How Nocturn Sport calculates its odds

The models behind every price on our sportsbook, explained for a curious bettor — no black box.

Updated October 4, 2026

Nocturn Sport calculates its odds in-house. We don't copy prices from other bookmakers and we don't buy them from an odds supplier. Each sport is priced by a model suited to the way it is scored: a goals model, a points model or an Elo rating. This page explains each one in plain words, how the margin is added, how we test the models and why markets close 1 minute before the start.

From rating to probability to price

Every price goes through the same three steps:

1. A rating for each team or player, learnt from past results.
2. A probability for every possible outcome of the event, derived from those ratings.
3. A price, which is the probability with a bookmaker margin applied.

What changes from sport to sport is step 2 — the shape of the scoring.

Goals model: football, ice hockey, baseball

In these sports, scores are small whole numbers that arrive irregularly. Each team has an attack strength and a defence strength. Home attack against away defence (plus home advantage) gives the home side's expected goals; the reverse gives the away side's.

Those expected goals feed a Poisson distribution, the classic model for counting rare events. With 1.6 expected goals, Poisson says the chance of scoring zero is about 20 %, one about 32 %, two about 26 %, and so on. Combining both teams gives a probability for every scoreline — a full score grid.

Two refinements:

• Football gets a Dixon-Coles style correction. Pure Poisson slightly underestimates low draws like 0-0 and 1-1; the correction adjusts those cells of the grid.
• Baseball uses a negative binomial distribution instead, because runs come in clusters (a four-run inning) and spread out more than Poisson allows.

From the grid we read 1X2, totals (1.5, 2.5 and 3.5 in football), both teams to score and handicaps. Because all markets come from one grid, they stay mutually consistent.

Points model: basketball, American football, rugby

When a game has dozens or hundreds of scoring events, the final margin and the total behave like a bell curve. Each team has attacking and defensive ratings in points; with the league's home advantage they give an expected margin (say home by 4.5) and an expected total (say 47 in the NFL).

We model each as a normal distribution. The share of the margin curve above zero is the home team's win chance (overtime included); the share above 4.5 prices the −4.5 spread; the total curve prices the over/under. Rugby also gets a 1X2, with the draw taken from the small slice of the margin curve around zero.

Elo ratings: tennis, MMA, League of Legends

In duels and esports, we use Elo, the rating system invented for chess. After each match, the winner takes points from the loser — more for an upset, fewer for an expected result. The rating gap converts to a probability: a 100-point gap is about 64 %, 200 points about 76 %.

Elo gives the chance to win one game, map or set. We then compute the series. Example: a 60 % chance per map gives about 65 % to win a Bo3 and 68 % to win a Bo5. The same calculation produces the series winner, the ±1.5 map or set handicap, total maps and correct score.

Adding the margin

A bookmaker's price includes a margin. We apply it proportionally: every outcome's probability is scaled up by the same factor, so price = 1 ÷ (probability × (1 + margin)). An outcome at 50 % and one at 20 % are shaded by the same relative amount, rather than loading the margin onto the long shots. The margin level is set per operator. Prices are also kept between a minimum and maximum odds limit.

Testing: does it work?

A model is only useful if it's calibrated — if events it calls 60 % happen about 60 % of the time. Measured on past matches:

• The football model was fitted on 10,783 matches, then tested on 1,973 later matches it had never seen.
• The basketball model was fitted on 8,665 matches and tested on 1,712 unseen ones.

On both tests the models beat a naive "base-rate" forecast — one that just uses how often home wins, draws and away wins happen in general — scoring a lower log-loss. Log-loss punishes confident wrong forecasts heavily, so beating the baseline means the ratings add real information. It does not mean every result is predictable; no model makes that true.

Why markets close 1 minute before the start

Our prices are pre-match forecasts. Once play begins, the information changes by the second, and a pre-match price would no longer be fair to anyone. Closing 1 minute before the start lets you bet on confirmed line-ups and late news while keeping every bet a pre-match bet. In-play pricing — the same models updated with score and clock — is in development; see live betting.

You can compare our probabilities with your own on nocturnsport.com and in our match previews. For reading prices as probabilities, see odds formats.

Also on this site