Demo -- read-only snapshot of the live system as of 2026-07-16T01:45:31Z; no live data, paper units only.
Skip to main content
← receipts / scoreboard

mlb

The ledger row, decomposed. Every figure below is read from a staged artifact on disk (cited per panel); this drill describes calibration and sharpness only -- no dollar edge, ROI, or bankroll is claimed.

model
1.3449 (CRPS)
market
2.4109 (CRPS)
delta (95% ci)
1.0660 [0.2656, 1.9602]
n
21

verdict UNDERPOWERED -- source scripts/platformkit/benchmarks/crps_market/last_run_ingame_mlb.json

Murphy decomposition (model)

RELIABILITYreliability = 0.0056 (lower better)0.0056RESOLUTIONresolution = 0.0872 (higher better)0.0872UNCERTAINTYuncertainty = 0.2499 (baseline)0.2499

Brier = reliability - resolution + uncertainty (10-bin). Lower reliability and higher resolution are better; uncertainty is the irreducible base-rate floor. Descriptive, not an edge claim. Source: murphy_decomposition.json.

reliability curve (model vs market)

as of 2026-09-16
.00.00.25.25.50.50.75.75.00.00market: predicted 0.049, observed 0.021, n=3233market: predicted 0.148, observed 0.121, n=2344market: predicted 0.248, observed 0.159, n=1876market: predicted 0.351, observed 0.237, n=1526market: predicted 0.453, observed 0.345, n=2703market: predicted 0.550, observed 0.493, n=4129market: predicted 0.649, observed 0.682, n=2957market: predicted 0.746, observed 0.704, n=2579market: predicted 0.852, observed 0.863, n=2449market: predicted 0.949, observed 0.975, n=3555model: predicted 0.051, observed 0.063, n=2072model: predicted 0.156, observed 0.046, n=1900model: predicted 0.245, observed 0.106, n=1774model: predicted 0.358, observed 0.235, n=2217model: predicted 0.458, observed 0.371, n=3859model: predicted 0.543, observed 0.552, n=5636model: predicted 0.647, observed 0.643, n=4026model: predicted 0.746, observed 0.822, n=2024model: predicted 0.849, observed 0.941, n=2406model: predicted 0.952, observed 0.980, n=1437PREDICTED

Predicted probability (x) vs observed frequency (y) against the perfect-calibration ideal (dashed). Amber = model, blue = market. Hover a point for its bin count. Source: calibration_stability.json.

bootstrap CI per bin (model)

as of 2026-09-16
binngapgap 95% ciexcludes 0?
0.0-0.120720.0119[-0.037, 0.082]no -- within noise
0.1-0.21900-0.1099[-0.153, -0.046]yes -- gap real
0.2-0.31774-0.1384[-0.216, -0.045]yes -- gap real
0.3-0.42217-0.1232[-0.230, 0.014]no -- within noise
0.4-0.53859-0.0867[-0.220, 0.051]no -- within noise
0.5-0.656360.0087[-0.103, 0.122]no -- within noise
0.6-0.74026-0.0043[-0.147, 0.134]no -- within noise
0.7-0.820240.0752[-0.038, 0.177]no -- within noise
0.8-0.924060.0927[0.030, 0.138]yes -- gap real
0.9-1.014370.0282[-0.006, 0.049]no -- within noise
3/10 eligible bins have a calibration-gap 95% CI (cluster bootstrap on game_id) that excludes 0; the rest are within noise. Model Brier 0.1668. Brier skill vs climatology 0.333, vs market -0.104 (at or below 0 -- the model does not beat the market Brier; that null is the point). Source: calibration_stability.json + brier_skill_scores.json.