How Accurate Are AI Sports Predictions? Read the Graded Record

The honest answer to “how accurate are AI sports predictions” is that it depends what the number means, and most sites do not say. A hit rate of 70% can describe a method that loses money and a hit rate of 45% one that makes it. This guide explains what accuracy measures, why it diverges from the return at short prices, what a market-anchored prediction should be expected to score, and where the live figures for Sportslyx are published. None of those figures are quoted here: a number printed in an article is out of date by the time it is read, and a record is something you check, not something you are told.

What “accuracy” actually measures

On Sportslyx, and on any site that grades honestly, accuracy is the hit rate: wins divided by wins plus losses, with pushes and voids left out because a push is a decision that did not resolve. It answers one question — how often was the side right — and it is silent about the price. That silence is the whole problem. Two records can both show 65% and describe completely different methods: one backing favorites at −200, which needs 66.7% just to break even, and one backing near-even prices, which is comfortably profitable at 65%.

There is a second thing the phrase hides when the word “AI” is attached. On Sportslyx the language model does not produce the probability. The percentage on every game is the market moneyline with the bookmaker’s margin removed — the same figure on the board, the game page and inside the breakdown — and the model writes the explanation around it. What the model contributes is prose and a data-quality tier; what gets graded is the market-derived read. So “how accurate is the AI” really asks two checkable questions: how well calibrated is the market the number came from, and how strictly are the results graded.

Hit rate versus ROI at short prices

The return at a fixed stake is where price re-enters. For a favorite at American odds of −X, the break-even hit rate is X ÷ (X + 100): a −150 favorite must win 60% of the time, a −250 favorite 71.4%, a −400 favorite 80%. The table shows what happens to the return on investment when the actual hit rate lands five points either side of break-even, at a flat $100 per read.

Two things follow. The shorter the price, the higher the hit rate has to be before the method earns anything, which is why an advertised “75% accuracy” on −300 favorites describes a method at break-even, not a good one. And the same five-point miss costs about the same money at every price, which is why a record built on short favorites can look excellent by hit rate and unremarkable by return. Always read the hit rate next to the average price or, better, next to the ROI at a fixed stake.

What a market-anchored prediction should score

If the probability is the de-vigged market line and the closing line is well calibrated — and for liquid markets it is, to within a point or two — then the expected hit rate of a set of reads is simply the average of their probabilities. Back sides priced at 62% on average and, over enough volume, you should hit about 62%. The expected return of doing so is not zero, either: it is roughly the bookmaker’s margin, negative, because the price you are paid is the market’s price with the margin still in it.

This is the caveat Sportslyx prints on its own prediction pages. The company board backs the market favorite by design, so its hit rate is naturally high and its return naturally modest. A high percentage at short prices is the expected shape of that record, not evidence of an edge, and the ROI can be negative in a given month without anything being broken. The record is published anyway — every month, losses included — because the alternative is a record you cannot check.

Where a published accuracy figure can mislead

• Streaks instead of samples — a run of short favorites lands by chance every few weeks on any large enough board
• Windows chosen after the fact — “since the method changed” or “last 30 days” selected because it looks good
• Pushes counted as wins, or pushes and voids left out of the settled count so the volume looks larger than it is
• No price on the record, so the hit rate cannot be turned into a return
• Grading rules that vary by game — a draw is a push one week and a loss the next
• Variance mistaken for signal — a 60% favorite loses four times in ten by construction, and a bad week says nothing about the method
• Samples too small to mean anything — at fifty settled reads the 95% interval around a hit rate is about fourteen points wide

The grading rules behind the Sportslyx figures

The published numbers are produced by a settlement job that reads official final scores and box scores, under rules fixed in advance and applied the same way to every read. The rules that move the numbers most:

• Game reads settle on the final score, including overtime and shootouts where the market settles that way
• A soccer draw is graded as a push — stated openly, because it flatters the record compared with a draw-no-bet price
• MLB player props — the only props the board reads — settle from the official box score; lines are half-numbered so a prop cannot push, and a player who did not play is void, not a loss
• Postponed games, cancelled fights and no-contests are void and removed
• Golf previews are published for information only and never enter the graded record
• Reads close fifteen minutes after the scheduled start; nothing is added once a game is under way
• Hit rate excludes pushes; ROI is measured at a flat stake per read; the period is one calendar month in UTC, the same everywhere in the product

Where to read the live numbers

Three public pages carry the record, all computed from the same rows a subscriber sees. The Pick Tracker shows the company board by calendar month and by sport, with settled count, wins, losses, pushes, hit rate and ROI side by side. The Results ledger shows every settled day in full — side, price, probability, score and outcome — with today and the following days visible only as counts, because those reads are the product. Each sport’s prediction page, such as the MLB board, repeats that sport’s monthly record beneath the day’s games.

Past results do not predict future results. The record exists so the method can be judged on volume, not so any single month can be sold as proof — and a good month is partly variance whichever way it went.

How to read them

1. Read the settled count before anything else; below a few hundred reads the percentages are mostly noise
2. Read the hit rate next to the ROI, never alone — the pair tells you whether the prices justified the accuracy
3. Look at each sport separately: a soccer row carries the draw-push rule, an MLB row mixes props with game reads
4. Compare months, not weeks, and expect negative ones from a market-anchored method
5. If you log your own reads, judge them the same way — the tracking guide walks through the arithmetic