Can AI Be Used to Help Win Your Bets?

AI keyboard

Whether it be extensive research, expert tips or betting systems, there are numerous ways in which punters try and score a solid run of winning bets at the bookies. Some do manage this, but only a minority as your average bettor is unable to overcome the house edge the bookmaker enjoys. For those losing more money than they win, the idea of using AI to help turn their fortunes around may well offer appeal. After all, the usefulness of AI across sectors is hard to ignore.

To give you a few examples of this, when looking across high-quality studies of AI in the medical field, one meta-analysis found that AI correctly diagnosed diseases 87% of the time (compared with 86% for healthcare professionals). Over to ChatGPT and it managed to outperform most humans when taking the Uniform Bar Exam in 2023, scoring within the 90% percentile. Lastly, bringing it back to sport, if you consider chess a sport, AI-powered opponents are now so strong that grandmasters rarely face them because it is not a fair fight.

Evidently, AI has a lot of real-world applications and it can perform at a higher level than many experts can. Given this, it is not an unreasonable assumption to think AI can help punters with their betting. This does not appear to be the case though, even though the likes of ChatGPT are publicly available. The situation may change in future but as things stand AI is not something that can help bettors secure winning bets in the long-term.

No Help from Chat Bots

Chat GPT AI bot

If you have used ChatGPT before, you may have noticed that there are certain requests it will not comply with. One of these is asking for predictions for sporting events. When we asked it if Liverpool would beat Brentford this weekend, this is the response we got.

As an AI language model, I do not have the ability to predict the future or know the outcome of specific events such as a football match between Liverpool and Brentford. The result of any given match is influenced by a wide range of factors, including the form of the teams, injuries, tactics, and many other variables that are difficult to predict with certainty. The best way to find out the result of the match is to watch it or follow the news and updates related to the match.

Wording our prompt differently, such as what is the percentage chance that Liverpool win, or who is more likely to win, always yielded a similar response from the ChatGPT. It is not programmed to make any sort of predictions nor will it share a ‘personal’ opinion about its favoured team or which makes for the most appealing bet. Due to this, ChatGPT is of virtually no help to any gambler.

Part of the reason for this is that ChatGPT does not have access to real-time information. When making our request, ChatGPT informed us that its knowledge cut-off date was sometime in 2021. As such, it could not tell us useful pre-match news such as injuries, suspensions and so on even if it was programmed to do so. This is not a limitation specific to ChatGPT either so using other language chat bots will not enable you to produce summaries of the latest team news.

What About AI Forecasts?

Five Thirty Eight predictions

Although ChatGPT refuses to make predictions about sporting events, this does not mean AI is incapable of doing just that. Various models have been created to try and calculate the percentage chance of sporting outcomes, whether it is winning a particular match or winning the league. One popular version of this comes courtesy of FiveThirtyEight. For leagues across the planet, they provide probabilities of various outcomes (such as league winner, a top four finish, or relegation) and update these regularly following real-life results.

A quick look at them suggests their model does a decent job. If you look at their Premier League predictions from before the 2021/22 season, you can see they called three of the top four correct and two of the relegated sides. This is a solid record although little more than you would expect from any decent tipster.

FiveThirtyEight Predictions (2021-22) Qualify for UCL % Relegation % Actual Finish
Man City 84% 1st
Liverpool 69% 2nd
Chelsea 64% 3rd
Man Utd 50% 6th
Crystal Palace 36% 12th
Brentford 36% 13th
Norwich 40% 20th
Watford 42% 19th

Let’s see how you would have fared had you stuck £10 on all their predictions when placing your bets. We will pick Crystal Palace as our third relegation selection as FiveThirtyEight predicted their goal difference would be one worse than Brentford’s.

Team Top Four Finish Odds Net Win/Loss Relegation Odds Net Win/Loss
Man City 1/20 £0.50
Chelsea 2/9 £2.22
Liverpool 2/9 £2.22
Man Utd 2/7 -£10
Crystal Palace 7/5 -£10
Watford 1/1 £10
Norwich 10/11 £9.09

Overall, by picking the selections FiveThirtyEight had the most confidence in, you would have been up £4.03. Not a big return from £70 wagered but a net win nonetheless. Let us take a look at the season before (2020/21) too to see if they could have guided punters to a net win then as well.

Team Top Four Finish Odds Net Win/Loss Relegation Odds Net Win/Loss
Man City 1/20 £0.50
Liverpool 1/12 £0.83
Man Utd 8/15 £5.33
Chelsea 1/2 £5
Fulham 1/1 £10
Leeds United 4/1 -£10
West Brom 11/10 £11

An even better display here with £22.66 net win going from their predictions. To see if this is something of a trend, we will take a look at one previous season (2019/20).

Team Top Four Finish Odds Net Win/Loss Relegation Odds Net Win/Loss
Man City 1/33 £0.30
Liverpool 1/10 £1
Chelsea 1/1 £10
Tottenham 4/6 -£10
Brighton 2/1 -£10
Norwich 5/4 £12.50
Sheff Utd 5/6 -£10

Here you would have been down £6.20 from your bets, so it would be wrong to think models like FiveThirtyEight are consistent money earners. This is not to say they are not of value though, but you have to interpret their data in the right way.

Finding Value in Computer-Generated Predictions

AI mobile phone

When looking at FiveThirtyEight and their predictions, we compiled the list of the teams they thought were most likely to secure a top-four finish and those most likely to be relegated. To simply back these teams, however, would be to miss the potential value of their statistical model. What you should instead be doing is attempting to find where the value lies. Let us take a look at how this works in practice.

Team (2021-22 PL season) FiveThirtyEight predicted chance of relegation Odds Odds converted into %
Fulham 32% 1/1 50%
West Brom 36% 11/10 47.6%
Aston Villa 25% 2/1 33.3%
Crystal Palace 23% 9/4 30.8%
Newcastle 31% 5/2 28.6%
Burnley 20% 7/2 22.2%
West Ham 14% 4/1 20%
Sheff Utd 21% 4/1 20%
Brighton 22% 4/1 20%
Leeds 33% 4/1 20%

Where FiveThirtyEight’s predicted chance of relegation is higher than the implied probability, this makes for a ‘good value’ bet. If you were huge fans of their model, this means you would end up backing Leeds, Brighton, Sheff United and Newcastle for the drop. Had you done this, you would have won £10, assuming you bet £10 on all of them.

This is the best approach to treat AI-generated predictions, not simply to take the names at the top of the list. The difference in the estimated relegation chances between Fulham (32%) and Newcastle (31%) was negligible but the Magpies were available at double the price. This makes it a far more sensible bet, should you want to put your faith in the data.

You can also apply this same approach to individual matches as some models do compute this. Going back to our very first example of Liverpool versus Brentford, here are the win probabilities provided by Google (unlike FiveThirtyEight, there is no information on how the probabilities are calculated).

Premier League Win Probability

It would be short-sighted to look at this and say, Liverpool have the best chance of winning, so I will back them. Instead, you need to look at where the value lies.

Outcome Predicted chance Best Odds Odds converted into %
Liverpool win 66% 1/2 66.7%
Draw 19% 4/1 19%
Brentford win 15% 11/2 15.4%

In this case, all outcomes have approximately the same value, although a draw is marginally the best as the other two have predicted chances slightly lower than the implied probability. Across most games we looked at, the Google probabilities largely reflected the odds available, usually no more than a couple of percentage points either side. Here is another example taken from another Premier League match.

Outcome Predicted chance Best Odds Odds converted into %
Brighton win 47% 21/20 48.8%
Draw 24% 3/1 25%
Man Utd win 29% 5/2 28.6%

If you were to assume the model was incredibly accurate at calculating percentages then a Man United win makes for the most appealing bet as it gives you a 0.4% edge. This is not much use to the punter though because not only will the model be reliable enough in its calculations, but even if it was, this is a very tiny margin to be working with. If you were to take the probabilities as perfectly accurate, and repeated this game 100 times, here is your net win if you were to bet £5 each go.

Number of Winning Bets (29) = £362.50 Net Win – £355 Lost Bets = £7.50 Total Net Win

When watching a match on TV, you may sometimes see a ‘live’ win probability, which is a real-time calculation made during the match. Opta is one provider of this and their calculations take into account a number of factors, such as historic performance, recent performance, home advantage, in-game performance, current score, red cards and time remaining.

Win probability

This update was provided seven minutes into a clash involving Burnley and Manchester City. By converting these into implied probabilities, you could have checked to see if any bookmakers were offering good value bets on any of the outcomes. We would, however, be surprised if any were offering more than 2% better than the probabilities on screen. So, whether you want to follow a reputable pre-game or live model, using it to find value in the market is a difficult task indeed.

The Bottom Line

Rigorous long-term analysis of football predicting models has not been carried out to prove their accuracy. Even if there was, you have the problem that models are constantly tweaked to improve their accuracy, making long-term analysis unreliable. It is fair to assume that the most highly-reputable models are able to provide decent probabilities but they are not perfect. This is partly because they are purely reliant on data and cannot factor in less tangible elements of the game.

If you did want to put your faith in a particular AI model yourself, and only back selections that provided good value, it is possible, certainly for outright bets. For individual game bets, the probabilities of statistical models follow the odds so closely that margins are razor-thin. Even when there is a slight advantage, this is dependent on you getting the market-leading price in order to gain your tiny advantage. This would inevitably mean registering and depositing at numerous bookmakers, making it something of a faff.

It might be worth the effort if AI was flawless at producing probabilities for football matches but it is not. Being accurate, but not perfect, is not enough when working at such small margins. By all means play around with some models but from our research it is hard to see how that can be relied upon for guaranteed long-term net wins.