The Defense Problem
Why WAR doesn't work for soccer yet, and what we're learning on the way
October 2026 · Draft Chalkboard
We've spent a lot of time trying to build a WAR (wins above replacement) model for soccer. We've built off-ball positioning models, on-ball action models, and xG systems, but it just isn't coming together. It's just too difficult to attribute individual actions to players in such a team dependent sport. That doesn't mean we aren't getting closer. Here's what we're learning and why we need more data.
The WAR Instinct
Baseball WAR works because the sport decomposes into isolatable confrontations. Pitcher vs. batter is literally one-on-one. You can slice a team's total production into individual pieces, and the pieces add up. That's what makes WAR possible: the sport's structure supports individual attribution.
Soccer offense has a version of this. The ball flows through individuals (passes, dribbles, shots, carries), so you can trace value through touches. This is what American Soccer Analysis's goals-added (g+) model does, and it works: over 125 NWSL team-seasons (2016–2026), offensive g+ predicts league points at r = 0.810 (R² = 65.6%).
Defense is different. The best defensive contribution is usually preventing an attack from developing at all, and it's difficult to assign individual value to invisible actions. Let alone invisible actions that rely on your nine other outfield teammates' positions.
g+ Interrupting and the Volume Problem
g+ measures defense through a single category called Interrupting. Every on-ball defensive action (tackles, interceptions, clearances, blocks, recoveries, pressures) is rolled into one number. Each action is valued by where it occurs on the pitch and what happens next.
The problem is that this framework captures volume, not quality. On the offensive side, touching the ball more in dangerous areas IS the skill. On the defensive side, needing to make more tackles and interceptions often means the defense has already failed upstream.
Breaking it down further: g+ tracks seven action types (shooting, dribbling, passing, receiving, fouling, interrupting) for both your team and opponents. When you add them one at a time in a stepwise regression to predict league points, shooting alone explains 64.7%. The additional variance comes from opponent dribbling quality against you (+7.0%), your team's foul discipline (+2.6%), and opponent passing quality against you (+2.5%). Interrupting never enters the model. Your team's defensive actions, as measured by g+, add virtually nothing.
Volume bias shows up directly: a team's own Interrupting g+ correlates positively with goals conceded (r = +0.413 across 125 team-seasons). Teams that accumulate more defensive action value tend to concede more goals, because making lots of tackles and interceptions means you are spending a lot of time under pressure.
ASA has been open about this limitation and has tried to address it. In 2022, they introduced g- (goals subtracted), which allocates defensive blame by dividing the field into 30 zones and estimating where each position should be based on league-wide frequency data. In 2023, they combined g+ and g- into net goals added. One of their findings: players who rate highly on defensive g+ (interrupting skill) do not tend to rate highly on g- (positioning quality). The two dimensions measure different things. Both papers acknowledged the core limitation: g- estimates where defenders should be without knowing where they actually are. We found something that might help with that.
What We Tested on 360 Data
We started with the only publicly available dataset that shows where defenders are standing during play: StatsBomb's 360 open data. This is 426 matches where every event includes a freeze frame with the location of all visible players on the pitch. The matches are a grab bag. Men's and women's international tournaments (World Cup, Euros, Copa América), some MLS, some domestic cups. Different levels of play, different tactical environments, different eras. It's not a controlled sample. It's what exists publicly with 360 coverage, and we used it to test whether off-ball defensive positioning contains any signal at all.
We tested every team-level defensive metric we could construct: compactness, defensive line height, back line gaps, pressing intensity, transition response speed, counterpress rates, pass lane coverage, and shot-moment positioning. Most of them turned out to be proxies for territorial control rather than defensive quality. Teams that pressed higher and stayed compact were winning teams, not necessarily better-defending teams.
Two categories survived. Pass lane coverage measured whether defenders were positioned within 3 yards of progressive passing lanes during buildup play. Passes into covered lanes completed at 77% vs. 96.5% into uncovered lanes, and teams with better coverage conceded fewer shots. Shot-moment positioning measured where defenders were standing when a shot occurred: goal-side rate, in-lane rate, nearest defender distance. Together with transition defense (what fraction of turnovers lead to opponent shots, computed from event sequences), off-ball defense added 11 percentage points of explained variance in match outcomes beyond offense alone.
The problem was getting any of this to the individual level. Pass lane coverage and buildup positioning require full 360 frames, which don't identify players by name outside of shot moments. We tried working backward from shot freeze frames (where players are identified) to track individuals through earlier 360 frames in the same possession. It got sloppy fast. Players merge, swap positions, and move unpredictably between frames. The tracking broke down within seconds.
So pass lane coverage is stuck at the team level for now. Shot-moment positioning and transition defense, on the other hand, don't need 360 data at all. Shot freeze frames are part of standard StatsBomb event data, and transition defense is computed from event sequences. To build individual profiles, we moved to NWSL 2023 (173 matches across one league season): the same players, the same teams, enough matches per player to test whether their habits are real or noise.
Positioning and Outcomes, Not Volume
Off-ball positioning at shot moments
This is conceptually what g- was trying to do, but with actual positions instead of estimates. Shot freeze frames show every player's location when a shot is taken. For each non-penalty shot in the NWSL 2023 season (4,515 total), we extract every identified defender and measure three things. Goal-side rate: is the defender between the shooter and the goal? In-lane rate: are they goal-side and within 3 yards of the shooting lane? DGS per shot: how much does their position reduce xG, computed by rerunning the model with that defender removed?
None of this is conceptually new. These are standard inputs to xG models. The contribution is flipping the question: instead of using positioning to score shots, we ask whether individual defenders are consistentlywell-positioned across all the shots they face. The answer confirms g-'s intuition: positioning quality is a repeatable individual skill (split-half r = 0.68–0.87) that predicts GA at the team level.
On-ball defensive action outcomes
Press no-shot rate and transition no-shot rate each ask the same question: when a player pressures during a possession or is the first defender to respond after a turnover, does the attacking team get a shot off?
Both on-ball metrics are zone-normalized. A press at your own goal line is a very different situation than one at midfield, and the league-average no-shot rate varies substantially by zone. We divide the pitch into 12 ten-yard bands and compute each player's rate relative to the league baseline for the specific zones where they actually press and respond. This prevents defenders who play for high-possession teams (and therefore press in less dangerous areas) from being artificially inflated.
Are These Real Skills?
| Method | Split each player's matches into odd and even halves |
| Question | Does their rate in one half predict their rate in the other? |
| High r | Metric captures a persistent individual trait |
| Low r | Metric is noise |
| Metric | Category | Split-half r |
|---|---|---|
| Goal-side rate | Off-ball | 0.868 |
| In-lane rate | Off-ball | 0.781 |
| DGS per shot | Off-ball | 0.679 |
| Press no-shot rate | On-ball | 0.515 |
| Transition no-shot rate | On-ball | 0.406 |
NWSL 2023, 173 matches. Minimum thresholds per half: 20 shots (off-ball), 25 presses, 10 transition responses.
The off-ball positioning metrics are the most stable (0.68–0.87), which makes sense: positioning habits are deeply ingrained. A center back who consistently gets goal-side does it in every match. The on-ball outcome metrics are moderately stable (0.41–0.52), reflecting the fact that pressing and transition outcomes depend partly on teammates and opposition.
Head-to-Head: Our Metrics vs. g+ Interrupting
The core question: which approach better predicts how many goals a team concedes? We compared g+ Interrupting per game against each of our metrics on the same 12 NWSL 2023 teams, correlating each with goals against per game.
| Metric | r vs GA/game | R² |
|---|---|---|
| g+ Interrupting / game | −0.032 | 0.1% |
| Transition no-shot rate | −0.451 | 20.3% |
| Press no-shot rate | −0.360 | 13.0% |
| Avg in-lane defenders / shot | −0.361 | 13.0% |
| Avg goal-side defenders / shot | −0.350 | 12.2% |
| Avg DGS / shot | −0.285 | 8.1% |
| Combined composite (5 metrics) | −0.715 | 51.2% |
NWSL 2023, N = 12 matched teams. Combined composite is the average position-adjusted z-score across all five metrics. GA and g+ data from ASA; defensive metrics from StatsBomb event data.
Every individual metric outperforms g+ Interrupting on its own. The weakest (DGS per shot, 8.1%) still explains 80x more GA variance than Interrupting (0.1%). The combined composite of all five reaches 51.2%.
What Defense Adds to the Points Model
The previous section shows our metrics predict goals against better than g+ Interrupting. But the article opened with a bigger question: g+ offense explains 65.6% of league points. Defense adds nothing. Do our metrics change that?
We correlated each metric with league points for the same 12 NWSL 2023 teams:
| Metric | r vs Pts/game |
|---|---|
| g+ Interrupting / game | −0.360 |
| Our composite (5 metrics) | +0.591 |
| Press no-shot rate | +0.761 |
| Transition no-shot rate | +0.666 |
| Goal-side rate | +0.119 |
| In-lane rate | +0.097 |
| DGS per shot | −0.029 |
NWSL 2023, N = 12 matched teams. Positive r = more points.
Our defensive composite correlates with points at r = +0.591. Press no-shot rate alone reaches r = +0.761. g+ Interrupting points the wrong way: teams that accumulate more interrupting value tend to win fewer games, consistent with the volume bias described above. When we combine offensive g+ with our defensive composite in a two-predictor model, the model explains substantially more variance than offense alone. Replacing our composite with g+ Interrupting actively hurts the model.
These numbers come from a single season of 12 teams and will shift with more data. Offensive g+ is noisy on small single-season samples too (R² = 65.6% across 125 team-seasons, but much weaker within a single year). The directions, however, are consistent with every other test: outcome-based defensive metrics add real signal where action-volume metrics do not, and both offense and defense point the right way.
Not WAR Yet
We haven't given up on WAR for soccer. We've prototyped combining ASA's offensive g+ with our five defensive metrics at the individual level, and the rankings look plausible. Defense meaningfully reshuffles player order: positioning-first defenders climb the table, volume-first defenders fall. The directions make sense.
But the calibration isn't trustworthy yet. Setting the relative scale between offense and defense requires regressing both against the same outcome, and we only have one season of matched data (12 teams) for the defensive side. Offensive g+ is well-calibrated across 125 team-seasons. Our defensive composite has one year. Averaging across sample sizes gets you a prototype, not a model.
The data gap is real too. Shot freeze frames show defender positions at the moment of a shot, but not during the other 95% of the match. Full 360 tracking would let us measure off-ball positioning throughout entire possessions, not just at shot moments. We had 173 NWSL matches with shot freeze frames and 426 international matches with full 360 coverage. Neither dataset is large enough or consistent enough to build a complete individual valuation model from. More seasons of NWSL event data would let us calibrate properly. Until then, the five metrics tell you which defenders are doing the right things. They just can't tell you exactly how much it's worth.
Event data from StatsBomb Open Data (NWSL 2023, 173 matches with shot freeze frames). Goals-added and standings data from the ASA public API (125 NWSL team-seasons, 2016–2026). ASA defensive model references: Introducing Goals Subtracted (2022) and Net Goals Added (2023). Defensive xG model trained on 11,862 shots.