How AI Sports Predictions Work

An AI sports prediction sounds like a black box: data goes in, a confident number and a paragraph of analysis come out. In practice the pipeline behind a good one is unglamorous and inspectable, and understanding it is the difference between using the output well and being misled by it. This guide walks through the five stages Sportslyx uses to produce a prediction — from the market line to the graded result — and explains what the AI actually does at each one.

The single most important idea comes first: on Sportslyx the win probability you see is not the language model’s opinion. It is derived from the market. The model’s job is to explain the matchup, not to invent the number.

Step one: the market line becomes a probability

Every prediction starts with the public moneyline for the game, which is the price on each side to win outright. A moneyline is a statement of probability with a margin built in. American odds of -150 mean you would have to risk $150 to win $100, which implies a 60.0% chance; the other side at +130 implies 43.5%. Add the two and you get 103.5%, not 100%. The extra 3.5 points is the bookmaker’s margin, usually called the vig or overround.

Removing that margin is called de-vigging. The simplest method, and the one Sportslyx uses for two-way markets, divides each implied probability by the sum of both so the pair adds to exactly 100%. The -150 side becomes 58.0%, the +130 side 42.0%. That 58.0% is the number shown on the game page as the win probability. It is a statistical estimate of what the market believes once its margin is stripped out — a market-derived probability, not a house opinion.

A worked example

Why anchor to the market at all? Because the closing line of a liquid market is the best public estimate of a game that exists. It aggregates injury news, sharp money and everything else faster than any single model. A prediction that starts from the market and explains around it is far harder to fool than one that starts from scratch and hopes to outdo it.

Step two: the data package

The probability is the anchor; the data package is what the analysis is written from. For each game on the slate Sportslyx builds a per-game package from public feeds — the ESPN scoreboard is the primary source for schedules, scores, base lines and box scores, with optional enrichment from other providers where available. The package typically contains:

• Current market lines — spread, moneyline and total — and how they have moved
• Recent form for both sides and the head-to-head record where one exists
• Availability: injury reports, probable starting pitchers, confirmed goaltenders
• Rest and schedule context such as back-to-backs or short weeks
• Venue
• Team identity — name, abbreviation and crest — from a profile refreshed daily; every statistic above lives in the per-game package itself, not in the profile

Not every game gets a full package. Feeds are delayed, some competitions publish less, and an early-week line may exist before the injury report does. Sportslyx scores each package with a data-quality tier — rich, partial or weak — and passes that tier to the model. A weak tier tells the model to be more cautious, and it is shown to the reader so that a thin breakdown is recognizable as one.

Step three: what the language model actually does

With the probability and the package in hand, a large language model writes the breakdown. It is called through a provider-agnostic gateway, so the specific model can change without the pipeline changing. The prompt contains the fair probability, the lines, the data package and the quality tier, plus rules about what a breakdown must cover: the strengths and weaknesses of each side, what the market appears to be pricing and what it may be underweighting — all written under the data-quality tier attached to the package, which is shown with the read.

The model is good at synthesis. It can turn a dozen inputs into a readable argument, notice that a favorite is playing its third road game in four nights, or that a starting pitcher’s recent strikeout rate does not match his season line. It is not good at arithmetic or at knowing what it does not know, which is why the number is fixed before it starts writing and why the data tier is fed in rather than left for the model to guess.

Step four: pre-generation and caching

Breakdowns are not written when you click. AI breakdowns are pre-generated for the upcoming slate and cached; the win probability on the card is derived from the current market line. Scheduled jobs run several times a day and write breakdowns for games that do not yet have one, so the page loads with the analysis ready; each breakdown is cached against the game and the version of the prompt that produced it. The breakdown text is not rewritten when the line moves. What follows the market is the number: a two-point swing on a spread or a confirmed goalie change moves the win probability on the card, because that figure is derived from the current line every time the page is served.

Two consequences follow. First, a breakdown you read in the evening may have been written that morning, before a late scratch; the probability on the card is current, the prose describes the matchup as the package stood when the job ran. Second, a breakdown written from a weak package stays flagged as such — the tier is shown with it rather than hidden.

Step five: grading and the public record

A prediction that is never scored is an opinion. Every read Sportslyx publishes is graded after the final result by a settlement job that reads official scores and box scores, and the results are aggregated into a public record by calendar month, in UTC, per sport. The rules are fixed in advance:

• Game reads are graded on the final score, including overtime and shootouts where the market settles that way
• A soccer draw is graded as a push — neither a win nor a loss — and this is stated openly because it flatters the record compared with a draw-no-bet price
• Player props — covered for MLB only — are graded from the box score; a player who did not play is void, not a loss
• Postponed games and cancelled fights are void and removed
• Hit rate is wins divided by wins plus losses, with pushes excluded; ROI is measured at a flat stake per read

The public record on the Pick Tracker uses the same population and the same period as the product, so the figure a visitor sees is the figure a subscriber sees. That matters more than the figure itself: a track record you cannot audit is marketing.

Where AI predictions fail

Being honest about the failure modes is part of the method.

• Stale prose: a breakdown written before a late scratch describes the earlier matchup — the win probability on the card follows the line, the cached text does not
• Confident prose about thin inputs: the quality tier reduces this but does not eliminate it
• Narrative over-fitting: a model can build a convincing story from a coincidence, such as a three-game head-to-head streak
• Market efficiency: when the line already contains the information, explaining it adds understanding but not an edge
• Variance: a 58% favorite loses more than four games in ten by construction, and a bad week is not evidence that the method is broken

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.

How to use an AI prediction well

1. Read the probability first and ask whether it matches your own sense of the game — if it does not, the breakdown should explain the gap
2. Check the data tier; a weak package deserves less weight regardless of how assured the prose sounds
3. Look at what the breakdown says the market may be missing, then verify it against the statistics page rather than taking it on trust
4. Log your own read before the game starts — the tracker closes fifteen minutes after kickoff — and let the grader score it
5. Judge the method on months of graded volume, not on the last three results

Summary

An AI sports prediction on Sportslyx is a market-derived win probability, a data package with a quality tier, a language-model breakdown that explains the number rather than inventing it, a pre-generation and caching cycle in which the probability follows the market line while the prose stays as written, and a grading job that publishes the result. Each stage is inspectable, and each has documented limits. For the full set of rules, see the methodology; to see the pipeline running on today’s games, start with the predictions board.