Since January 2024, Benfica, Porto and Sporting have met each other 16 times in the league, in Clássicos and Lisbon derbies, and seven of those games ended level. Porto put five past Benfica in March 2024. The following season Benfica went to the Dragão and won 4-1. In May 2025 the derby at the Luz finished 1-1 with the title on the line. If you were hoping for a pattern, the big games aren't going to give you one.
Against everyone else, the same three clubs are a different story. Over the same period they played 260 league matches against the rest of the Primeira Liga and won 207 of them. They lost 16. That's a win rate of 79.6 per cent and roughly one defeat every 16 games. It's also why, in more than 90 years of league football in Portugal, only two other clubs have ever been champions: Belenenses in 1946 and Boavista in 2001.
I wanted to know what that gap looks like to a machine, and it turns out there's a neat way to find out. NerdyTech Hub, a company based in Câmpina, Romania, runs two prediction models side by side: NerdyTips, which has published AI football predictions since 2021, and BananaPoint, which does the same job for tennis. Both put their full archives online, misses included, and both attach a confidence score to every call. That makes it possible to set two sports next to each other and ask a simple question. Which one is easier to read, and why?
Easier to read than the Premier League
The NerdyTips archive holds 880 Primeira Liga matches, from January 2024 to early November 2026. Counting only the games where the model named a winner outright, it was right 57.2 per cent of the time. That may not sound like much until you set it against the rest of Europe over the same stretch. In the Bundesliga the figure was 55.5 per cent, in Serie A 55.4, in the Eredivisie 54.9, in La Liga 54.7, in Ligue 1 53.3 and in the Premier League only 52.0. Of the seven leagues I compared, Portugal's was the easiest to call.
The gap grows when the model is sure of itself. Calls it rated at 80 out of 100 or higher were right 75.8 per cent of the time in Portugal, compared with 63.8 per cent in England.
But that predictability runs exactly three clubs deep. When Benfica, Porto or Sporting played anyone else, the model's outright calls were right 77.9 per cent of the time. Take all three out and look only at games between the other 15 clubs, and the figure falls to 47.0 per cent, with almost a third of those matches (30.3 per cent) ending in a draw. Portuguese football isn't really one league. It's a procession at the top and a scrum everywhere else.
Going down the pyramid only thickens the fog. In the Segunda Liga the model's outright calls landed 43.0 per cent of the time, in Liga 3 it was 44.6 and in the Campeonato de Portugal 49.2, with draws making up between 27 and 30 per cent of results across those divisions. In Portuguese football, the higher you climb, the clearer the picture gets.
Tennis, turned upside down
Now try the same exercise with tennis. BananaPoint puts out AI tennis predictions for nearly every professional match on the calendar, from the biggest ATP and WTA events down to the smallest ITF tournaments. Its match-by-match file, when I downloaded it, covered 7,515 completed matches between 29 September and 12 November 2026. A tennis match always produces a winner, which makes life a little easier for a forecaster, and overall the model named the right one 63.0 per cent of the time. That's higher than in any of the football leagues above.
The real surprise is the shape. The easiest matches to call were on the lowest rung of the women's game, in ITF events, at 67.2 per cent. The WTA tour came next at 64.5, then men's ITF events at 63.4. Once you reach the Challenger and ATP level, things get hard: 57.6 per cent on the ATP tour and 56.5 on the Challenger circuit. In Portuguese football, the elite is where certainty lives. In tennis, the men's elite is where it goes missing.
That ATP figure sits almost exactly on top of the Primeira Liga's 57.2 per cent, which is a lovely coincidence, but the reasons behind the two numbers point in opposite directions. Portuguese football is predictable because of inequality: three clubs with far deeper squads and far bigger budgets than anybody else. Men's tennis at the top is unpredictable because of equality. Of the ATP matches in the sample, 43.7 per cent had at least one tiebreak, against 16.8 per cent in women's ITF events, and 39 per cent went to a deciding set. When two big servers hold for an hour, a set comes down to a few points, and nobody can tell you in advance which way those will go. Ben Shelton needed three tiebreaks to get past Aleksandar Kovačević in Basel in October, and he was one of four ATP players in six weeks to win a match that way. The tiebreak, in other words, is tennis's Clássico: the moment when everything you know stops counting for quite so much.
Knowing when you don't know
The most interesting thing in both archives, though, isn't the hit rate. It's the confidence score.
A forecast is only useful if its confidence means something. If the weather service says 70 per cent chance of rain, it ought to rain on roughly seven of every ten days it says so. BananaPoint's tennis model is remarkably good at this. When it rated its call for the winner at 90 or above, it was right 87.1 per cent of the time, across 443 matches. In the 80s it was right 82.4 per cent of the time. Below 20, across 1,852 matches, it was right 49.2 per cent of the time, which is a coin toss, and a coin toss is precisely what the low rating had been telling you. Around a quarter of all matches fell into that zone, and the model said so plainly instead of dressing up a guess as insight.
Football makes the same job harder, because a match has three possible results rather than two. Even so, the pattern holds in Portugal. NerdyTips' outright calls rated 80 or higher landed 75.8 per cent of the time, while those rated below 40 came in at about 43 per cent.
Put the two side by side and tennis turns out to be the more decisive sport once the evidence is clear. BananaPoint's calls rated 80 or more were right 85.4 per cent of the time, almost ten points better than the best football league in the comparison. More than anything, it's the draw that holds football back.
Portugal on the court
The two sports cross paths in Portugal more often than you might think. In those six weeks alone, BananaPoint rated 272 completed matches at nine tournaments in the country: Challengers in Braga, Maia, Porto and Lisbon, a WTA event in Lisbon and ITF tournaments in Quinta do Lago, Loulé, Lagos and Vale do Lobo. When the model was confident there, rating a call 80 or higher, it was right 18 times out of 21.
It also kept a close eye on Henrique Rocha through the autumn. The model backed the young Portuguese to beat August Holmgren in Porto and Thiago Monteiro in Lisbon, and he won both.
Two archives worth reading
BananaPoint is the younger of the two projects. Its model, BananaCore, made its first calls in November 2025; the website opened to the public in September 2026. Since that first day it has graded 45,223 predictions, 67.1 per cent of them correct, counting the call it was most confident about in each match, and every month from January to October 2026 finished between 66.8 and 69.7 per cent. Each prediction is time-stamped before the first serve and can't be changed afterwards. There are pages for 1,214 tournaments, from the four Grand Slams to 789 ITF events, and player pages that bring together serve and return numbers, results on each surface and recent form. It's a deep and carefully built resource, and the openness is what sets it apart from almost everything else in the field.
NerdyTips has been at it longer. Its public record goes back to May 2021 and covers close to 295,000 football matches from over 700 competitions.
What the two archives show together is that predictability isn't a fixed property of a sport. It depends on where you look. Portuguese football is easy to read when a giant meets everyone else and almost impossible to read when the giants meet each other. Men's tennis is hardest to read from the Challenger circuit upwards, and gets clearer the further down the rankings you go.
The best thing about both models is that they seem to know which world they're in. Ask BananaPoint about two big servers on a fast indoor court and it will tell you, in numbers, how unsure it is. Ask NerdyTips about a Clássico and you'll often get the same honest answer. Given that seven of the last 16 ended level, that may be the smartest prediction of all.
