METHODS & PERFORMANCE
Model Research
Explore model evidence, forward tests, and first-points rankings. Historical replay and saved pregame performance are different kinds of evidence.
Checking pipelineWaiting for API status
SIX-MONTH HOLDOUT PASSED
first-basket-fga-proxy-v1
Legacy field-goal target only · not validation of the new research previews
HOLDOUT822Jan–Jun 2026 · 3 excluded
TOP-ONE14.60%120 correct
BRIER0.8963uniform 0.9000
LOG LOSS2.2850uniform 2.3026
LOCKED ≥15% RULE
Legacy field-goal watchlist
Each selection is saved before scheduled tip and never recalculated. Results are attached only after the opening possession is processed.
PROSPECTIVE VALIDATION
New-season evidence
Exploratory threshold locked at 15.00% · not a high-confidence label| Game | Model selection | Probability | Captured | Result |
|---|---|---|---|---|
| No prospective watchlist selections yet. Qualifying picks will appear when confirmed starters are available. | ||||
This is a forward test of a rule chosen from historical results. The model, 15% threshold, lineups, jumpers, timestamp, and full ten-player ranking are preserved with each snapshot.
V2 FORWARD TEST
Legacy field-goal model comparison
V1 remains the production model. Two frozen candidates record their own picks before tip-off so we can measure how they perform on the same games.
PREGAME COMPARISON
V1 alongside two V2 candidates
Longer history gives a player's older opening shots more weight. Scorer blend adds a player's past first-basket rate. Both use 75% model signal and 25% equal probability across the starters.
Loading saved pregame predictions…
| Model | Top pick hits | Accuracy | Brier | Log loss | ≥15% picks: hits / picks |
|---|---|---|---|---|---|
| Loading model comparison… | |||||
Lower Brier and log loss mean better probability estimates. Changes shown beneath each score compare V2 with V1 on the same games. These candidates were selected using last season's data; this forward test starts with new pregame captures.
Latest saved picks
| Game / captured | V1 · Production | V2 · Longer history | V2 · Scorer blend | Actual first scorer |
|---|---|---|---|---|
| No three-model captures yet. | ||||
WALK-FORWARD REPLAY
Choose a stored slate
The latest available historical date loads automatically. Each replay excludes the selected game and every game after it from training.
HISTORICAL SCHEDULE
Games by date
Finding stored dates…| Matchup | Final | Evidence | |
|---|---|---|---|
| Loading the historical slate… | |||
FIRST POINTS · FREE THROWS INCLUDED
Three opening predictions
Select any stored game.Earlier dates only. These previews are not saved pregame snapshots. Possession and method estimates have not yet been validated prospectively.
FIRST POINTS · FREE THROWS INCLUDED
First points ranking
The same first-points forecast as the three opening cards.
The complete ranking will appear here after a game is selected.
| Rank | Starter | Team | First points % |
|---|
The same first-points probabilities as Three opening predictions; made free throws count. The verified actual scorer is highlighted. Historical replays were reconstructed later, not saved before tipoff. Ties are not forced into picks and unknown results remain unresolved.
LEGACY FIELD-GOAL MODEL METHODOLOGY
Evidence first. Probability second.
This describes the legacy field-goal model, not the first-points ranking above. That candidate was frozen after the development window, then beat the 10% uniform baseline on both proper scoring metrics across 822 unseen games from January through June 2026.
- 01Opening tipEstimate the selected jumper matchup from prior tip results.
- 02Team pathCondition each team’s first-score rate on winning or losing the tip.
- 03First shotEstimate which confirmed starter takes the team’s first field-goal attempt.
- 04RestrainNormalize ten starters, then blend the signal 50/50 with uniform probability.