AI/MLTwo systems, one pipeline
From the play
to the forecast.
STRYKE learns what each player does in each situation. MACHINE uses those models to play the game thousands of times before it starts. Both are built and run in-house, on data we collect ourselves.
- 01 · Source Play-by-play League event feed, box scores and availability, polled as they publish.
- 02 · STRYKE Players & lineups Possessions, shots, rotations and conditional player models.
- 03 · MACHINE Game simulation 10,000 possession-by-possession games per matchup.
- 04 · Output Distributions Win probability, margin, total, quarters, leads and comebacks.
Player and lineup intelligence Stryke Possessions rebuilt from raw events, empirical-Bayes shooting, lineup status that never passes a projection off as confirmed. 
Open STRYKE → Full-game Monte Carlo Machine Regular season so far: favourites 34–14 (70.8%), Brier 0.217. Every forecast is graded day by day. 
Open MACHINE →
Open STRYKE → Full-game Monte Carlo Machine Regular season so far: favourites 34–14 (70.8%), Brier 0.217. Every forecast is graded day by day. Open MACHINE → 01Rules both systems follow
01
No look-ahead
Every model and every run carries a cutoff. Nothing published after it can enter, and the guard is in the code, not in a checklist.
02
Labelled status
A forecast is either an original forecast, an experimental live run, or retrospective research. The label is assigned by the system and travels with the number.
03
Pre-registered changes
A new default has to beat the old one on a test written down before the results are seen, on games neither model trained on.
04
Uncertainty shown
Intervals, sample sizes and evidence weights sit next to point estimates. A thin sample is marked as thin.