The model

The NBL model

Five short notes on how the NBL model works, where its forecasts turn into a real edge, and where they don't. Written for anyone who wants to check the working.

Summary

The model prices every NBL game to a full probability distribution, and those probabilities are well calibrated. On a market this efficient, being calibrated is not itself an edge, and we don't claim it is.
One pocket of betting edge clears our bar: the handicap bets our system serves. On the pooled 2024–2025 sample that is about +16% per dollar staked, with a bootstrap 95% interval of [+6%, +20%] that excludes zero, and it clears the same bar on the 2025 season on its own, which the model never saw during development.
There is no demonstrated edge on the head-to-head market or on the bets the system declines to serve.
Every number here is from the systematically backtested model, which prices each game off default line-ups and public information only. It is the weakest-informed version of what we run. Live, we feed it player-outs, replacements and team news by hand before it prices a bet: information the backtest never sees. Read these results as a lower bound.

Served handicap bets

Served handicap bets at best cross-book odds, out of sample, 2024 to 2025. Stake sized by model conviction.

Everything on these pages is out of sample. Each season is predicted by a model trained only on earlier seasons, and 2025 was set aside and scored exactly once, after every design decision was frozen. That is about as close to a live record as a backtest gets, but it is still a backtest. The record we publish and log accurately starts the day you subscribe. See the record →

Notes

If you want to know more about our processes, email [email protected].

A basketball arena lit for a night game, the court empty

Abstract illustration of many grey market lines and one amber model line diverging

Where the model has an edge

The one result that clears our bar, the ones that don't, and what kind of edge it is.

$A hand working through equations on paper in blue ink$

How the model is built

The predictive distribution, the information it runs on, and how a signal earns its place.

A symmetrical building atrium under a glass ceiling

How the model is tested

The data split, walk-forward scoring, the leak audits, and the bar a result has to clear.

A hand writing entries in an old expense ledger

The numbers

A decade of out-of-sample accuracy, the 2025 holdout, the equity curves, and the caveats.

A price chart in pale blue on a dark monitor

Closing-line value

Why we can post strong CLV on demand, why we don't, and why it's a poor gauge of our signal.