Model Accuracy — NFL

How well do ranking models predict regular-season game outcomes?

Fraction of games where the predicted favourite won

higher is better

Mean squared error of probability predictions (lower is better)

lower is better

Cross-entropy loss of probability predictions (lower is better)

lower is better

Results by Season

Season	Games	Pick Accuracy	Brier Score	Log Loss
2026	17	64.7%	0.2191	0.6289
2025	271	63.5%	0.2262	0.6432
2024	272	68.8%	0.2132	0.6145
2023	272	59.9%	0.2342	0.6625
2022	269	62.5%	0.2259	0.6414
2021	271	58.3%	0.2344	0.6626
2020	255	66.3%	0.2176	0.6286
2019	255	64.3%	0.2232	0.6392
2018	254	63.0%	0.2214	0.6308
2017	256	64.8%	0.2164	0.6210
2016	254	64.2%	0.2205	0.6290
2015	256	65.2%	0.2258	0.6460
2014	255	68.6%	0.2069	0.6003
2013	255	65.5%	0.2164	0.6200
2012	255	62.0%	0.2186	0.6242
2011	256	64.1%	0.2109	0.6072
2010	256	62.5%	0.2319	0.6555
2009	256	69.5%	0.2042	0.5949
2008	255	64.7%	0.2161	0.6226
2007	256	65.2%	0.2128	0.6111
2006	256	57.4%	0.2376	0.6660
2005	256	67.6%	0.2106	0.6083
2004	256	64.1%	0.2237	0.6378
2003	256	64.1%	0.2207	0.6323
2002	255	59.6%	0.2273	0.6452
2001	248	65.3%	0.2288	0.6529
2000	248	62.9%	0.2223	0.6315
1999	248	62.5%	0.2223	0.6338

Each point is a 5% probability bucket. On the diagonal = perfectly calibrated. Above = model underestimates; below = overestimates.