How Sportslyx Analytics Works

Market de-vig to fair probability. A moneyline pair carries the bookmaker’s margin: the two implied probabilities add up to more than 100%. Sportslyx removes that margin proportionally across the two sides so the pair sums to exactly 100%, and the favoured side’s share is the win probability shown on the game. For soccer the read is priced on the 1X2 line: the two side prices from the three-way market are de-vigged against each other, the draw is left out of the probability, and a drawn match is graded as a push. UFC and golf are individual events; there is no home side and no venue adjustment, only the market for each fighter or the outright market for each player.

Team and player statistics. The per-game package aggregates the public statistics that matter for the sport — efficiency, pace and rest in basketball; probable starting pitchers, run scoring and pitching in baseball; goaltending, special teams and shot metrics in hockey; form, goals and shots per match and fixture congestion in soccer; striking and grappling profiles in MMA; strokes gained, recent finishes and made-cut rate in golf. A baseball analyst also thinks about the bullpen and the park, and a soccer analyst about expected goals; those are not feeds Sportslyx ingests, so they appear in a breakdown only as context, never as data. These numbers describe the matchup; they do not override the market probability. They are shown on the statistics pages for every team the feeds cover.

Studio custom weights. Studio lets a subscriber choose the statistics they believe in, assign weights and run a live game through that model. The projection is the subscriber’s own; it is saved alongside the market number and the AI read, and it is graded after the game like anything else. Studio does not add hidden factors — the weights you set are the whole model.

Combos from one score distribution. A same-game combo is priced from a single model of how the score is likely to unfold, so correlated legs — a favourite’s moneyline and an under, a run line and a team total — are priced together rather than multiplied as if they were independent. The product page on Combos explains the pricing in more detail.

What it is. The percentage shown on the overview, on the matches board, on a game card and inside the opened breakdown is the market-derived win probability described above — the de-vigged share of the favoured side. All four surfaces read the same source, so a game cannot show one number on the board and a different one in its analysis.

What it is not. It is not the language model’s self-reported certainty, it is not a proprietary rating, and it is not an estimate of how likely the analysis is to be right. Language models are poorly calibrated about their own certainty; Sportslyx therefore does not surface that figure at all, and asks the model to write around the market number instead. What the model output does carry is a data-quality tier — rich, partial or weak — which says how complete the inputs were.

Calibration. A calibrated 60% should win about six times in ten over a large enough sample. Because the number comes from an efficient market, it inherits the market’s calibration, which is good, and the market’s errors, which exist. The monthly track record is the calibration check: if reads at a given probability win noticeably more or less often than the probability implies, that is visible in the record rather than hidden behind a rating.

• Delays. Public feeds lag the real world. A late scratch, a goaltender change or a pitching change can arrive after the breakdown was written, and the text is not rewritten when it does: the percentage on the card follows the line, the written read does not. Breakdowns are pre-generated for the upcoming slate and cached, and nothing regenerates them when the market moves. Check the availability panel and the line before you rely on the text.
• Incomplete feeds. Some competitions carry no shot data, some cards no full fighter records, some golf fields no strokes-gained history for a debutant. The quality tier records this; it does not fix it.
• Coverage edges. Player props are covered for MLB only. UFC is priced on two markets — the fight moneyline and the round total — so the method of victory is discussed in the prose but never priced. Golf has one market, the outright winner, and its previews are published for information only: golf never enters the graded record.
• Model error. Language models can misread an input or assert something that is not in the package. Structured inputs, a fixed brief and output validation reduce this; nothing eliminates it.
• Small samples. A month of reads in one sport can be a few dozen games. Hit rates over samples that small swing widely for reasons that have nothing to do with skill, which is why the record is shown by month and by sport rather than as a single headline number.
• Market efficiency. Because the probability is derived from the line, it cannot systematically outperform the market it comes from. The value of the product is in the reading, the consistency and the record, not in a claimed edge over the price.
• The draw-push rule. Soccer reads are priced on the 1X2 line but a draw is graded as a push. A true draw-no-bet price would be shorter, so the published soccer ROI is higher than a two-way market would produce. This is a deliberate, disclosed convention, not an oversight, and it is repeated on every soccer page.