AI Sports Betting Predictions: What Holds Up
December 31, 2025
What changed on this page. The earlier version printed a table giving artificial-intelligence models an average return of +8% to +18% against −2% to +3% for human experts. Nobody ever published those figures; they were invented for the page, and the table is gone. What replaces it is the arithmetic any prediction has to clear before it is worth paying for.
Machine learning really is used in betting — mostly by bookmakers and trading firms, to price markets faster than a person can. What is sold to the public under the same word is usually a subscription to a list of selections. The test for both is identical and easy to state: does the price you took beat the price the market settled on?
What a model can and cannot do
It can estimate a probability from structured inputs — recent form, lineups, rest days, travel, venue, pace — react to team news in seconds, and stay consistent. It never falls in love with a team, never chases a loss, and never gets tired at the end of a slate. Those are real advantages over a human reading the same numbers.
What it cannot do is see what is not in the data, or beat a price that already contains everyone else’s estimate. A bookmaker’s odds are not an opinion; they are a market aggregate with a fee added. Being right more often than not is worthless if you are right at prices that were already better than your estimate.
There is also a limit on what any winning system can do in a retail account. Bookmakers restrict or close accounts that beat them consistently, so a public track record built on ordinary accounts tends to end exactly where it starts working.
The margin is the bar
Implied probability is 1 divided by the decimal odds. Add both sides of a two-way market and whatever exceeds 100% is the bookmaker’s margin — the amount the price is shaded by to pay for the book.
| Two-way price | Implied probabilities | Sum | Margin | Break-even on the first side |
|---|---|---|---|---|
| 1.90 / 1.90 | 52.6% and 52.6% | 105.3% | 5.3% | 52.6% |
| 1.95 / 1.95 | 51.3% and 51.3% | 102.6% | 2.6% | 51.3% |
| 2.10 / 1.72 | 47.6% and 58.1% | 105.8% | 5.8% | 47.6% |
| 1.50 / 2.50 | 66.7% and 40.0% | 106.7% | 6.7% | 66.7% |
So a predictor that wins 52% of its bets at 1.90 is not profitable: it loses about 1.2% of everything it stakes. The margin also compounds on accumulators — four legs priced with 5.3% each carry roughly 23% combined, which makes the four-fold the most expensive way to be right.
Why an accuracy percentage means nothing
Accuracy without prices is not a result. A service that wins 75% of selections priced at 1.25 is losing money: 100 stakes of one unit return 75 × 1.25 = 93.75 units, a loss of 6.25% of turnover, because break-even at 1.25 is 80%. The same 75% at odds of 2.00 would be extraordinary. The number is identical; only the price tells you which case you are looking at.
The first question for any published record is therefore not how often it was right, but at what odds, staked how much, and recorded when. A screenshot of winners with no prices and no timestamps measures nothing at all.
How to test a predictor before paying for it
- Out-of-sample only. Results on the data a model was trained on are a description of the past, not a forecast.
- Closing-line value. Record your price and the price at kick-off. Consistently beating the closing line is the fastest available evidence that an estimate is ahead of the market.
- Enough bets. With flat one-unit stakes at even money the swing after 100 bets is around ten units, while a 2% edge is two. A hundred-bet record proves nothing in either direction, and a three-week trial is a marketing device rather than a test.
| The claim | What evidence would be needed |
|---|---|
| 87% accuracy | The odds of each selection and the return on turnover; accuracy at unknown prices cannot be judged |
| +12% ROI last season | A bet-by-bet log with the price taken and a timestamp before each event |
| Beats the bookmakers | Closing-line value over several hundred bets in the same market |
| Trained on millions of data points | Out-of-sample results on data the model never saw; the size of a training set proves nothing |
| Accumulator of the day | The margin on each leg — four legs at 5.3% carry about 23% together |
| Risk-free or guaranteed | Nothing can support it; the subscription is paid whether the selections win or lose |
Where a model genuinely helps
As a filter rather than an oracle. It can screen a slate of two hundred markets down to the handful where your estimate and the price disagree enough to be worth a stake, and it enforces the habit of writing a number down before the market moves — the same discipline the esports betting guide describes without any machine learning in it at all.
In-play the advantage is timing rather than insight, and there the obstacle is structural: your stream runs behind the server while the bookmaker’s data does not, so a suggestion that arrives three seconds after the round has ended is already priced in. The live esports playbook covers what that delay does to any in-play system. This site, for its part, sells no predictions and has no platform behind the name — betiyo.com is a parked domain offered for sale — so there are no odds here that these tests would be asked to flatter.
A model changes your estimate, never the margin you are paying. Set a deposit limit and a stake size before the first bet, keep stakes flat at around one per cent of the bankroll, and treat a losing week as information rather than as something to win back. Never bet with borrowed money or on credit, and if betting has stopped being entertainment, every market this site is read in has a free national support service.
Frequently Asked Questions
Can artificial intelligence beat the bookmakers?
Occasionally, and mostly where the market is thin: syndicates with their own data and fast execution do it. The test is whether the price taken beats the closing price. A public subscription that genuinely managed it would move the very prices it was selling, which is why the claim is far more common than the evidence.
What does a 90% accuracy claim tell me?
Nothing on its own, because it omits the odds. Winning 75% of selections at 1.25 loses 6.25% of turnover, since break-even at that price is 80%. The same rate at 2.00 would be remarkable. Ask for the prices taken and the return on turnover instead.
How many bets do I need before judging a predictor?
Hundreds, with every price recorded. After 100 flat bets at even money the ordinary swing is about ten units, larger than any realistic edge, so a short record is noise. Closing-line value gives a faster read than profit does.
Treat any predictor the way you would treat a price: check what it is built on, compare it with the closing line, and assume that anything sold as certain is being sold rather than measured.