TopBets

← Dashboard

How to read a bet

For every game, the model works out each result's chance. It suggests a bet only when the price pays more than that chance is worth.

1Barça win -150 · Barça v Getafe 21:00
2Model 68% · DraftKings 58%
3Value +13% · stake 2% · lower risk
  1. What to bet, and the price. -150 means you stake $150 to win $100.
  2. Its chance: the model's, then the bookmaker's. The model gives Barça more chance than the price implies. That gap is the reason for the bet.
  3. How good, how much, how safe. Value is the expected return per dollar. The stake is a share of your betting money. Risk comes from how likely the bet is to win.

The words you'll see

Value
What the bet returns on average, by the model's chance. Positive means the price looks generous.
Stake
How much of your betting money to put on it. Never more than 2%.
Risk
Lower: the model gives it 55%+. Medium: 40%+. Higher: less.
Confidence
For a prediction: strong at 60%+, lean at 45%+, otherwise a toss-up.
Prediction only
The model's likeliest result, but the price isn't worth a bet.
Skipped long shot
The bookmakers give it under 25%. That's where the model is least reliable, so it isn't sent as a pick.

Before you bet

How the chances are made

Results so far are on the Track record page.

Technical details

1. Team ratings

A Maher (1982) attack/defense model, not Dixon-Coles yet at this stage. Each team gets one attack multiplier and one defense multiplier, relative to the league average — not split by home/away venue; home advantage is captured separately, as the league's own real average home and away goals, fit directly from results. Built fresh from every finished match of this season and the five before it:

  • Recency. Each match is weighted by exp(-ln2 × age_days / 365) — a one-year half-life, fading out smoothly rather than dropping off a hard cutoff. A team's strength is mostly stable from one season to the next; the 45-day half-life used before left six or seven matches per team, which is mostly noise.
  • History. The scores of the five previous seasons, plus this season's matches and the previous season's last 50 with their xG. A long half-life needs the history to go with it.
  • Expected goals (xG) blending. Where a provider has xG for a match, the goals feeding the rating are 0.5 × actual + 0.5 × xG, so a scoreline built on one great chance and two deflections counts as weaker form than the result alone suggests. Falls back to the actual score when xG isn't available (real coverage is ~84% on API-Football, not 100%, and the earlier seasons are scores only).
  • Starting from average. Every team's goals are averaged together with three matches at the league average, so one or two results can't speak for a whole team: without it, a promoted side whose only match was a 0-0 is rated as never scoring and never conceding. The pull fades to nothing once a team has a season on record. A team with no matches at all gets a neutral rating.
  • Transfermarkt prior (currently enabled, weight 0.15). Nudges attack up and defense down together (or vice versa) for a team whose squad market value sits away from the league's average — log-scaled, capped, and fading as evidence builds up (to about a third of its starting size after 8 recent matches' worth, and under a tenth after 20; a result from a year ago counts as half a match). Only applied to a team that already has at least one finished match in the ratings window; a team with zero games is still purely neutral until its first result arrives. The project's docs/model.md has the snapshot format and freshness rules.

2. Expected goals

λ_home = league_avg_home_goals × home_attack × away_defense (and the mirror for λ_away) — each side's own attack times the opponent's defense, scaled onto the league's real home/away scoring baseline.

3. Availability

A side missing key players gets its λ discounted. Attackers and midfielders reduce attack; defenders and goalkeepers worsen defense — each missing player weighted by their share of the squad's minutes this season (never more than one eleventh: nobody plays more than that), the closest available proxy for "how much do we miss them." A player out since before the season has no minutes in it and carries last season's instead, as a share of a full season. The combined discount is capped so several simultaneous injuries can't overcorrect. This step can't be checked against past seasons (nobody records who was unavailable on a past date), so the track record keeps every prediction with and without it and reports which scored better.

4. Scoreline grid

Independent Poisson(λ_home) × Poisson(λ_away) for every scoreline up to 8-8, with the Dixon & Coles (1997) ρ correction applied to the four low-scoring cells (0-0, 1-0, 0-1, 1-1), where plain independent Poisson is known to under-predict draws and over-predict 1-0s. This same grid is what the same-game parlays and the "most likely score" pick read directly, so those always stay consistent with the Home/Draw/Away percentages by construction — they're never a separate calculation.

5. How it is checked

scripts/backtest.py replays past seasons match day by match day, each match predicted from earlier days only, and scores the result against the bookmakers' closing odds. Over La Liga 2015-16 to 2025-26 (4,180 matches, on scores alone) the model's log-loss was 0.979 against 0.960 for the closing odds and 1.065 for the league's base rates; it called 52.8% of results against the odds' 54.4%. Its percentages are honest — outcomes it gives about 85% happen 84% of the time — it is just less sharp than the market, which beat it in all eleven seasons. Staking the gaps between the two lost more than betting blind.

Live matches are kept on the same footing: every prediction is recorded before kickoff beside the books' lines and scored on the track record page.

6. What was removed, and why

Head-to-head. The model used to shift expected goals toward whichever side had dominated a pair's past meetings. Those scores mostly restate which side is stronger. With the short memory the ratings had then, that was useful; with ratings that remember, adding it double-counts strength, and the replay scored worse with it. (The custom scoring parlay still shows a pair's head-to-head goals; the model does not use them.)

The 70/30 blend with the league average. Every prediction used to be pulled 30% of the way to "what usually happens in this league", to tame the overconfidence of short-memory ratings plus head-to-head. It also put a ceiling and a floor on everything — no home win above about 83%, no away win below about 9% — so a long shot the market priced at 6% always looked like value at 11%. Without the two things it was correcting, the raw percentages come true about as often as they say.

What's tuned, what isn't

The half-life, the number of earlier seasons and the three league-average starting matches were chosen on the 2015-19 seasons of that replay and checked on 2020-25. The 0.5 xG blend and the -0.1 Dixon-Coles ρ were tested there and left as they were. The injury discount and the Transfermarkt prior could not be tested — history has no record of either — and are provisional. A "strong" pick is not automatically a good bet: a heavy favourite's short price often gives negative expected value even when the model agrees with the market.