MET-AL · Ensemble Verification Views (SYNTHETIC)

Global controls

Dispersion knob δ 1.00 (calibrated)

δ rescales member offsets about the ensemble mean (mᵢ′ = mean + δ·(mᵢ − mean)). δ<1 under-dispersive · δ>1 over-dispersive.

Lead time spread grows with lead

Region subset

Member count N 20

Case subset (first-k) 180 cases sampling sensitivity

Event & binning

Event threshold P(obs ≥ τ) 288.0 K drives reliability + contingency

Reliability bins 10

Spread-skill bins 8

Decision prob. p* (contingency) 0.50

apply √((N+1)/N) spread correction MET small-ensemble parity option

Rank (Talagrand) histogram

MET RHIST

U-shape = under-dispersive · dome = over-dispersive · flat = calibrated · slope = bias. Ties broken MET-style (random placement). Expected = total/(N+1).

Spread-skill

MET ECNT

Bin by ensemble spread; per-bin RMSE via ratio-of-sums on SL1L2 MSE. Point area ∝ case count. On 1:1 = calibrated; above = under-, below = over-dispersive.

Reliability diagram

MET PCT · PSTD

Forecast prob = member fraction ≥ τ. Diagonal = perfect; dashed = no-resolution (climatology ō) & no-skill. Inset = sharpness. Brier with REL − RES + UNC decomposition.

Stretch · contingency + CRPS

MET CTC · CTS · ECNT

2×2 from thresholding forecast prob at p*: a=hits, b=false alarms, c=misses, d=correct neg. Scores derived from raw counts (correct aggregation), exact brief formulas. CRPS minimized near calibrated.