00Sports analytics · case-based inference
What you know about the game, weighed against every game like it.
You bring the situation: a starter ruled out, a back-to-back on the road, a coach down to a seven-man rotation. BayesBook searches the historical record for the closest analogues and shows you how those situations actually resolved. Your context becomes a read, not a hunch.
P(H) The starting point
E What you know
~ Games that looked like it
P(H | E) Your read
Where the number comes from.
Start with how often the outcome happens in general: the betting line, or the long-run record. Add what you know about tonight. Your read is the first, moved by the second. The methods behind it include Bayesian updating, which is the formal name for that move.
P(H | E) =
P(E | H) · P(H)
P(E)
posterior = likelihood × prior / evidence
your read = the starting point, moved by the analogues
your read = the starting point, moved by the analogues
- P(H)
- The starting point. How likely the outcome is before tonight's details are in: the betting line, or how often it happens across the whole record. The prior.
- E
- What you know. The situation you described. BayesBook treats it as a fact about tonight, not an opinion. The evidence.
- P(E | H)
- What the analogues say. Among past games that looked like tonight's, how often did it go that way? The likelihood.
- P(H | E)
- Your read. The starting point, moved by what the analogues say. Shown with the games that produced it. The posterior.
Fig. 01 · starting point → your readwin probability, home side
Prior · the starting point
Posterior · your read, after the analogues
Fig. 02 · the twelve analogueshome-side margin, one bar per game
Home side won · 8
Home side lost · 4
Mean margin · +5.1
Every game, in one space.
Lay every past game out as a point, with games that resembled each other sitting close together. Pick k centres. Hand each game to its nearest centre, then move each centre to the middle of the games it was handed. Repeat until nothing moves. The methods behind it include k-means clustering, which is the formal name for that loop.
J =
kΣi = 1
Σx ∈ Si
‖x − μi‖2
cost = every game's distance to its centre, squared, added up
goal = the centres that make it smallest
goal = the centres that make it smallest
- k
- How many groups. Chosen before anything runs. The games get sorted into exactly k groups, each gathered around one centre.
- Si
- A group. Every game whose nearest centre is centre i. A game belongs to one group at a time and can switch between rounds.
- μi
- A centre. The average position of the games in its group. Each round it moves to the middle of whatever it currently holds.
- ‖x − μi‖2
- The miss. How far a game sits from its centre, squared so that far-off games count for more. Add every miss up and you have the cost.
Fig. 03 · k-means, k = 4each game tied to its nearest centre
Game
✕Centre · final
Centre · first guess
Path between rounds
Fig. 04 · the loop settlingcost by round
Sports analytics for your reads.
BayesBook is a tool, not a tout. We are not betting on it. The point is to make your own knowledge of a game legible against the record.
i.
You supply the context
The system has no opinion about tonight's game until you give it one. Whatever edge exists is yours; BayesBook measures it.
ii.
Evidence over narrative
Instead of a confidence score, you get the list of games. If the analogues are weak, you see that they are weak.
iii.
Priors are explicit
The baseline is stated, sourced and adjustable. A posterior is only as honest as the prior it started from.
iv.
Small samples, said plainly
Twelve analogues is twelve analogues. Interval widths are shown, never hidden behind a single number.
01Research preview
Bring a situation. See the record.
Access is limited while the situation taxonomy is still being refined. Tell us what kind of context you tend to have and what sports you follow.