Successful Key Moment (SKM): A Football Metric for Real Match Influence

Key takeaways

• Most football metrics still overweight final outcomes such as goals and assists - even advanced ones inherit this bias.
• Existing frameworks like xT and VAEP address parts of this problem, but none combine difficulty, context, role, and downstream influence in a single unified score.
• SKM is designed to score actions by expected value, context, pressure, role expectations, and downstream match influence.
• The metric is intended to better reflect role-sensitive contribution across systems and positions.
• The real value is not replacing existing metrics but connecting them to how people actually watch football.

1) Why Another Football Metric Is Needed

Football analytics has become highly effective at quantifying certain kinds of value: shot quality, passing progression, possession value, defensive actions, and on-ball involvement. Yet a major gap still remains between what analysts, coaches, and viewers see in a match and what mainstream data systems tend to reward.

Goals and assists are still culturally dominant. Even more advanced metrics often inherit the same bias by remaining tightly centered around terminal events - shots, final passes, or possession endings. But football is a chain-reaction sport. Many of its most important moments are neither the final action nor a conventional box-score event.

A midfielder slowing tempo under pressure, a center-back stepping into space to break a line, a full-back making an underlapping run that distorts shape, or a forward occupying two defenders to create a lane for someone else - these are often decisive moments in the structure of the match without appearing decisive in the stat sheet.

That is the motivation behind Successful Key Moment (SKM): a metric that attempts to quantify the quality and influence of important actions based not only on what happened immediately, but on what was expected, what the context demanded, and what changed afterward.

2) How SKM Relates to Existing Work

Before introducing SKM’s architecture, it is worth being precise about what already exists and where the gaps remain. This field is active, and SKM should be understood as building on prior work rather than ignoring it.

Expected Threat (xT) - Singh, 2019

Karun Singh’s xT model divides the pitch into a grid and assigns each cell a probability that a possession starting there will yield a goal within the next few actions. Passes and carries are then valued by the xT difference between their start and end zones. xT is elegant and interpretable - its biggest strength is that non-technical audiences can grasp it immediately.

Its limitations are well-documented. It is purely location-based: two passes covering the same pitch zones score identically regardless of the pressure on the passer, the speed of the move, or the game state. It cannot value defensive actions. It treats a 1-0 pass in the 90th minute the same as one at 3-0 in the 20th. SKM’s difficulty (D) and context (C) components are a direct response to these gaps.

VAEP - Decroos, Bransen, Van Haaren & Davis, 2019

VAEP (Valuing Actions by Estimating Probabilities) is the closest existing cousin to SKM and deserves direct engagement. Rather than grid-based zones, VAEP trains a classifier on the entire action sequence context - action type, location, game state, recent history - to estimate how each action shifts the short-term probability of scoring or conceding. It covers all on-ball action types, not just passes and carries.

VAEP is a significant advance. It already does a lot of what SKM proposes to do. The comparison paper by Van Roy, Robberechts, Decroos & Davis (2020) shows that VAEP more accurately values risk and contextual failure than xT, because it can assign negative value to actions that shift possession probability against the team.

Where VAEP falls short, and where SKM attempts to go further, is on three dimensions:

1. Difficulty relative to expectation: VAEP measures what happened to match state. It does not isolate how improbable the successful action was relative to what a typical player in that situation would achieve. A line-breaking pass through a compressed block and a routine sideways switch can produce the same ΔP. SKM’s difficulty multiplier (D) explicitly captures this gap.

2. Role-based normalization: VAEP applies the same framework to every player regardless of tactical role. A holding midfielder and an attacking midfielder are compared on the same distribution. SKM’s role/system factor (R) adjusts for what is expected given a player’s role, formation, and system - rewarding deviation that consistently improves match state, rather than treating all high-value actions as equally impressive regardless of how naturally they fall within a player’s responsibilities.

3. Interpretable decomposition: VAEP produces a single number. SKM produces a decomposed score - V, D, C, R, and ΔP are visible and attributable. This matters for communicating findings to coaches and scouts who need to understand why an action was rated highly, not just that it was.

OBV (On-Ball Value) - StatsBomb

StatsBomb’s proprietary OBV metric follows a similar philosophy to VAEP: every on-ball action is valued by its effect on the probability of the team scoring or conceding. OBV is used internally by clubs and has been partially described in public research, but its full methodology is proprietary. SKM is designed as an open, reproducible framework - the full pipeline should be replicable from public event data.

DxT - 2025

A 2025 paper introduced DxT (Dynamic Expected Threat), which extends xT by incorporating real-time off-ball player positioning. Rather than assigning static zone-level threat values, DxT adjusts them dynamically based on where defenders and attackers are at the moment of the action. This is directly aligned with SKM’s Layer 3 (tracking and pressure modeling). DxT performs worse than VAEP on global player ranking tasks but is more interpretable for coaching-level tactical analysis - which is exactly the use case SKM also targets.

xSuccess - Paul, Klemp & Memmert, 2026

The most recent adjacent work isolates completion probability as a correction to outcome-driven models like VAEP. xSuccess estimates how likely a given action was to succeed given context, then adjusts the raw VAEP value accordingly. A risky action that happened to succeed is deflated; a well-executed action that failed due to variance is partially credited. This overlaps meaningfully with SKM’s difficulty multiplier. The difference is that SKM incorporates difficulty as one of several adjustments, rather than as a correction applied post-hoc to a VAEP-derived score.

Where SKM fits

SKM is not claiming to outperform VAEP on player ranking accuracy - that is an empirical question that requires building and testing the model. The claim is narrower and more tractable: that by combining role-normalization, contextual weighting, and an explicit difficulty term into a single decomposable framework, SKM produces scores that are more interpretable to coaches and scouts, fairer across positions, and more aligned with how football observers actually evaluate influence.

3) The Core Idea Behind SKM

SKM is built on a simple principle:

An action should be valued not only by whether it succeeded, but by how difficult it was, how contextually important it was, and how much it changed the likely future of the possession or match state.

This creates a bridge between:

• event data
• tracking-derived context
• tactical role awareness
• match situation pressure
• downstream impact

Instead of treating all successful passes, dribbles, presses, recoveries, or carries as equal, SKM asks:

• What was the expected outcome of this action in this situation?
• How much pressure or risk was present?
• How valuable was this moment in the context of game state?
• What happened next because of this action?
• Was the player acting inside or outside normal role expectations?
• How much of the change in expected team advantage can be attributed to this moment?

That is what makes SKM closer to a match-influence framework than a conventional stat.

4) A Proposed Mathematical Framing

SKM can be framed in two complementary ways, which illuminate different aspects of the metric.

4.1 Additive framing (for transparency and debugging)

\[SKM_i = \Delta P_i + \alpha \cdot D_i + \beta \cdot C_i + \gamma \cdot R_i\]

Here, ΔP is the primary signal - the downstream change in expected positive outcome. Difficulty, context, and role are adjustments that augment the baseline. This formulation makes it easy to decompose any score and explain which factor drove it. It is the recommended starting point for implementation.

4.2 Multiplicative framing (for interaction effects)

\[SKM_i = \Delta P_i \times D_i \times C_i \times R_i\]

The multiplicative version captures the intuition that difficulty should amplify value rather than merely add to it. A difficult action in a critical game state under high pressure should score meaningfully higher than a routine action under the same conditions - not just additively higher. The trade-off is that a zero in any factor zeroes the whole score, which creates fragility. This version makes more sense once the components are well-calibrated and validated individually.

Component definitions

• ΔP_i: downstream change in expected positive outcome probability (scoring minus conceding), estimated from action sequences - the anchor of the metric
• D_i: difficulty multiplier, derived from completion probability models - how improbable was success given the context
• C_i: contextual importance multiplier, incorporating game minute, scoreline state, opponent quality, field zone, and pressure phase
• R_i: role/system adjustment, reflecting how expected or surprising this action type is for this player’s role and tactical context

5) Component Design

5.1 Baseline action value and ΔP_i

Rather than assigning manual baseline values to action types, SKM grounds itself in the downstream probability shift. This is computed as the difference in expected positive outcome probability before and after the action, estimated over the next N events in the sequence (typically 10-15 actions, following the VAEP approach).

Actions worth capturing beyond the standard pass/shot vocabulary include:

• line-breaking passes
• progressive carries into tight space
• recoveries under pressure
• press-resistance actions
• defensive cover movements
• aerial duel wins in transition
• off-ball decoy runs (estimated from tracking data where available)

The ΔP model can be trained using a gradient-boosted classifier on event sequence data, with StatsBomb open data providing a viable public starting point.

5.2 Difficulty multiplier (D_i)

Difficulty is central to the correction SKM makes over existing metrics. A sideways pass with no pressure is not equal to a same-distance pass through a compressed block.

Difficulty is estimated as the inverse of a completion probability model - how likely was success given this context? Inputs:

• under-pressure flag and pressure distance
• number of nearby defenders (from 360 data where available)
• passing lane width and obstruction
• body orientation and weak foot usage
• speed of play and available support options
• zone-specific turnover base rate

D_i should be normalized to a bounded range (suggested: 0.5 to 3.0) to prevent outlier actions from distorting season-level aggregates.

5.3 Context multiplier (C_i)

Context includes:

• match minute (increases in weight as the match tightens)
• scoreline state (drawing or trailing by one at home inflates stakes)
• tournament and league importance
• possession phase and field zone
• opponent quality
• current pressure wave and fatigue state
• home versus away conditions

A successful action in the 88th minute at 1-1 against a top-four side in a compressed defensive block is not the same as the same action at 3-0 in open play. C_i should be normalized to a bounded range (suggested: 0.5 to 2.0).

5.4 Role/system factor (R_i)

One of the persistent weaknesses in football evaluation is applying identical standards across all positions. A center-back, a holding midfielder, and a winger should not be assessed against the same action-value distribution.

SKM’s role factor works in two steps:

1. Cluster players into role archetypes based on their per-90 action profiles - roughly six to eight clusters (deep defender, ball-playing CB, holding mid, progressive mid, wide attacker, central attacker, and hybrids).

2. Compute role-expected action distributions for each cluster, then score each action by how surprising it is relative to role expectations. A center-back completing a line-breaking progressive pass into the final third earns a higher R_i than a midfielder doing the same. A striker shooting is role-expected and earns R_i ≈ 1.

The adjustment should be bounded (suggested: 0.8 to 1.5) to avoid rewarding tactical chaos. The goal is to credit genuine cross-role influence, not to penalize specialists.

5.5 Downstream effect (ΔP_i)

This is the most important component.

Rather than rewarding only direct endpoints, SKM measures the change in expected positive outcome probability after the action. That includes changes in:

• probability of entering the final third
• probability of generating a shot
• probability of retaining control under pressure
• probability of stopping opposition progression
• probability of shifting match control

ΔP is the “future effect” term that moves the metric beyond shallow event counting. It should be trained on the full action sequence, not just the immediate next event.

6) A Worked Example

To make the above concrete, consider the following action from a hypothetical Champions League knockout match:

Situation: Minute 76. Score 1-1. A holding midfielder receives the ball in his own half under pressure from two forwards. He turns away from pressure and plays a line-breaking pass through two defensive lines to a forward’s feet in the opponent’s half. Possession is retained.

Component	Score	Reasoning
ΔP_i	+0.07	Pass moves team from low-threat defensive zone into high-threat transition; estimated from sequence model
D_i	2.4	Under heavy pressure, line-breaking pass through compressed block - completion probability ~30% in this context
C_i	1.8	76th minute, tied score, knockout match, high pressure phase
R_i	1.1	Holding midfielder playing line-breaking pass - within role but toward the impactful edge of role distribution

Additive SKM: 0.07 + (0.3 × 2.4) + (0.3 × 1.8) + (0.3 × 1.1) = 0.07 + 0.72 + 0.54 + 0.33 = 1.66

Multiplicative SKM: 0.07 × 2.4 × 1.8 × 1.1 = 0.333

By contrast, a statistically identical pass - same start and end zones, same action type - made at 3-0 in the 20th minute with no pressure and no defenders nearby would score:

Additive SKM: 0.07 + (0.3 × 0.8) + (0.3 × 0.6) + (0.3 × 1.0) = 0.07 + 0.24 + 0.18 + 0.30 = 0.79

The first pass is approximately twice as valuable - even though both produced the same ΔP. That gap is the point of SKM.

7) SKM as a Bridge Between Metrics and the Eye Test

People often say a player “ran the game,” “kept the team alive,” or “changed the momentum” even when the stat line looks ordinary. That language is not irrational. It reflects hidden structure.

The eye test works because humans are good at perceiving:

• pressure release
• rhythm control
• shape manipulation
• threat escalation
• game-state intelligence

But the eye test is subjective, inconsistent, and often biased toward reputation. SKM’s purpose is not to romanticize intuition. It is to formalize its strongest parts in a measurable way.

In that sense, SKM is not anti-data. It is an argument for better data logic.

8) Modeling Pipeline

A practical SKM system can be built in layers, with each layer adding meaningful power while remaining independently useful.

Layer 1: Event-based prototype

Use public event data (StatsBomb open data, FBref) to create a first approximation with the core action types: passes, carries, interceptions, duels, shot-creating actions, turnovers, and defensive recoveries. Compute ΔP from event sequences using a gradient-boosted classifier. Build initial versions of D from the under-pressure flag and difficulty proxies available in event data.

This version is limited but publishable and useful for initial validation.

Layer 2: Possession value integration

Integrate established expected possession value models:

• xT for attacking zone transitions
• VAEP-style sequence classifiers for ΔP estimation
• OBV-like state transition values as a cross-check

This improves ΔP accuracy significantly, particularly for defensive and transition actions.

Layer 3: Tracking and pressure modeling

Add tracking-derived inputs where available:

• defender density at point of action
• pressure cones and body orientation
• pass lane obstruction and off-ball structure manipulation
• team compactness disruption

StatsBomb 360 data provides partial tracking coverage for selected competitions and is publicly available for some datasets. This layer unlocks most of SKM’s intended value for the difficulty component.

Layer 4: Role-aware normalization

Model expected action distributions by role cluster, formation, coach/system, game state, and opponent type. This prevents the metric from systematically rewarding attacking positions, which produce more high-ΔP events by default, and preserves tactical realism in cross-position comparisons.

9) Example Use Cases

Player evaluation

SKM surfaces players who consistently improve team state without dominating visible output - central midfielders, ball-playing center-backs, pressing forwards, and hybrid defenders whose contribution rarely appears on a traditional stat line.

Recruitment

Recruitment departments want players whose influence survives context shifts and opponent quality changes. SKM can surface players who make difficult positive actions repeatedly, operate effectively under pressure, and elevate possession states without needing high-volume final-third output.

Tactical review

Coaches can use SKM to examine where important state changes are created, which actions break structure, and whose decisions stabilize or destabilize key phases.

Broadcast and media

A broadcast-ready version of SKM could improve football storytelling. Instead of centering all narratives on goals and assists, analysts could explain - with data - why one action in the 77th minute mattered more than three passes in the 30th.

10) Guardrails and Limitations

Attribution is difficult

Football is heavily interdependent. Many high-value actions are enabled by earlier positioning, movement, and spacing that precede the credited event. Over-attributing value to a single touch weakens the framework.

Data availability

A public-data version of SKM will be meaningfully less powerful than a tracking-integrated version. The difficulty component in particular depends on defender density and pressure geometry that event data can only approximate.

Tactical bias

Any metric risks rewarding styles that are easier to measure. Low-tempo controllers, off-ball manipulators, and non-ball-dominant defenders remain hard to capture fully. SKM’s Layer 3 and Layer 4 are designed to reduce this bias, but not eliminate it.

Interpretability matters

A metric this layered cannot become a black box. Coaches, scouts, and fans need understandable decompositions: why was this moment scored highly? Which factor drove the score - pressure, difficulty, context, or downstream effect? The additive formulation exists specifically to preserve this transparency.

11) Why This Matters

Football has enough metrics that describe events. What it still needs are metrics that describe influence.

SKM is an attempt to move from:

• final action to action chain
• outcome to expected value shift
• raw event count to tactical significance
• stat line to match reality

That is the deeper opportunity in sports analytics: not just measuring more, but measuring more truthfully.

12) Future Directions

The most interesting extensions from here:

• graph-based possession influence modeling to capture multi-player coordination
• role-conditioned neural sequence models for ΔP estimation
• momentum-sensitive state value that responds to run-of-play patterns
• off-ball credit assignment using tracking data
• real-time broadcast overlays for influential moments
• player style clustering based on SKM component signatures

A production version could eventually offer:

• per-match SKM timelines
• role-adjusted season leaderboards
• “hidden influence” scouting dashboards
• tactical breakdowns by phase of play

13) Final Thought

A great football metric should not flatten the sport into numbers. It should make the numbers worthy of the sport.

That is the ambition behind Successful Key Moment: a framework that respects the complexity of football while making its invisible contributions more legible.

References

• Singh, K. (2019). “Introducing Expected Threat (xT).” karun.in/blog
• Decroos, T., Bransen, L., Van Haaren, J., & Davis, J. (2019). “Actions Speak Louder Than Goals: Valuing Player Actions in Soccer.” Proceedings of the 25th ACM SIGKDD Conference on Knowledge Discovery & Data Mining.
• Van Roy, M., Robberechts, P., Decroos, T., & Davis, J. (2020). “Valuing On-the-Ball Actions in Soccer: A Critical Comparison of xT and VAEP.” AAAI Workshop on AI in Team Sports.
• Fernández, J., Bornn, L., & Cervone, D. (2019). “Decomposing the Immeasurable Sport: A Deep Learning Expected Possession Value Framework for Soccer.” MIT Sloan Sports Analytics Conference.
• Paul, Y., Klemp, M., & Memmert, D. (2026). “Beyond Outcome Bias: Incorporating Action Completion Probability and Risk-Return Into Soccer Evaluation Models.” Machine Learning and Data Mining for Sports Analytics (MLSA 2025), Communications in Computer and Information Science, vol. 2833.
• Pleuler, D. (2021). Soccer Analytics Handbook. GitHub.
• StatsBomb. (2023). “Measuring the Value of Actions in Football.” StatsBomb Blog.

Related Work

• Severance-Style Segmentation and Geo-Visualization
• Interactive Cultural Timeline Map
• Indian Media Strategy Shift