before the final card is dealt, a player knows their own cards and the shared cards, but cannot see an opponent’s exact holding. Previous actions narrow the plausible range without reducing it to one certain hand. For an Australian player engaging in an online poker hand through Ignition Australia, any percentage attached to a choice belongs to the information available at that moment. A later card removes some possibilities and creates others, so the calculation changes with the state of play.
The probabilities shifting due to a new card being revealed does not make the earlier estimate wrong. They remain the best interpretation of the outcomes given all information available at the time. An unlikely card was always a possibility, while a likely outcome was never guaranteed. Reviewing the choice therefore requires the earlier cards, action sequence, and estimated range, not merely the eventual reveal. A high-value rugby possession can end with no score, while a low-value position can produce 7 points. The result records one realised path through a larger set of possibilities.
Suppose a simplified possession has a 20% chance of producing a converted try, a 10% chance of producing a penalty goal, and a 70% chance of producing no points. Multiplying probability by value gives 0.20 × 7, plus 0.10 × 3, plus 0.70 × 0, for 1.7 xP. A full calculation must also include any route by which the opposition could score next, giving those outcomes negative values.
Absolute xP and the change between two states answer different questions. A possession valued at 2.0 xP may look strong in isolation, but an action that began from 2.6 has reduced the expected value by 0.6. Moving from 0.2 to 1.0 adds 0.8, despite leaving the team in a less threatening position overall. Attribution remains difficult because the change may depend on ruck speed, support lines, a decoy run, or a defensive mistake, rather than the ball carrier alone. The before-and-after states must use the same model and outcome definitions.
- What the Model Leaves Out
Two possessions beginning on the same patch of grass may carry different value. A quick ball against a retreating defence offers options that a slow ball against a set line does not. A general model may group those situations together unless its inputs capture defensive shape, ruck speed, and available attackers. Adding detail can improve specificity, but every extra variable needs enough reliable examples. Team-specific data may reflect a particular kicker, lineout, or attacking pattern more closely while producing a smaller sample and less stable estimates.
After five rounds of the 2025 Super Rugby Pacific season, The Guardian reported 62.4 points per match, alongside faster games, law changes, and teams favouring attacking kicks to the corner. That pace changes how often possession states and scoring routes recur. An xP relationship trained on an earlier Premiership season would not automatically retain the same calibration in that Australian scoring environment. Changes in tactics, officiating, player availability, and opposition quality alter the situations feeding the average.
Score and time can also affect the meaning of an action. A penalty goal worth 3 points has the same numerical reward throughout a match, but its strategic value changes when a team trails by 5 with 2 minutes left. Expected points can describe the average scoring value of the possession while missing the match objective that governs the decision.
- Four Checks Before Quoting xP
An xP figure needs four pieces of information beside it:
- Game state: field position, possession, phase type, score difference, and time remaining.
- Outcome set: tries, conversions, penalties, drop goals, no score, and any opposition score included by the model.
- Weighting: each outcome’s probability multiplied by its point value, with the products added together.
- Model boundary: the competition, seasons, sample, and variables used to estimate those probabilities.
The next event belongs outside that calculation. A try does not prove the possession was worth 7 points in advance, just as a turnover does not prove it was worth nothing. When the model, sample, and game state are visible, xP can compare possession quality. When only the decimal survives, it carries less information than the match situation it replaced.
Image Source: unsplash.com
