Model performance

Two kinds of evidence, kept apart on purpose: the LIVE ledger below grades published picks against real results as they settle; everything under it is the out-of-sample record each model was promoted with.

The Ledger

Every pick

Every published pick is recorded before its event and graded after it. Nothing is deleted, nothing is regraded.

2,396picks recorded before their event
2,314 graded from official results, 82 still open — none edited, none deleted

Accuracy: 50%
Return at the prices published: -163.7u
Brier score: 0.231

Sport	Record	Hit	Brier	CLV
MLB	1094–1100	50%	0.231	+1.0
NFL	16–16	50%	0.268	—
UFC	9–8	53%	0.225	-0.6

82 picks recorded and awaiting settlement. Brier: 0.25 is a coin flip, lower is sharper. CLV: probability points gained on the closing line — positive means the market moved toward our number.

The governor — every market judged by the closing line

A market earns the right to claim an edge by beating closing prices over at least 40 measured picks with a confidence interval clear of zero — thresholds set in code before any data existed. Verdicts advise; nothing promotes itself silently.

! marks a market where the closing line and the money disagree. Both columns are measured; when they conflict, trust the return and treat the verdict as suspect — the “closing” price is the last one archived before the start, a median 3.07 hours early, and a line taken that far out is easy to beat and says little.

Market	Verdict	Record	CLV	Return	n
Runs batted in · MLB	edge rejected	3052–3026	+1.24pp	-8.0%	487
Runs scored · MLB	edge rejected	2931–2680	+0.81pp

Unvalidated

no significant edge over baseline

CRPS improvement over baseline

Higher is better. Muted bars did not reach significance at p < 0.05 on a paired block bootstrap over weeks.

Walk-forward results per market against an empirical baseline
Market	Learner	CRPS	Baseline	Delta	p	Log loss	Calib. error	Cov 80	Rows
Anytime TD anytime_td	lgbm	0.167	0.183	+9.0%	< 0.0005	0.4997	0.0165	0.984	28,938
dst sacks dst_sacks	lgbm	0.958	1.019	+6.0%	< 0.0005	0.6354	0.0137	0.899	3,230
dst takeaways dst_takeaways

Did the claimed edge mean anything?

Every recommendation is ranked and gated by an estimated edge. These are the graded picks grouped by the edge we claimed at the time, against what a flat unit on each actually returned. If the estimate carries information, the bottom row beats the top one.

Edge we claimed	Picks	Hit	Return	95% interval
under 3%	211	47.9%	-10.7%	-25 to +4
3–5%	706	50.7%	-5.4%	-13 to +2
5–8%	908	51.9%	-4.3%	-11 to +2
8–10%	418	45.2%	-15.4%	-25 to -6

No relationship worth naming. A bigger claimed edge did not return more than a smaller one, on the sample so far. No threshold has been changed on the strength of this table. The bands were fixed before the numbers were read and are not re-cut; a rule moved to fit what has already happened is not a rule.

What we claimed, against what happened

Every curve on this page comes from the walk-forward harness and describes a model. This describes the published numbers — scored against the append-only record of what was issued and how it settled.

Every published pick

claimed 53.4% · hit 49.8% [49.1–50.5] n=18,831

Over 18,831 settled picks over 39 event days, these claimed 53.4% and hit 49.8%. The claim sits OUTSIDE the [49.1%, 50.5%] interval on the outcome, so the 3.5 point overstatement is more than sampling noise. Read a published percentage with that in mind.

Recommended picks only

claimed 56.3% · hit 49.9% [47.8–52.0] n=2,243

Over 2,243 settled picks over 35 event days, these claimed 56.3% and hit 49.9%. The claim sits OUTSIDE the [47.8%, 52.0%] interval on the outcome, so the 6.4 point overstatement is more than sampling noise. Read a published percentage with that in mind.

Nothing here is corrected for. Correcting means fitting, and fitting on a record this shallow means fitting to a handful of slates — the record spans 39 event days, and the floor set for touching a published probability is ten. The gap is shown instead.

Calibration, drawn

The curves behind the scalars, from out-of-sample walk-forward predictions — never re-scored in-sample, which would draw flattery. Reliability: claimed probability against observed frequency; a calibrated model hugs the diagonal, and point size is sample count, because a bin of nine and a bin of nine thousand are different kinds of evidence. PIT: twenty bars that sit flat at the dashed line when the whole predicted distribution is shaped right — a U is overconfidence, a hump is timidity, a slope is bias.

anytime_td

ECE 0.018 · PIT 0.026

reliability

Probabilistic metrics are computed at a grid of realistic lines, not at sportsbook prices. They measure whether the model's probabilities are calibrated. They do not measure profit.

Model performance

The Ledger

The governor — every market judged by the closing line

CRPS improvement over baseline

Did the claimed edge mean anything?

What we claimed, against what happened

Every published pick

Recommended picks only

Calibration, drawn

anytime_td

dst_sacks

dst_takeaways

extra_points

fg_made

first_td

idp_sacks

kicking_points

last_td

pass_att

pass_cmp

pass_int

pass_longest

pass_td

pass_yds

points_allowed

rec_longest

rec_rush_yds

rec_yds

receptions

rush_att

rush_longest

rush_yds

solo_tackles

tackles_assists