How the model works

1. Every side gets a rating

Each nation starts at the same rating. After every match the winner takes points from the loser. Beat a side you were expected to beat and you gain a little; beat a side rated far above you and you gain a lot. Home sides get a fixed bonus before the comparison. The size of the swing and the home bonus were chosen using matches played before 2017 and then frozen, so the record from then on is an honest test, not a fit.

2. Ratings become a win probability and a fair price

The gap between two ratings maps to a win probability: a side rated 100 points above another is expected to win about 64% of the time on neutral ground. The fair price is simply one divided by that probability. At 64% the fair price is $1.56; a bookmaker offering $1.70 is offering more than the model thinks the side is worth, and one offering $1.45 is charging over the odds.

3. Margins and totals

A straight line fitted to past matches turns the rating gap into an expected margin, and another into an expected total. The spread around each line comes from how past matches with a similar gap actually finished. From that spread the site works out the chance of covering any line or beating any total you type in. These rest on smaller samples than the win probability, so treat them as rougher.

4. The record is walk-forward

Every prediction in the record used only matches played before it. The model never sees a result before it predicts it. The record page shows accuracy, calibration (when it says 70%, does the side win about 70% of the time?) and the typical miss on margins and totals, with the bad stretches left in.

• 489 matches from 2008 to 2026
• Favourite won 65.6% of the time
• Brier score 0.203 against 0.250 for a coin toss (lower is better)

What it cannot see

• Team lists, injuries and suspensions. It rates the jersey, not the seventeen players in it on the day.
• Weather and ground conditions, which matter most for totals.
• Any bookmaker's price. A gap between the model and the market can mean value, or it can mean the market knows something the model doesn't.
• Anything about club rugby league yet. The NRL model goes live for Round 1, 2027, with its own record page from the first round.

Caveats on the record

• This is a record of predictions against results, not a betting record. No historical bookmaker odds for internationals are in the data, so it says nothing about profit.
• International sides play a handful of Tests a year, so ratings react slowly and every figure here rests on a small sample compared with a club competition.
• The model sees only scores, dates and venues. It knows nothing about team selection, injuries, suspensions, weather or how much a side has changed since it last played.
• Many mismatched games are easy to call, which flatters the headline accuracy. The calibration table is the fairer test.
• The rating settings were chosen using matches before 2017. Results from 2017 onwards were not used to choose them.