Skip to content

NFL model accuracy

National Football League

Every final NFL game, forecast with only what was known the day before and then scored. Lower Brier and log loss are better; a coin flip scores 0.250 Brier. All leagues

Scorecard

Accuracy of each model on final games
ModelGamesBrierLog lossAccuracySkillCalib. error
Game simulator (backtest)1200.2400.67755.0%+5.1%8.4%
Elo1200.2430.67854.2%+4.0%9.9%
Base rate1200.2530.70251.7%—2.2%
Published forecasts15small sample0.2260.64866.7%+7.9%25.8%

Calibration

When a model gives the home side a 60% chance, does it win about 60% of the time? Dots on the dashed line mean yes; whiskers show the 90% range the observed rate could land in by chance.

Brier score by month

Early-season forecasts lean on last season's regressed ratings, so expect them to improve as games accumulate.

Head to head

Game simulator (backtest) vs Elo

120 games both forecast

Brier
0.240 vs 0.243
Log loss
0.677 vs 0.678
Accuracy
55.0% vs 54.2%

Game simulator (backtest) has the lower Brier score here.

Published forecasts vs Game simulator (backtest)

15 games both forecast

Brier
0.226 vs 0.194
Log loss
0.648 vs 0.581
Accuracy
66.7% vs 73.3%

Game simulator (backtest) has the lower Brier score here, but on too few games to call.

Market consensus isn't connected: it needs a licensed odds source, which an admin can import. Until then the base rate and Elo are the benchmarks.

How this is measured

The backtest replays the game simulator for every final game from the efficiency ratings and Elo as they stood before that date, the league's scoring and home advantage so far, rest and travel, the venue and the game's weather. Injuries and lineups of past games aren't on record, so the replay is team level; the published forecasts, which include them, are scored separately as they accumulate.

Brier = ½ Σ (probability − outcome)² over home, away. Log loss = −ln(probability given to the result), with probabilities clipped at 0.5%. Tied finals are left out. Full details are in the model methodology document.

Pick record

Every forecast the site published before a game, scored once it finished. A pick is the likeliest outcome. A well-calibrated model's 70% picks come in about 70% of the time, so compare each row's hit rate with its confidence.

Record

10–5 67%

Average confidence 56%

Hit rate by confidence
Confidence PicksHit
Under 60%1362%
60%–70%2100%
Latest picks
GamePickResult
Nov 1 TEN @ CINCIN 68%36–43
Nov 1 BAL @ BUFBUF 57%10–36
Nov 1 ARI @ DALDAL 58%33–30
Nov 1 MIN @ DETMIN 56%15–19
Nov 1 IND @ JAXJAX 58%7–42
Nov 1 LV @ NYJNYJ 53%38–33
Nov 1 CLE @ PITCLE 53%6–27
Nov 1 ATL @ TBTB 50%0–16
Nov 1 LAC @ LALA 60%9–24
Nov 1 KC @ DENKC 53%24–28
Nov 1 NE @ MIANE 64%24–19
Nov 1 PHI @ WASWAS 54%10–27
Oct 29 CAR @ GBGB 55%19–27
Oct 26 DAL @ PHIDAL 52%38–21
Oct 25 KC @ SEASEA 51%3–20

Data as of Source: nflverse

Cookies

Cookies keep you signed in. Optional ones remember your settings. No ads or tracking. Cookie policy