The Most Accurate NFL Prediction System

Last week I posted a link to ThePredictionTracker.com, a site that tracks predictions of many NFL prediction systems. I love this site because so many people are out there making predictions, but rarely do we ever get an honest look at how accurate they've been. It even includes accuracies from previous years.

I'm especially pleased because based on the records of the other systems, the one used here at NFL Stats has been the most accurate NFL prediction system since its inception. (Please forgive the self-congratulations.) The NFL Stats logistic regression efficiency model has been estimating game outcome probabilities over the past two seasons with the highest total accuracy rate of any system.

Admittedly, there is some amount of luck involved in predicting sports outcomes, which is why various systems should be compared over more than a single season. You might notice that the system that is doing the best so far this year (Greenfield's Team Rankings system) was among the very worst in 2006. This would suggest the system is prone to a very high degree of luck. Systems with true predictive value are still susceptible to luck, but should still rank above average fairly consistently.

In 2006 the NFL Stats model accurately predicted the winning team in 122 out of 195 games for a rate of 62.6%. So far in 2007, the model has been correct in 58 out of 82 games for a rate of 70.7%. The two-year total is 180 correct out of 277 games for an accuracy rate of 65.0%. (Although 65% doesn't sound very impressive, 2006 was an unusually difficult year to predict.)

I added the two-year totals for all the other systems listed at ThePredictionTracker.com. I included systems that made predictions in both 2006 and 2007 with at least 100 total games predicted. Only 3 of the systems were excluded. In total, there were 61 systems compared. The table below lists each system and its 2-year record.

System	Correct	Total Games	Accuracy %
BBNFLStats	180	277	65.0
Deslock Rankings	160	249	64.3
Pigskin Herald	170	269	63.2
ARGH Power Ratings	212	337	62.9
Hank Trexler	249	397	62.7
Least Squares	249	397	62.7
Logistic Regression	190	303	62.7
Mike Greenfield	172	275	62.5
Eric Hollobaugh	163	261	62.5
Roger Johnson	189	303	62.4
Brian Gabrielle	247	397	62.2
Sonny Moore	247	397	62.2
Beck Elo	246	397	62.0
Line (updated)	246	397	62.0
Pythagorean Ratings	246	397	62.0
Sagarin	246	397	62.0
Sagarin Elo	245	397	61.7
Stephen Kerns	245	397	61.7
System Median	245	397	61.7
Computer Adjusted Line	245	397	61.7
CPA Rankings	244	397	61.5
Least Abs. Val Reg	244	397	61.5
Steven Jens	244	397	61.5
System Average	244	397	61.5
John Coffey	169	275	61.5
CPA Retro Rankings	243	397	61.2
Lee Burdorf	243	397	61.2
Sagarin Predictive	243	397	61.2
Sports Report Elo	233	381	61.2
Viacheslav	233	381	61.2
NutShell Retro	223	365	61.1
Kambour Rating	242	397	61.0
Reistertown Index	242	397	61.0
Line (opening)	241	397	60.7
Scoring Effeciency	240	396	60.6
Pigskin Index	240	397	60.5
PerformanZ Ratings	239	396	60.4
Jeff Imes	237	393	60.3
Matthews Grid	161	267	60.3
Jeff Self	168	279	60.2
Covers.com	229	381	60.1
Kasulis Enhanced Spread	238	397	60.0
Dokter Entropy	228	381	59.8
The Sports Report	228	381	59.8
What If Sports	171	286	59.8
JFM Power Ratings	237	397	59.7
Stat Fox	237	397	59.7
Computer Adjusted Li	159	267	59.6
Bihl Rankings	147	247	59.5
Ed Bemiss	155	263	58.9
Ashby AccuRatings	233	397	58.7
Dunkel Index	233	397	58.7
Sport Trends	233	397	58.7
Geoff Rout	122	210	58.1
Tom Benson	230	397	57.9
LS - w/ team HFA	228	397	57.4
Kasulis Enhanced Spr	153	267	57.3
Grid Iron Gold	192	340	56.5
NutShell Sports	220	397	55.4
TSR Slots	206	381	54.1
Sports Report Pred	68	130	52.3

(One caveat: my model is based solely on current season data so no public predictions were made prior to week 4. Many of the other systems make predictions beginning at week 1. However, my model would have been 33/47-- 70.2% correct for weeks 1-3 in 2007, which would have actually improved the model's record. These weeks were not published and therefore I did not include them in the comparison above. Note: Week 1 was predicted using just home teams, and one game in week 3 was predicted 50/50 and was excluded. )