How accurate?

A plain answer with the numbers behind it: our own backtest, the home-team baseline from 6,140 games, and our live record as it builds.

The short answer

In an honest replay of the last three NBA seasons, where the model only ever saw games already played, our current model (v2.3) picked the winner 68.0% of the time. Just picking the home team gets 55.3%. A good NBA model is right about two games in three. Anyone claiming far more over full seasons is counting selectively or measuring something else.

67.5%

2023-24 walk-forward · Brier 0.207

68.6%

2024-25 walk-forward · Brier 0.207

67.8%

2025-26 walk-forward · Brier 0.206

55.3%

just pick the home team, 6,140 games since 2021

0–0

our live 2026-27 record starts Tue Oct 20

What the numbers mean

One season proves little

An NBA regular season has 1,230 games. A model whose true hit rate is 68% will land anywhere from about 66.7% to 69.3% in a season on luck alone (one standard deviation is about 1.3 points). So a model that went 70% once has told you little; a model that holds 67 to 69% for years has told you a lot.

Accuracy is the wrong headline anyway

Two models can both go 68% while one says "55%" on every game and the other says "85%". The second claims to know more than it does. The better measures are calibration (when we say est. 70%, do those teams win about 70% of the time?) and the Brier score, which punishes confident misses more than cautious ones. Lower is better: a coin flip scores 0.250; our backtest scores 0.207.

Why not 80%?

Injuries announced an hour before tip-off, rest days, and plain shooting luck. Many NBA games are decided by a few possessions. In our tests, knowing exactly who played added little on top of good ratings; the game itself is the limit.

How to judge any model's claim

1. Were the picks published before tip-off, and can you see all of them? Ours are stored before every tip-off and every graded pick is on the Record page.

2. Over how many games? Anything under a few hundred is mostly noise.

3. Was the backtest honest? "Walk-forward" means each day's picks used only games already played. Ours also passes a leak check that rebuilds sampled games from earlier data only.

4. What's the baseline? "68%" means little until you know the home team won 55% of games since 2021.

5. Are the chances calibrated? Ask for the table: stated chance by bucket against how often those picks won. Ours fills in live on the Record page from opening night.

We don't show market prices on NBA pages. Every number here is our own model's estimate, never betting advice.