How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 2810 | 0.596 | 67% |
| Base rate | 2810 | 0.690 | 55% |
| Recent form | 2810 | 0.636 | 63% |
Against the betting markets
Each market's price before the start, scored on the same finished games as the model. A negative figure means the market was more accurate.
| Market | Since | Priced | Model log loss | Market log loss | Model vs market |
|---|---|---|---|---|---|
| Bookmakers (Pinnacle, DraftKings, bet365) | 19 Jan 2026 | 839 | 0.561 | 0.540 | −4.0% |
| Polymarket | 24 Apr 2024 | 2774 | 0.597 | 0.581 | −2.7% |
| Kalshi | 15 Apr 2025 | 1569 | 0.595 | 0.576 | −3.2% |
The playoff odds play the real format: seeds 7 to 10 go through the play-in, then a fixed bracket of best-of-seven series with home court to the better record. A series already under way is continued from its current score.
Offensive and defensive ratings, pace and the four factors are in the Teams explorer, from box scores. They sit beside the rating; they do not move the forecast.
Rest, back-to-backs and travel are not in the model.
Who is playing is judged from box scores and the injury report, not from lineups announced before tip-off, so a late scratch can be missed.
Before a season starts, the player and team stats are last season's.
Questions and answers
How accurate are PodiumOracle's NBA predictions?
On 2,810 games from held-back seasons (2024–25 to 2026–27), which the model never saw while it was tuned, the side it favoured won 67% of the time. Its log loss was 0.596, against 0.690 for the base rate and 0.636 for recent form. Against the betting markets on the same finished games, its log loss was 0.561 against the bookmakers' 0.540 over 839 games, 0.597 against Polymarket's 0.581 over 2,774 games and 0.595 against Kalshi's 0.576 over 1,569 games (lower is better).