Tipsters vs AI, Algorithms and Crowds: Who Made Money?

We sorted 155 of the 172 football prediction services in the Tips.GG Accuracy Index by who makes the picks: people, AI, algorithms or the crowd. Two in five of the human-run ones finished in profit, and the services built on an algorithm managed fewer than one in ten.

Thirty-five of the algorithm-run services published more than 1,000 predictions each between January and September 2026. None of them finished ahead.

The top of the table belongs mostly to people. Four of the five services that returned more than 10% are run by human tipsters.

In full, the share in profit runs: people 41.3%, AI 16.0%, the crowd 15.4% and algorithms 7.1%.

This piece uses three measures: the share of services in profit (how many in a type finished ahead), median ROI (how the typical service in a type did) and ROI per stake (what backing every pick of a type would have returned).

That is the headline. The rest of this piece is about why it is less simple than it looks.

Who Makes the Picks

In our first audit of football prediction services, we ranked them by results. Here we sort the same index by who makes the picks.

There are four types.

Human: a named expert, an editorial team or a single tipster, as on bets.co.za and wettfreunde.net.

AI: the service says artificial intelligence, machine learning or neural networks produce its picks, as StaTips on X (Twitter) and betideas.com do.

Algorithm: the service describes a statistical model or algorithm and does not claim AI, as betimate.com and TheyScored (Twitter) do.

Crowd: the picks are aggregated from users, through votes, as on sofascore.com, or community tipster leagues such as olbg.com and bettingexpert.com.

The rule is simple. A service is whatever it says about itself where its predictions are published: its website, About page, app store listing or profile bio. We recorded a quote and a link for every row. That placed 160 of the 172 services.

Twelve services say nothing about who or what makes the picks we track. For these, the type is an editorial decision: three were ruled Human and nine Algorithm, marked with an asterisk in the table below. None of the twelve finished in profit, and they don’t change the result. Leave them out and the order stays the same: human services rise to 43.1% in profit and algorithm services to 9.1%.

Seventeen services are not part of the comparison. They describe a mix of methods and appear as Mixed in the table below. Forcing them into one type would mean inventing a method they don’t claim.

The Gap Opens Among Small Services

The difference between people and algorithms is real. It is also a difference in output.

The median algorithm service published about 5,500 predictions over the nine months. The median human-run service published 423. Thirty-five of the 42 algorithm services passed 1,000 predictions; 20 of the 75 human ones did.

Profitability falls away as volume grows. Split each group at 1,000 predictions and the gap closes at the top but stays open below it.

Above 1,000 predictions, almost nobody makes money, whoever or whatever is picking: 5 of the 73 compared services, and not one of them an algorithm. Three of the five are human, and none of the three finished more than 0.85% ahead. Below 1,000, people lead. Of the human services, 50.9% finished in profit, against 3 of 7 algorithm services, 3 of 15 AI services and 1 of 5 crowd feeds. These groups are small, so read that as what this dataset shows, not as a law.

What the split does not show is that people predict football better than models. A tipster posting 200 picks has far more room for a lucky run than a model posting 9,000. Most algorithm services choose to publish at a scale where luck averages out, and the bookmaker’s margin does the rest.

Tips.GG’s own predictions come from an editorial team, so our row counts as Human. It reads 5,836 predictions, 53.9% accuracy and −4.05% ROI.

That places us among the 20 high-volume human services. Three of them finished ahead, none by more than 0.85%. Publishing at scale is where human tipsters look most like everyone else.

AI Is a Wide Bet, Not a Better One

Twenty-five services describe their picks as AI-driven. Four of them finished in profit, or 16.0%, below the index as a whole and barely ahead of the crowd.

The group’s results run in both directions. StaTips (Twitter) says its team develops AI predictions. It returned +10.11% on 582 predictions. At the other end, Scorebet24 (Twitter), which describes its tips as AI-driven, lost 17.66% on 473.

The label itself is a poor guide. Seven services in the table have “AI” in their name. Only one finished ahead: Football Predictions AI (Twitter) at +3.18%. footballpredictions.ai, which has AI in its name and no methodology we could find, lost 15.10%, the second-worst result in the group.

AI services are not more accurate than human-run ones either. The median AI service called 51.0% of its picks correctly, against 54.5% for human-run services and 50.0% for algorithm services.

“People ask us whether AI has finally beaten the tipsters. The honest answer is that the label on a prediction tells you far less than the record behind it. Over a few hundred picks almost anyone can look brilliant — a person, a model or a crowd. Over thousands, almost everyone ends up paying the bookmaker’s margin. The question worth asking a service isn’t who or what makes its picks. It’s how long it has been right at the prices it was given.”

Anton Malyutin, CEO of Tips.GG

The Crowd Loses Slowly

Crowd feeds behave differently from both people and models. They rarely go far in either direction.

Thirteen services aggregate what their users pick. Their results run from −5.02% for Football Predictions Community Voting (Twitter) to +5.20% for betblog.com, and no other type stays in such a narrow range. Only two finished ahead: betblog.com and BetExplorer (Twitter) at +2.22%.

That narrow range shows when every pick is counted. Pooled together, crowd picks returned −3.27% per stake, the smallest loss of the four types. It is still a loss: backing the whole crowd would have cost money.

The biggest crowds all lost. sofascore.com, betclan.com and olbg.com together carry more than 35,000 predictions, and none finished ahead. A vote across thousands of users is an average, and averages rarely stray far from the price in either direction.

About the Data and Its Limits

The Tips.GG tracker logs football match-result (1X2) predictions published by websites, tipster operations and individual channels. This study covers 1 January to 30 September 2026. Every prediction is settled at a flat $10 stake at the price Tips.GG records for that match, bookmaker margin included (8.33% on average), so a service has to beat the margin to finish ahead.

Accuracy is the share of predictions that won. ROI is profit as a percentage of total stakes, and a service counts as profitable when its ROI is above zero. To qualify, a website needs 50 graded predictions and an account listed under a platform needs 200. Names run by one publisher are merged into one row. That leaves 172 services, 43 of which made money. The full methodology is published at Tips.GG Accuracy Index: Methodology.

The type is a claim, not a method we verified. We sorted services by what they say about themselves. We did not audit what runs underneath. theanalyst.com describes an “AI-powered Opta supercomputer”, whose predictions come from match simulations built on Opta ratings; it counts as AI because that is how it presents itself. The same service can read differently in different places. forebet.com’s app listing cites artificial intelligence, but its website, where its predictions are published, describes “mathematical algorithms and statistical models”, so it counts as Algorithm. We typed every service by the pages its predictions come from.

Some rows contain more than one kind of pick. In six compared services outside the Crowd type, part of the feed is signed as community voting: xgscore.io 41.6%, livetipsportal.com 41.0%, betmines.com 37.2%, sport.ua 13.5%, zagranie.com 7.9% and goal.pl 7.3%. Three more carry under 1%. Each service is counted as a whole; we did not separate the voting share from the rest of the feed. One channel appears as two rows. Matt SB (YouTube) runs the channel, and Sophie SB (YouTube) gives her own picks in a video series on it; our tracker logs her picks separately, under her name.

Some groups are small. Crowd has 13 services. A single result moving across zero shifts its rate by nearly eight points.

Every pick is settled the same way. The flat stake is a basis for comparison, not a claim about how anyone bet, and the single price Tips.GG records for a match is not necessarily one any reader could have taken.

So Who Made Money?

No type did as a whole. Pool every pick, or take the typical service, and all four types finished behind the bookmaker. What made money were individual services: 43 of the 172 in the index, most of them small, human-run operations.

The three measures show where each type stands:

• Share in profit: people lead, with 41.3% of human-run services ahead, against 16.0% for AI, 15.4% for the crowd and 7.1% for algorithms.
• Median ROI: people lead again, with the typical human-run service at −1.22%, ahead of the crowd at −2.67%, AI at −3.76% and algorithms at −4.89%.
• ROI per stake: the crowd edges ahead at −3.27%, with people at −3.74%, AI at −4.35% and algorithms at −4.91%.

The measures agree at the bottom: algorithm services come last on all three. At the top they split between people and the crowd, and on ROI per stake the two sit less than half a point apart.

So the answer to the headline is not a type. A service that beat the margin is most likely to be a human tipster with a few hundred picks. But a type is not a track record: before following any service, check how many picks it has published and how it did at the prices it was given.

Reusing the Data

The full table is below, with all 172 services, each service’s type, predictions, accuracy and ROI. It is free to cite and reuse with attribution to Tips.GG and a link back to this research. The type of every row, with the quote and link it rests on, is available on request, as are methodology questions and the data in other formats: [email protected].

All 172 Services by Type

All 172 services, ranked by ROI. Football match-result predictions, 1 January to 30 September 2026, settled at a flat $10 stake. Accuracy is the share of predictions that won. ROI is total profit as a percentage of total stakes. Type reflects each service’s own public description; 12 rows were typed by editorial decision, as explained above. Rows marked Mixed are not part of the type comparison. * Type decided by Tips.GG editors; the service does not describe how its picks are made.

You can view all Tips.GG research here.