
Expected threat values some of the ball progression that happens before a shot. In a common location-based xT approach, a completed move receives the difference between the modelled values of its end and start locations. That can recognise a useful pass or carry even when the next action is not a shot.
The value is a model output, not the actual causal increase in a team’s chance of scoring in a particular situation. Read it with the model definition, action coverage and football context. It cannot establish, on its own, that a player made the best decision or that a team will score next.
Start with the version of the model
Karun Singh’s introduction develops a location-value approach to attacking build-up. Its important distinction is that a location can have value through future moves, not merely through shooting immediately from that spot.
The socceraction documentation describes an implementation for valuing successful ball-moving actions through the difference in location values. Check the implementation you actually use rather than assuming every statistic labelled xT covers the same events.
Record the provider or implementation, training data, pitch-coordinate convention and action filters where available. Different value surfaces or event definitions can produce different totals from what looks like the same match.
Some products add information beyond location, or use another action-value framework. A shared interest in build-up play does not make those outputs interchangeable. Preserve the product’s own name and definition instead of treating every possession-value statistic as xT.
Follow one invented move
Suppose the chosen model assigns 0.02 to the starting location and 0.05 to the destination. A qualifying successful move from the first to the second receives 0.05−0.02, or+0.03 xT.
These values are hypothetical. They do not describe a real fixture, a trained OptaTips model or an observed improvement in a specific team’s winning probability. They illustrate subtraction within a stated modelling convention.
The positive value means that the destination has a higher modelled attacking value than the origin. It does not mean that the pass itself was easy, that the receiver had no defender nearby or that a goal became inevitable.
A successful return move from 0.05 to 0.02 would receive−0.03 under the same location-difference convention. The negative sign describes the model’s value change, not a complete assessment of whether retaining the ball was sensible in that situation.
Why this is different from shot value
A pass can put a receiver in a position from which a later pass creates a strong chance. If the receiver does not shoot, that first move may not appear in a shot-based chance-creation column. A location-value approach is intended to describe a different part of the attacking process.
Expected goals, by contrast, assigns modelled value to recorded attempts under its own definitions. A useful progression sequence can occur without producing a shot, while a high-value shot can arise after very little progression.
The expected-goals explanation covers the limits of reading chance quality from a match total. xT can add a view of eligible ball movement; it does not repair every limitation of shot data.
Do not simply add xT and xG and call the sum a team’s true scoring probability. They can describe different stages of related possessions, so an unexplained sum risks counting parts of the same process more than once.
The action filter changes the interpretation
A successful-move-only total does not automatically charge the player for every failed attempt. A player who completes one valuable progression and loses several other balls may still look different once those losses are considered separately.
Check whether unsuccessful passes, turnovers, shots, carries and set pieces are included, excluded or treated through another model. Do not infer their treatment from a heading that merely says attacking contribution.
Retain attempted and successful action counts where available. A high total can reflect frequent opportunities, a high average value per completed move, or both. The total alone cannot identify the mixture.
Likewise, do not call a completed pass’s xT a pre-action estimate of how valuable attempting it was. A review of the decision before execution needs to consider available alternatives and success risk. A value attached after a successful move answers a narrower question.
Location is useful information, but incomplete information
Two receptions in the same grid location can occur under very different conditions. One receiver may face an organised defensive block; another may have space after a turnover. A simple location-only value surface need not distinguish them.
Body orientation, available passing options, pressure and the positions of other players can affect the football situation. Do not write as though a location-based model directly observed those details when they were not among its inputs.
That limitation does not make the model useless. It defines the evidence that supports the number. Use video, richer event information or tracking data when available to investigate what the summary leaves uncertain.
The field-tilt guide similarly separates territorial activity from chance quality. xT adds a modelled location-value change; the remaining context still needs attention.
Net movement and positive-only totals differ
Consider an invented sequence of successful moves from values 0.02 to 0.05, then to 0.04, then to 0.06. The signed changes are+0.03,−0.01 and+0.02. Their net sum is+0.04.
A positive-only sum would retain 0.03 and 0.02, giving 0.05. Neither number should be substituted silently for the other. The first preserves the negative movement; the second reports only increases.
For compatible consecutive location differences within this simplified sequence, the net sum equals the final value minus the initial value:0.06−0.02=0.04. That is a property of the arithmetic, not proof that all the football between those points was irrelevant.
Filtering events, splitting actions between players or resetting possessions can change the reported aggregation. A dashboard’s net, positive, received and created values may represent different operations. Ask for the definition before comparing their sizes.
Coordinate handling can create apparent threat
Location-based calculations require a consistent attacking direction. If one half’s coordinates are not transformed correctly, a harmless move can appear to progress towards goal or a genuine progression can appear to move away.
Check pitch dimensions and coordinate ranges too. A dataset using one scale cannot be mapped directly to a grid built for another without a valid conversion. A visually attractive heat map is not evidence that this conversion was correct.
Grid resolution affects the representation. Two points within one cell may share a value even though their positions differ; a small movement across a boundary can enter another cell. Interpolation and finer grids are modelling choices, not automatic guarantees of better judgement.
For a published review, state these methodological limits without pretending to have run checks that were not performed. If the supplied data lack start or end locations, do not invent them to complete an action-value table.
Keep role, volume and minutes visible
A deep passer, a ball-carrying midfielder and an advanced receiver can have different opportunities to add modelled location value. Compare their responsibilities before ranking a total as evidence of general superiority.
Season totals also reflect minutes and action volume. Per 90 rates standardise one exposure measure, while value per completed move uses another. A player can lead one measure and trail another without a contradiction.
Show the underlying sample beside the rate. One unusually valuable move in a short appearance can produce an impressive per 90 number with little evidence about repeatability. A larger sample is helpful only if its role, event coverage and model remain relevant.
Changing teams or competitions can alter both the opportunities and the model’s fit to the data. Do not assume a value surface learned in one setting is equally well calibrated in every other setting merely because the coordinates look familiar.
Do not turn description into an untested forecast
A team progressing the ball into higher-valued locations may give an analyst a useful question to investigate. It does not establish a guaranteed future advantage, a permanent attacking ability or a profitable market price.
A forecasting claim needs its own method, information cut-off and evaluation on later unseen matches. Preserve whether the xT values were available at that time; a revised historical model can make a retrospective analysis differ from what a forecaster actually knew.
Keep measured outcomes, model descriptions and hypotheses separate. The fact that a move received positive value does not tell you that every alternative was worse. Nor does a later goal prove that the model assigned credit perfectly.
Write the claim at the level the evidence supports
A reproducible xT review names the model, action filters, period, minutes and aggregation rule. It reports counts alongside totals and explains whether negative values and failed actions are included.
For the opening example, the supported statement is that an eligible completed move increased the chosen model’s location value from 0.02 to 0.05 and received+0.03. It should not be described as a measured three-percentage-point causal improvement in a real player’s scoring chances.
That distinction keeps a useful build-up metric from becoming an inflated claim. It lets xT explain what it was designed to describe while preserving the football and uncertainty outside the model.
For adults aged 18 and over. All numerical examples are hypothetical. No actual OptaTips model results, data-supply relationships or guaranteed outcomes are asserted.
Comments 0
No comments yet.