Mithrandir Metrics
Valinor Calibration research lane Refreshed Jun 1 06:00 AM ET
Calibration / Probability Refinement / Reliability · Calibration research lane

Valinor Model

Calibration and probability refinement. Not necessarily the highest-AUC lane - the one built to ask whether probabilities are shaped correctly.

Markets
6
Active
2
Date
2026-05-28
Method CatBoost or LightGBM with explicit calibration, stacked probability repair, CQR, beta/isotonic calibration, or Bayesian pooling. Status Calibration research lane Coverage 2 of 6 markets currently have a Valinor lane.
Definition

What this lane is

01
Method family
Valinor asks whether the probability is shaped correctly: center, tails, interval width, and reliability by bucket. HR is now the first production Valinor lane; other markets remain honest missing or research status.
valinor
02
Not this
Not a renamed Fangorn output, not a high-variance upside model, and not a hand-adjusted probability column without validation.
guardrail
Distinctness

Falsifiable test

What calibration or uncertainty problem is this lane solving? If it does not improve probability shape, tail behavior, interval coverage, or reliability, it is not Valinor.

Comparison first

The value of the three-lane system is comparison. No single lane is the right one in isolation.

Coverage matrix

Valinor across six betting families

MarketStatusMethodDate
Home Runs Production Production non-stacked LightGBM with beta calibration; promoted for HR after replay-window ECE and distinctness validation. 2026-06-01 Open
Strikeouts Missing Missing; planned boosted residual / tail-calibration layer over K distributions. 2026-06-01 Open
Total Bases Experimental Missing; planned calibrated ordinal/count distribution over total bases. 2026-06-01 Open
Hits Missing Missing; planned calibrated binary/ordinal hit probability model. 2026-06-01 Open
Pitcher Outs Missing Missing; planned conformal/calibrated recorded-outs distribution. 2026-06-01 Open
Futures Missing Missing; planned simulator calibration or Bayesian model averaging after exploratory work. -- Open