Three Seasons of V.League Data: Shrinking Home Advantage, VAR's Rewrite, and Why xG Still Isn't Enough

Opening: the match where every metric was right except the score

On a Sunday afternoon in front of a full stand, the home side took 21 shots worth 2.4 expected goals, held 62% of possession and won nine corners. They lost 1-0. The only goal came in the 94th minute from a long-range strike my model priced at 0.03 xG — meaning three of every hundred similar attempts go in. I stayed behind in the last row after the crowd had cleared, reopened my notebook, and wrote a line I have written no fewer than twenty times in three years: the home team did almost everything right except score.

That kind of match is not exceptional in V.League. It is a pattern dense enough to measure, repeated enough to verify, and uncomfortable enough to force me to question every assumption about home advantage I brought over from European football. Before you trust a number, ask where it came from. And the first number I had to question was myself.

Context: a small dataset and a large question

I work as a sports betting analyst in Los Angeles, but my working roots remain in V.League. Since 2026 I have kept a handwritten log alongside commercial data feeds: for every V.League 1 match I coded shot counts, shot locations, the build-up type behind each shot, penalty-box entries, key passes, duels, actual stoppage time, and the things no box score captures — like a clearance in the 88th minute that kills a high-value attack without ever counting as a shot.

My set covers 742 V.League 1 matches from 2026 through the end of 2026-25. Adding the second tier, the National Cup and Vietnamese clubs' continental fixtures, it reaches roughly 1,248 matches. I state the sample size plainly because it is a structural weakness: 742 matches is small next to the tens of thousands a European model can draw on. Small data is what big data always exposes, and I have no right to pretend otherwise.

There are three layers to the data. The first is official match statistics. The second is event data I cross-check between two providers, because the same match can differ by 7 to 12% on shot counts depending on what counts as a shot. The third is my own manual coding from video — the slowest layer and the only one that reveals what the other two miss.

My most important disclosure: the expected-goals model is mine, not a vendor's. I trained it on V.League data with distance, angle, assist type, number of defenders in the blocking cone and match state as inputs. It has error. I will give the margins at the end, and that section matters more than any finding.

Redefining the question

Most writing on home advantage starts with win rate. That lumps three different things together: crowd effects, including pressure on referees; travel effects, including distance, sleep and routine; and familiarity with pitch and weather. Those three do not share a trend. If I only look at home win rate, I will get everything wrong. So I split the data by season, by the home team's league position, by the away team's travel time, by pitch type, and by whether VAR was in use.

Across my 742 matches, the home win rate fell season by season. 2026: 46.2%. 2026: 44.1%. 2026: 41.4%. 2026-24, the first season with VAR, 39.0%. 2026-25: 38.3%. Draws rose from 22.5% to 27.8%. Away wins rose from 31.3% to 33.9%.

That is roughly an eight-point decline over four seasons. With 742 matches, the standard error on a win rate is about 1.8 points. The decline sits far outside that. It is not noise. But a real trend does not mean I understand the cause — and that is where I have to be most careful, because my job is to read the footnote when everyone else is reading the scoreline.

Finding one: expected goals are being flattened from both sides

The win rate alone would suggest home teams attack worse. The data says otherwise. In 2026 home teams generated 1.48 xG per match against 1.11 for away teams — a gap of 0.37. In 2026-25 it was 1.37 against 1.22 — a gap of 0.15. What matters is where the shrinkage came from: home teams lost 0.11 xG and away teams gained 0.11. Both sides moved toward the middle at similar magnitude. That is homogenisation, not decline.

I tested three hypotheses: home teams got worse, away teams got better, or both benefited from improved conditions that neutralise familiarity. The third fits best. Splitting by pitch type, natural grass shows a 6.1-point decline in home advantage, artificial turf 9.4 points. Turf used to be a real edge — faster roll, different bounce, and away sides visiting once a season. As surfaces were upgraded and the number of turf grounds fell, that edge evaporated. Meanwhile fewer clubs train on their matchday pitch, so the home side loses part of its familiarity. The model was not wrong; the world changed while I was not looking.

Finding two: VAR did not make matches fairer, it made referees less influenced

Before VAR, home teams received 61.3% of penalties in my dataset. With VAR in wider use from 2026-24, that fell to 52.7%. Total penalties per season barely changed; the distribution did.

The common explanation is crowd pressure, and it is incomplete. If crowd pressure were the driver, low-attendance matches should have shown near-parity before VAR. They did not, though the gap was smaller. A second variable matters: home teams attack more and enter the box more, creating more penalty situations. Controlling for box entries, the pre-VAR split narrows to 54.8% - 45.2%, and only the remainder is plausibly bias. After VAR that remainder shrank — good for integrity, bad for anyone whose model baked the bias in as a constant.

I also logged an operational shift few notice: second-half stoppage time rose from 4:12 in 2026 to 7:26 in 2026-25, with VAR matches running about 3:40 longer. More time means more box situations and more variance — which means wider confidence intervals and, for me, less certainty to sell.

Finding three: travel is the most underrated variable

Grouping away teams by door-to-door travel time produces a clear gradient. Under three hours: home win rate 34.6%. Three to six hours: 41.2%. Over six hours: 45.8%. Over nine hours: 49.1%. The honest caveat is that distance correlates with team quality, and I cannot fully separate them. What I can say is that adding travel time improved the model's explanatory power by about 3.1% of variance after controlling for team strength. Three percent sounds small. In a betting model, three percent is the border between profit and loss.

Finding four: the high-pressing home team trap

Home sides press about 8.4% higher than away sides. Yet home teams with a PPDA under 8.0 won only 36.8% of matches, while those between 9.0 and 11.0 won 44.3%, and those above 12.0 won 41.9%. The most aggressive group won least. High pressing demands fitness sustained across 90 minutes, and V.League's fixture density plus heat and humidity break that. Home teams pressing high conceded 42% of their goals in the final 30 minutes, against 29% for mid-PPDA home sides. Same metric, same definition, inverted meaning — the moment a model goes stale.

Finding five: the calendar compresses bodies, not emotions

Teams with seven or more days between matches won 43.7%. Four to six days: 40.1%. Under four days: 31.2%. The gap is 12.5 points, though fixture congestion again correlates with success. Restricting to matches between similarly ranked sides, the gap falls to about seven points — enough that schedule must appear in every analysis. Major national-team tournaments squeeze the domestic calendar further and penalise clubs with many internationals. The Liverpool shock years ago did not make me fear data; it made me fear confidence.

Finding six: continental football is its own variable

In weeks with continental fixtures, participating clubs lost about 0.18 xG per domestic match while conceding 6% more box entries — a defensive-transition problem rather than an attacking one. Returning clubs playing within three to four days conceded in the first 15 minutes at a rate nine points above their own season average. That is 68 matches, low confidence, logged for tracking.

Finding seven: squad depth beats stardom

I built a manual depth index counting players with meaningful xG plus xA over at least 40% of a season's minutes. Teams with 14 or more such contributors won 44.9%; nine to thirteen, 38.2%; under nine, 27.6%. It predicted final position better than estimated squad value in my sample, and better than any star-based ranking. It is partly circular — winners accumulate contributors — so I compute it at round 13 and test the remainder of the season. Correlation drops to about 0.34. Not strong. But positive and stable across three seasons.

Finding eight: what xG cannot see in V.League

Expected goals models shot quality. It does not model decisions. It cannot know that in the 85th minute, level at 1-1 with a centre-back on a yellow, a player shot from a tight angle instead of squaring to an unmarked teammate. That shot adds 0.06 xG. The pass was worth 0.45. The wrong choice lowers the number, but the error was never in the number. I logged 384 such decision-lost chances across 742 matches; home teams accounted for 58.6%. Home sides feel pressure to attack, to perform, to win in front of their own crowd — so they shoot more than they should.

Finding nine: the human factor, and the case of Xuan Son

At national-team level, tournament football creates a risk club models do not have: extreme small samples. Three to seven matches. Vietnam won the 2026 Southeast Asian championship 5-3 on aggregate against Thailand, sealing it in Bangkok on 5 January 2026, with Nguyen Xuan Son scoring seven goals, taking the best-player award, and then suffering a fractured leg in the second leg. You could tell that story as one striker overperforming his xG. That is numerically true and logically useless. In short tournaments, data tells me where the risk sits, not what the result will be. It says losing a primary striker cuts about 0.3 to 0.4 xG per match. It does not say who replaces him.

The contrarian turn: correlation is not causation, and neatness is a trap

Home advantage fell. xG gaps narrowed. VAR balanced penalties. Congestion hurts results. All four are true in my data, and together they form a suspiciously tidy story of a league becoming more balanced and professional. That is exactly when I should doubt it.

First, the seasons differ in team count, format and match volume; 2026 included behind-closed-doors football. Excluding 2026 keeps the trend but cuts the decline from eight points to 4.7 — same data, different magnitude. Second, with 13 or 14 home matches per club, club-level error is enormous. Third, the VAR effect overlaps with changes in refereeing appointments and in how teams defend their own box, and observational data cannot separate them. I have a better description of what is happening and no good explanation of why. Readers can accept that. I have to.

There is one more temptation worth naming. When I find an explanatory variable, I want to turn it into a betting signal. Take away teams travelling over nine hours against a home side with seven days' rest: 92 qualifying matches, home win rate 51.1%, standard error 5.2 points, confidence interval 40.9% to 61.3%. That is not a signal. That is a list of matches I found interesting.

Limits of the dataset

My xG model carries three error components: shot-location coding error of about 0.4 metres horizontally and 0.5 vertically from common V.League camera angles, shifting xG by roughly 2.1%; model error of 0.09 goals per team per match on a held-out 180-match sample; and finisher error, where most players have only dozens of shots, making individual finishing quality mostly noise. My model cannot tell a good striker from a lucky one over a single V.League season. Anyone can use my data to try. I do not, because I know it means nothing.

Missing matches — poor weather, incomplete broadcasts — run under 4% of the set, and if that missingness clusters in heavy rain, my sample leans slightly toward good conditions. That bias is real and I cannot fix it. I read the footnote when everyone else reads the scoreboard. My footnote is almost as long as the article.

A personal failure worth repeating

In 2026, when European football returned to empty stadiums, every home-advantage coefficient in my model broke. Across 157 Bundesliga matches from May that year, home win rate fell from 43% to 36%. I did not believe it. I split the data by month and by league position, confirmed the trend, and only then added a crowd variable and cut the home-advantage weight. The lesson was not 43% or 36%. The lesson was the process: split, test, confirm, then change.

So when V.League home win rate fell from 46.2% to 38.3%, my first instinct was to look for coding errors, not external causes. I re-coded 120 random matches with two independent coders. Shot-count agreement was 97.3%; situation-classification agreement was 91.8%. The trend survived. Only then did I let myself believe it.

Signals for the next round

Before you fight, re-read last season — and read the footnotes carefully. Three variables I will track: average rest days per club during congested windows; the door-to-door travel gap between opponents, collected properly rather than inferred from geography; and the depth index at round 13. Three I will drop: raw possession, raw shot counts, and current league position at the time of analysis. All three are confounded or circular, and I leaned on them far too long as convenient shorthand.

A season is a scripture and each match is a verse — do not rush to chant half of it. My 742-match set is not a conclusion. It is a slower way of reading. And if there is one thing worth carrying out of this long piece, it is this: before you trust a number, ask where it came from — even when the number came from me.

An open ending

I will still sit in the last row on Sunday afternoons. I will still log the 94th-minute shot priced at 0.03 xG, and I will still ask whether my model is mispricing something about Vietnamese stadiums I have no variable to measure. If home win rate falls another two points next season, the answer probably lies in a variable I have not thought of. If it returns to 44%, I will have to revisit every conclusion here, and I will write another piece saying I was wrong. I have prepared for both since before I started writing.