ApexPredict football 1X2 model card
Versioned evidence and limitations for the analytical system behind published calls. Updated 26 August 2026.
Intended use
Pre-match analysis of football fixtures. As of 30 September 2026 the public competition list shows 65 active competitions; predictions are generated only in those that passed the league readiness check, so a listed competition can show fixtures without predictions. Published markets are the 1X2 match result, over/under 2.5 goals and both teams to score. Over/under 1.5 and 3.5 goals are shown for information only: they are never a shortlist pick, a Telegram pick or an accumulator leg. The model ranks outcomes and may abstain. It is not a guarantee, wagering service or authority for corners, cards or penalties.
Match result (1X2)
Where bookmaker prices exist, the published 1X2 probabilities are those prices with the bookmaker's margin removed; since September 2026 the market carries the full weight (1.0). Where no price is available, the probabilities come from a chronological Elo rating built on point-in-time fixture history. Goal, roster and language-model context is advisory unless it passes the stored evidence and publication policy. PostgreSQL publication records—not an explanation generator—are product truth.
Over/under 2.5 goals
When at least three bookmakers quote the market, the probability is their consensus: each book's margin is removed proportionally and the results are averaged, and a side becomes a pick only at 60% or more. Such a bookmaker-priced pick is made only in leagues whose own past bookmaker prices passed a calibration check: enough history, with picks at 60% or more landing about as often as they stated. Elsewhere the probability is shown without a pick. Without three quotes, the probability comes from the goal model, and a pick is allowed only in leagues where the goal model passed its readiness check. Every published over/under 2.5 and both-teams-to-score pick is checked again at a fresh price in the last hour before kickoff. If it has fallen below its threshold, it is replaced by a no-pick version; the original pick stays on record and is never edited. If the prices are missing or the refresh fails, the pick stands as published and is recorded as not re-checked. Performance figures grade every pick as it was first published, including picks a later check replaced.
Both teams to score
Uses the league-gated goal model. Bookmaker prices for this market have been recorded since 25 September 2026; a calibration check in mid-October 2026 decides whether it moves to bookmaker consensus as well.
Validation evidence
1X2: over 760 Premier League fixtures the published blend is indistinguishable from the margin-free market baseline (log loss +0.00044, with a confidence interval that includes zero). We have not shown an advantage over bookmaker prices. Over/under 2.5: across 20,783 fixtures in 18 leagues, seasons 2023/24 to 2025/26, market consensus probabilities scored a log loss of 0.6763 against 0.6868 for the goal model, which is why the market is used whenever three books quote it.
Untouched holdout
On the 380-match 2025/26 Premier League holdout, daily walk-forward Elo selected 181 outcomes correctly (47.63%); the closing market selected 188 (49.47%). These historical samples do not predict future accuracy.
Known limitations
The Elo rating used when no prices exist structurally does not rank a draw first under its current bounded draw equation, although its draw probability is well calibrated over the larger validation set. Current injury, lineup and timestamp-matched market evidence may be absent. Missing required evidence causes watchlist or abstention, never an invented default.
Publication and change policy
Only a recommended result that passes every evidence check may enter the main feed. Original probabilities, model identity and cutoff are preserved. Withdrawals and supersessions remain visible with time and reason; material post-kickoff edits are rejected.
Evaluation
Public reporting includes sample size, accuracy, Brier score, log loss, calibration gap and losing streak. ROI is shown only for samples with complete immutable publication-time decimal odds. Results are never hidden because they lost.
Method details and reproducible research commands are maintained with the prediction-engine source. A model replacement requires versioned holdout evidence, shadow evaluation and explicit promotion review.