How the Ghost predicts
Every probability on SoccerGhost is worked out from football: results, the chances behind them, the players on the pitch. None is worked out from a bookmaker's price. Prices are compared against the Ghost's numbers, never mixed into them, so when the two disagree you are seeing two different opinions rather than one echoing the other.
The result, goals, both teams to score and the correct score
Each side gets an attacking and a defensive rating from three years of its competition's results, the most recent counting most (a two-year half-life), with a home advantage measured for that competition. Ratings are blended with what each side's chances say: its expected goals (xG) and shots on target over the last year. Recent form moves them a little; once both line-ups are confirmed, a side missing its regulars is marked down by how much of its usual strength it names. A side new to its competition starts where promoted sides have tended to, not at the average.
The ratings give a grid of every scoreline and its chance (a Dixon–Coles model). Every goals market is read off that one grid: the result, double chance, over/under, both teams to score, the correct score and each team's goals. That is why they never contradict each other.
Corners, cards and shots
Corners, cards, shots and shots on target have models of their own: each side's figures for and against in the same competition, pulled toward that competition's level where a side has few matches. For cards, the named referee's record moves the forecast, where we hold enough of it. These forecasts are shown for reading; they are never published as calls on the Ghost Record until their own record says they should be.
International football
National sides are rated in a pool of their own, apart from club football, with the venue decided match by match: a home match, a neutral ground, or a side playing at home away from home.
How it is checked
Nothing ships until it beats what it replaces on matches it has never seen: every change is tested walk-forward on held-out seasons, and kept only if the improvement is larger than the noise. The live numbers are then watched every night: what the Ghost said against what happened, by kind of call and by competition.
Every call the Ghost publishes is written to the Ghost Record before kick-off, in a chain in which each line seals the one before it, so no call can be edited or quietly removed afterwards. The losses stay on it with the wins.
What it cannot do
No model knows the result. A 60% call loses four times in ten, and it should. On the calls settled so far, the bookmakers' own probabilities have been slightly more accurate than the Ghost's. We say so here and on the Ghost Record rather than hide it: the Ghost is a calibrated second opinion built from football, not a way to beat the bookmakers.
Betting loses money on average. If it stops being entertainment, here is where to get help.