# Football Forecasting - Summary & Final Thoughts - EdsCave

Go to content

## Football Forecasting - Summary & Final Thoughts

Football Forecasting

In the previous pages we have examined a couple of football game forecasting methods, ranging from utterly trivial to moderately sophisticated, with the underlying feature that the only input information is the final scores for each game, and which teams played.  The following table summarizes the accuracies of the different methods when applied to the 2012,2013 and 2014 NFL seasons.

 Method % Games Called Correctly (wks 5-20) 2012 2013 2014 2012-14 Flip a Coin (presumed accuracy) ~ 50 % ~ 50 % ~ 50 % ~ 50 % Home Team wins 56 % 58 % 57 % 57 % Previous Games won % 63 % 60 % 66 % 63 % "        " with home-team adjustment 64 % 61 % 68 % 64 % Mean Points Per Game 66 % 54 % 63 % 61 % "        " with home-team adjustment 68 % 60 % 66 % 64 % Margin-of-Victory Model 66 % 63 % 68 % 66 % Offense-Defense Model 66 % 65 % 64 % 65 %

One thing that jumps out is that even though a season has ~250 games to predict, random chance still plays a part in prediction accuracy. One can see a few percentage points variation in accuracy for each method from season to season. Evaluating a prediction algorithm's effectiveness over even shorter periods, such as a single game week, or even half a season will give even more sporadic results.  To get a reliable measure of predictive accuracy,  any algorithm being considered for serious use should be back-tested over at least several years of data.

Another question I am sure you are asking yourself, (and I have asked myself) is how good these algorithms really are? To this point I have just compared' algorithm vs. algorithm'. The real test is 'algorithm vs. football expert'.  Fortunately, ESPN provides a series of game predictions from a group of human experts and a crowdsourced expert (Pick'em). This data can serve as a meaningful benchmark.  Here is a summary of % correct calls for the 2015 season covering weeks 5-16.  Although ESPN's experts started at week 1, I adjusted the performance from week 5 to provide a direct comparison to the algorithms, which require a few weeks of startup data.

 2015 Weeks 5-16 Forecast RankingForecaster data from ESPN NFL Expert Pickshttp://espn.go.com/nfl/picks Forecaster % Correct Calls Mortenson 65.0 % Hoge 65.0 % MOV Model 63.8 % Jaworski 63.3 % Pick'em 62.7 % OD Model 62.7 % Schlereth 61.0 % Caplan 61.0 % Joyner 60.5 % Ditka 60.5 % Carter 60.5 % Golic 59.9 % Johnson 58.8 % Jackson 58.2 % Allen 57.1 % Wickersham 55.9 %

As can be seen, both the Margin-of-Victory and Offense/Defense models provide comparable performance to human experts. The major downside of the algorihtmic models is that they can't provide the play-by-play analysis or game narrative that are such a big part of the sport.  Somehow, I have the feeling that 'robo-analyst' will not be part of the game for a long time.

And for the final time, the information provided on these pages is intended for entertainment purposes only, and not for purposes of gambling, as the accuracy of the models presented here is not adequate to 'win' (make money) on a consistent basis in Las Vegas style gambling. If any of these forecasting models were that good, what would I be posting them here for? I would be on the Boston-Las Vegas flight every Friday night during football season making some serious money!!!

2 JAN 2016