MLBmoneyline_ingame (murphy reliability decomposition)
Model 0.005619; reference 0.003425; n=27351.
View Score decompositionCalibration / comparability gate
Read the published comparability decisions first.
This view shows exactly which reliability-component comparisons are supported by the committed evidence, and which measurements remain outside that axis.
Reading trail
Read first: Read each sport's calibration reliability before comparing components across sports.
Next question: How closely do published forecast bins align with observed outcomes?
Read the analysis: What carries across sports
Published measurement capability
Availability is derived from the published row fields. A null value remains not published; it is never treated as zero.
| Sport | Model reliability score | Reference reliability | Reliability decomposition | Source count |
|---|---|---|---|---|
| MLB | Published | Published | Published | Published |
| International soccer | Published | Published | Published | Published |
| NBA | Not published | Not published | Not published | Published |
| Tennis | Not published | Not published | Not published | Published |
Supported reliability-component comparisons
These are the only rows with model and reference Brier reliability components. Each component is the bin-weighted mean squared difference between forecast probability and observed outcome frequency. Values use squared-probability units; lower means less measured calibration error within these bins.
Model reliabilityReference reliability
Comparable reliability components
Each n counts scored probability rows, not unique games.
International soccermoneyline_ingame (murphy reliability decomposition)
Model 0.191982; reference 0.088088; n=4265.
View Kernel Transfer evidenceRows held outside the comparison
A common scale does not establish matched populations or transferability. These published rows do not have the compatible components required for this chart.
Source counts have different units: MLB combines scored pregame rows and in-game checkpoints; NBA counts scored checkpoints; tennis counts states. These counts can include repeated observations from a game and do not measure independent game samples.
MLB: totals_margin (CRPS distributional, pregame+ingame)
Missing or incompatible measurement: Murphy reliability/resolution for a binary outcome probability
CRPS scores a full run-count distribution, not a binary outcome probability -- Murphy reliability/resolution is undefined for it here; never forced onto the Brier axis above.
View Kernel Transfer evidenceNBA: winprob_ingame (Brier, no reliability decomposition)
Missing or incompatible measurement: 10-bin Murphy reliability/resolution split
nba's benchmark stores per-checkpoint Brier MEANS only -- no 10-bin Murphy reliability/resolution split has been computed locally for nba win-prob, honest gap not an estimate.
View Kernel Transfer evidenceTennis: pregame_prior + ingame_surface_gate (no market side)
Missing or incompatible measurement: market probability
both tennis gate receipts are model-vs-model checks (prior vs base; surface-specific vs surface-blind) -- no market_prob to decompose against, reliability undefined by construction.
View Kernel Transfer evidence