NFL Performance

How the official forecasts have actually done.

Forecasts only — not betting advice Details

Confidence is the forecast probability and nothing else (probability × 100). "Model is N points higher/lower" describes disagreement with the market, not a recommendation or a stake. Forecasts are estimates, not guarantees.

Quick Read

194 of 356 correct (54.5%), plus 4 pushes (excluded)

356 official predictions decided so far.

HOW SURE

With only 356 results, the true accuracy could plausibly be anywhere from 49% to 60%. That range narrows as more games are graded.

BEST

Best market so far: Moneyline — 69 of 119 correct (58.0%), plus 1 push (excluded)

WORST

Worst market so far: Spread — 58 of 117 correct (49.6%), plus 3 pushes (excluded)

CALIBRATION

Too ConfidentTOO CONFIDENTAverage forecast confidence 60.2%; actual hit rate 54.5%; the model has been too confident by 5.7 points so far.

PROBABILITY

Probability accuracy 0.253 — no better than a coin flip. (0.25 is what you would score by saying 50% on everything; lower is better.)

TYPICAL MISS

Average margin miss: 10.1 points · Average total-score miss: 11.2 pointsModel has projected margins about 1.8 points too high so far. Model has projected totals about 1.4 points too low so far.

ENOUGH DATA?

MeaningfulMEANINGFUL356 decided predictions is enough to draw real conclusions about forecast accuracy — though not about profitability, which needs far more.

By market

Moneyline: 69 of 119 correct (58.0%), plus 1 push (excluded)
Spread: 58 of 117 correct (49.6%), plus 3 pushes (excluded)
Total: 67 of 120 correct (55.8%)

How far off are the projections?

Average margin miss: 10.1 points
Model has projected margins about 1.8 points too high so far.
Average total-score miss: 11.2 points
Model has projected totals about 1.4 points too low so far.

When the model says it is confident, is it right?

Each row reads: what the model claimed, against what actually happened.

50-54% confidence: 51 of 104 correct (49%); model expected ~52%; worse than expected, and the sample cannot yet rule out chance
55-59% confidence: 50 of 90 correct (56%); model expected ~58%; about as expected, and the sample cannot yet rule out chance
60-64% confidence: 44 of 74 correct (59%); model expected ~62%; about as expected, and the sample cannot yet rule out chance
65-69% confidence: 26 of 47 correct (55%); model expected ~67%; worse than expected, and the sample cannot yet rule out chance
70-74% confidence: 16 of 29 correct (55%); model expected ~72%; worse than expected, but 29 results is a small sample
75-79% confidence: 5 of 10 correct (50%); model expected ~77%; worse than expected, but 10 results is a small sample
80+% confidence: 2 of 2 correct (100%); model expected ~82%; better than expected, but 2 results is a small sample

By confidence, by market

Confidence is the forecast probability, never edge. Each row is the picks whose stated probability fell in that bucket; pushes are excluded. THIN marks fewer than 30 decided predictions.

Moneyline

Confidence bucket	n	Right	Log loss
50-54	34	16/34 (47.1%)	0.7019
55-59	27THIN n=27	16/27 (59.3%)	0.6689
60-64	23THIN n=23	14/23 (60.9%)	0.6700
65-69	16THIN n=16	10/16 (62.5%)	0.6583
70-74	14THIN n=14	9/14 (64.3%)	0.6714
75-79	4THIN n=4	3/4 (75.0%)	0.5944
80+	1THIN n=1	1/1 (100.0%)	0.2191

Spread

Confidence bucket	n	Right	Log loss
50-54	34	15/34 (44.1%)	0.7072
55-59	34	16/34 (47.1%)	0.7153
60-64	29THIN n=29	16/29 (55.2%)	0.7018
65-69	12THIN n=12	8/12 (66.7%)	0.6334
70-74	6THIN n=6	3/6 (50.0%)	0.7906
75-79	2THIN n=2	0/2 (0.0%)	1.4920

Total

Confidence bucket	n	Right	Log loss
50-54	36	20/36 (55.6%)	0.6844
55-59	29THIN n=29	18/29 (62.1%)	0.6809
60-64	22THIN n=22	14/22 (63.6%)	0.6532
65-69	19THIN n=19	8/19 (42.1%)	0.8222
70-74	9THIN n=9	4/9 (44.4%)	0.8468
75-79	4THIN n=4	2/4 (50.0%)	0.8515
80+	1THIN n=1	1/1 (100.0%)	0.1758

When the model disagrees with the market

Historical validation on 2,095 games found this disagreement is not a usable betting signal. This asks the same question of the live record.

When the model agrees with market (<2%): 29 of 51 correct (56.9%) — 51 results, too few to conclude anything
When the model mild disagreement (2-5%): 31 of 61 correct (50.8%), plus 1 push (excluded) — 61 results, too few to conclude anything
When the model strong disagreement (5-10%): 54 of 93 correct (58.1%), plus 1 push (excluded) — 93 results, too few to conclude anything
When the model extreme disagreement (10%+): 80 of 151 correct (53.0%), plus 2 pushes (excluded)

Technical details — raw statistics

By market

Market	n	W-L-P	Hit rate [95% CI]	Brier	Log loss
Moneyline	119	69-50-1	58.0% [49–66]	0.2391	0.6711
Spread	117	58-59-3	49.6% [41–59]	0.2615	0.7184
Total	120	67-53	55.8% [47–64]	0.2589	0.7132

Projection error

Projection	games	RMSE	MAE	Bias
Margin	120	12.84	10.05	1.75
Total	120	13.87	11.17	-1.43

Bias is mean(projected − actual): positive means the model projected too high. RMSE punishes large misses more than MAE does; both are in points.

Calibration table

Confidence	n	Claimed	Observed	Gap
50-54	104	52.4%	49.0%	+3.4
55-59	90	57.6%	55.6%	+2.1
60-64	74	62.2%	59.5%	+2.8
65-69	47	67.2%	55.3%	+11.9
70-74	29THIN n=29	72.2%	55.2%	+17.1
75-79	10THIN n=10	77.1%	50.0%	+27.1
80+	2THIN n=2	82.1%	100.0%	-17.9

Green band = the range the sample actually supports (95% Wilson interval). Green line = observed. Red line = what the model claimed. A red line outside the green band is a calibration miss the sample is large enough to demonstrate; a red line inside it means the sample is still too small to tell.

Model vs market, by size of disagreement

Disagreement	n	W-L-P	Hit rate [95% CI]	Brier
agrees with market (<2%)	51	29-22	56.9% [43–69]	—
mild disagreement (2-5%)	61	31-30-1	50.8% [39–63]	—
strong disagreement (5-10%)	93	54-39-1	58.1% [48–68]	—
extreme disagreement (10%+)	151	80-71-2	53.0% [45–61]	—

By week

Week	n	W-L-P	Hit rate [95% CI]	Brier
1	45	21-24	46.7% [33–61]	0.2658
2	48	28-20	58.3% [44–71]	0.2525
3	48	29-19	60.4% [46–73]	0.2488
4	46	24-22-2	52.2% [38–66]	0.2683
5	43	26-17-2	60.5% [46–74]	0.2202
6	42	16-26	38.1% [25–53]	0.2882
7	42	24-18	57.1% [42–71]	0.2482
8	42	26-16	61.9% [47–75]	0.2321

By model version

Version	n	W-L-P	Hit rate [95% CI]	Brier
NFL-v1.1	356	194-162-4	54.5% [49–60]	0.2531

By data quality

Status	n	W-L-P	Hit rate [95% CI]	Brier
DEGRADED	41	17-24-1	41.5% [28–57]	0.2874
HIGH	270	156-114-3	57.8% [52–64]	0.2458
Historical / quality not recorded	45	21-24	46.7% [33–61]	0.2658

"Historical / quality not recorded" means those predictions were written before the data-quality framework existed, so no status was recorded. It is not a quality classification, and no quality has been inferred for them.

Computed from the official prediction per game and market — the final valid pregame production snapshot — joined to final scores from the schedule source rather than to the log's own grade, so a grading bug cannot hide from this page.
Generated 2026-11-05 23:48 UTC.