Tennis Has More Data Than Football — But Its Modelling Is A Decade Behind

Tennis Has More Data Than Football — But Its Modelling Is A Decade Behind

2:14 in the morning, Brisbane. I re-ran the extraction routine on my second monitor and got back a blank sheet. Empty title column. Empty player column. Empty surface column. Not a single scoreline, not a single timestamp, not a single tournament recorded. The only field with content was the notes column, and it repeated the same sentence on every row: insufficient information, cannot assess.

That was a bad night for a data pipeline. But it taught me something that years of reading the ATP rankings never did: the hardest part of this job is not finding an answer. It is knowing when you have not been given the ingredients to answer at all.

The next morning I went back through the tennis coverage running in the Australian market. A great many fans were waiting for a prediction. A few dozen accounts had already published a champion's name. None of them included a confidence interval.

I kept thinking about that blank column, and about a paradox nobody says out loud: tennis is among the most thoroughly recorded sports on earth, and yet it is the worst modelled of the major professional sports. The gap is not in collection. It is in interpretation.

The infrastructure arrived long ago. The analytical language did not.

To see why, look at the plumbing. Tennis digitised its courts early. Hawk-Eye appeared at Wimbledon in 2026 as a challenge-based line-calling system, and within a little over a decade the same system shifted from umpiring tool to competition infrastructure. The 2026 US Open was the first Grand Slam to roll electronic line calling out at scale. The 2026 Australian Open did the same across every court. By 2026, Wimbledon ended 147 years of line judges.

The data thickened accordingly. At Wimbledon, IBM has held the technology partnership since 2026, bringing SlamTracker and pre-built analytical layers for broadcast. In Melbourne, Infosys attached its name to the Australian Open's digital infrastructure from 2026. On the tour side, Tennis Data Innovations was set up as a joint venture between the ATP and ATP Media to centralise data and build commercial products from that very stream. On the integrity side, the International Tennis Integrity Agency (ITIA) was established in 2026, succeeding the earlier Tennis Integrity Unit, with a remit to monitor suspicious betting patterns.

Which means every ball at a major is recorded: timestamped, coordinate-mapped, speed-measured, shot-typed, attributed to a named player. A three-set match generates 180 to 250 points, each carrying five to ten attributes. Set beside football, where a match produces roughly a thousand coordinate-bearing events, tennis is not behind on density at all. Structurally, tennis data is cleaner: every point has a defined winner, no contest over possession, no vague notion of control.

So why does an analyst who came out of football, now covering tennis for the Australian market, feel like he is working with fewer tools?

The point is the atomic unit, but points are not independent

There is one fundamental technical difference between the two sports. In football, events are largely independent of one another: a pass in the 12th minute does not directly determine whether a team scores in the 78th. In tennis, every point depends on the point before it. The same serve at 40-0 and at 30-40 is two different events — psychologically, tactically, and in expected value.

This is what traditional stat sheets destroy. When you collapse every service point a player hits into one number, you have thrown away the most important variable: scoreboard pressure. A player can win 72 percent of service points across a match and lose every service point in the deciding game. The final box score will look entirely normal.

It took me a while to see this, because my habits came from football, where match-level aggregation is the default. In tennis, match-level aggregation conceals the most.

The genuinely useful metrics sit lower: first-serve points won, second-serve points won, return points won, break-point conversion. Those four numbers together already sketch an almost complete picture of a player in a given match. But they only mean something when read game by game, not match by match.

The mathematics of randomness in a single-elimination sport

Football produces about 2.7 goals a match. Tennis produces about 200 points. That sounds far more stable. But tennis results are not decided by points. They are decided by sets, and there are very few sets: three, or five.

At ATP level, the average service-hold rate on hard courts tends to hover between 78 and 82 percent. Among the elite servers it clears 88 percent. Most games therefore carry no information about the gap in quality at all — they simply confirm the structural advantage of the server. The difference between world No. 5 and world No. 45 is usually settled by three or four points across an entire match.

Public simulations show that the relationship between point-win rate and match-win rate is a very steep curve. A player winning roughly 55 percent of points can win close to 90 percent of matches in a best-of-three format. Drop the point-win rate to around 52 percent and the match-win rate can fall into the low-70s. A small error at the raw data layer is amplified into an enormous error at the conclusion layer.

This is why tennis is harder to predict than football match by match, and easier to predict across a season. A stronger football team can control 60 minutes and win 2-0. A stronger tennis player can win more total points than the opponent and still leave in the quarterfinals. This is the tennis version of the xG story I used to write about constantly in football: winning the underlying numbers, losing the scoreboard.

I remember an evening in Brisbane, rewatching a match in which the winner finished 14 total points behind. He won the break points. The entire margin sat inside four points. Those four points decided a place in the next round, a prize-money payment, and a few dozen ranking points.

The serve is the best-measured thing in the sport — and that is the problem

Since electronic line calling became ubiquitous, placement data has become trustworthy to the centimetre. That opened an analytical direction previously confined to the video rooms of a few academies: serve-placement maps by box, by spin type, by scoreline.

But precisely because the serve is measured so well, tennis analysis tends to reduce everything to it. Broadcasts talk about serve speed, first-serve percentage, ace counts. Meanwhile, the part that actually decides an elite match usually lives in the second-serve return and the first shot after the serve.

This imbalance has a practical cost. A player can be framed by the media as owning a top-tier serve while the decisive metric — second-serve points won — sits in the bottom half of the draw. The second serve is a measurable weakness, but it rarely makes the front page because it produces no highlight.

The best metric almost nobody uses

If one metric deserves to sit at the centre of every tennis report, I would argue it is the Dominance Ratio, DR, developed by analyst Jeff Sackmann on the Tennis Abstract data set.

The calculation is compact: a player's return points won percentage divided by that same player's service points lost percentage. It measures the ability to generate pressure on the opponent's serve without surrendering your own. Sackmann has shown that DR correlates with match wins better than raw total points won, because it separates control of the contest from the scoreboard.

This metric has almost never appeared on a broadcast graphic. A player sustaining a DR above 1.30 across a season is almost certainly in the title-contender bracket. A player sitting around 1.00 while winning a lot of tight matches is living on variance — and variance always sends an invoice.

The power of DR is that it does not care whether you won. It measures only whether you were controlling the points. A metric indifferent to results is a hard sell to broadcasters, and that is exactly why it gets ignored.

Surface, ball, climate: three variables nobody declares

Here tennis loses to football clearly on the ability to normalise the competitive environment.

Hard court is not one surface. It is a family of surfaces. Bounce depends on the coating, the humidity, the surface temperature, and the ball pressure. Melbourne in January can push surface temperatures past 50 degrees Celsius. Indian Wells sits in dry desert air and the ball travels faster. Miami's humidity makes the ball slower and heavier. Some European events are staged indoors under near-total climate control. No standard conversion exists to bring these conditions onto one scale.

The ball compounds it. Each Grand Slam has its own supplier, and every change of ball introduces a new variable into the data that nobody declares in the model. Players have repeatedly complained publicly about the ball used at the Australian Open across the 2026-2026 period, arguing it fluffed up quickly and became heavy within a few games. The tournament has reasons to change. The manufacturer has reasons to defend the product. The analyst is handed a data set with an undeclared extra variable.

Based on my experience tracking matches in Queensland and Victoria, I have adopted a habit: whenever the ball or the surface changes, I tag the following three months of data as non-comparable with what came before. Not because the data is wrong, but because the rules changed and the comparison lost its validity.

The 2026 lesson, applied to tennis

Ahead of the 2026 World Cup I built a forecasting model on six major tournaments' worth of data, using Elo and qualifying records. It ranked Brazil as the number-one contender at 23.4 percent. I was confident enough to write that the data had identified the champion. Brazil went out in the quarterfinals. France — which my model ranked fourth at 11.2 percent — lifted the trophy.

After the 2026 World Cup, I removed the word "certain" from my analytical dictionary entirely.

The flaw was not in the mathematics. It was in the variable list. I was missing club minutes played before the tournament, missing travel schedules, missing undisclosed injuries. The same class of error shows up in tennis, but it shows up more starkly, because there are fewer matches and each one carries more weight.

In 2026 I learned that a 95 percent probability still leaves 5 percent that knows how to laugh.

In tennis I have added two variables football does not have. The first is cumulative movement load: a player reaching a semifinal after two five-set matches carries an entirely different set of legs from one who got there in two straight-setters, even though the scoreboard shows nothing. The second is cross-time-zone scheduling: moving from Melbourne to Europe inside two weeks carries a physiological variable alongside the airfare.

Empty stadiums: the cleanest laboratory

In 2026 and 2026, tennis went through what football went through: competing in empty arenas. The 2026 US Open was played without crowds. The 2026 Australian Open staged days of competition under Victoria's lockdown orders, with the only applause coming from the public-address system and the organising staff.

I once spent an entire semester comparing data before and after football's return to empty stadiums, and the results forced me to rewrite a fair number of assumptions. For tennis, that window was a rare opportunity: the first data set in which the crowd variable was almost entirely removed while every other variable — surface, ball, rules — stayed largely fixed.

From the empty stadiums, I could hear the match breathing.

What stands out is not how the results shifted. What stands out is how much the crowd affects small tactical decisions: whether a player dares to come forward more, whether a player dares to hit the line on break point. In tennis, crowds do not shape the tempo of a match the way they do in football. They shape risk appetite.

The crowdless season was the cleanest laboratory sport has ever had. Its biggest lesson: a great many psychological traits commentators assign to players by default are in fact responses to the stadium environment.

Live data and the betting market

There is one dimension rarely put on the serious analytical table: most of the economic value of real-time tennis data is consumed by the betting market, not by fans and not by youth academies.

The market structure shows it plainly. A point is recorded within a few hundred milliseconds. Within that window, some systems have already repriced the odds. Television viewers get the number seconds later. Analysts get the whole data set after the match ends. Information value decays with time, and in tennis a delay of a few seconds means the entire value has already accrued to a very small group of users.

This is the darkest side effect of the digitisation of sport, and tennis shows it most clearly. Not because the data is bad. Because the priority order of distribution has been inverted: speed gets paid, transparency does not.

Integrity bodies such as the ITIA do their work, and they maintain teams monitoring anomalous betting patterns. But monitoring is defence, not design. A system in which all users access data simultaneously at the same tier would eliminate most of the incentive to abuse it, and would create a public data foundation good enough for independent analysis groups to check one another.

The resource gap between the top ten and the top eighty

At the organisational level, tennis carries a structural paradox: a player competes as an individual but operates as a small business. Coach, fitness specialist, doctor, physiotherapist, analyst, commercial agent — each player builds their own apparatus.

Compared with football, where a large club maintains an entire data department, tennis allocates resources very unevenly. Players inside the world's top ten typically have a dedicated analyst and their own camera setups to record opponents at lower-tier events. Players around No. 80 often share services, or review footage on a phone.

This data gap does not show up in the rankings, but it exists. It surfaces in match-specific preparation: whether you know which box an opponent favours at 30-30, whether you know an opponent switches to the cross-court backhand on the third point of a game.

Four zones the data does not cover

Any analysis above has to come with its limitations. This section is mandatory in everything I write.

Public tennis data does not cover four zones. The first is undisclosed injury: a player may be competing with a wrist problem that only surfaces after the tournament ends. The second is off-court state: a coaching split measurably affects results over the following fortnight but appears in no data table. The third is motivation: a player already assured of a major entry may approach a smaller event with a completely different level of investment. The fourth, and the hardest, is shot quality that produces no point: an outstanding defensive scramble that forces the opponent to hit one more ball and miss is recorded as an opponent error, not as a winner for the defender.

The fourth zone worries me most, because it cannot be fixed by collecting more data. It requires a new labelling system, and that system has to be built by humans.

The addiction to picking a champion

The tennis analysis industry has an addiction: naming the champion.

The content market proves it. An article headlined with a question about who will win a Grand Slam draws many multiples of the readership of a piece explaining why a player's first-serve percentage drops six points when serving to the outer box on break point.

Yet naming the champion is precisely the lowest-information question available. With 128 players in a single-elimination draw, any individual's probability is compressed into a narrow band. Even the strongest player rarely exceeds 25 percent before the first ball. Put differently, the correct answer to the most exciting question is usually: probably not the name you have in mind.

The questions that actually carry value go largely unasked. What percentage of second-serve points does this player lose when the return comes back deep? What is the down-the-line backhand win rate in the third set? After roughly how many points does movement quality begin to degrade?

There is another temptation I have to warn myself against. Once you have data, it becomes easy to believe that digitisation solves everything and that objectors are simply short on numbers. I used to think that. Now I think most objectors are not short on numbers at all — they are short on reasons to trust the model, because nobody has explained to them where it fails.

Tennis has no transfer window in the football sense, but it has a market equivalent in consequence: the market for coaches and support staff. Deals there are also where people pay to buy a row in a data table. A coach who has improved the serve percentage of three different players gets paid on that number, even though the number may not be the cause.

In the end, data does not lie; it is the people reading it who make excuses.

Signals for the next round

The signal I will be tracking in the coming stretch is not a champion's name.

I will be watching whether tournaments begin publishing ball specifications — compression, diameter after use, number of games before a change — as part of official match data. If that happens, cross-tournament comparisons regain their validity. I will be watching whether point-level data is opened at a public tier deep enough for independent research groups to verify one another's conclusions. And I will be watching whether anyone dares to publish a model accompanied by a full account of what that model does not know.

If you ask me who will win this tournament, the most honest answer I can give is still the one from that Brisbane night: not enough information to conclude. The only difference is that now I know exactly what information I am missing.