clock menu more-arrow no yes

Filed under:

2013/14 PL: 8 Game Splits And Predictive Values

New, comment

This article is part 2 of a brief look at the first 16 games of the season and what splitting those first 16 games into two groups can tell us about certain statistics and their predictive value. It must be said that 8 games is an absolutely tiny sample. This is article wishes well; and is helpful rather than definitive.

In part 1 of 8 Game Splits I looked at how certain stats - shots +/-, SoT +/-, Scoring% etc - were repeatable from weeks 1-8 of the 13/14 season to weeks 9-16 of 13/14 season. Some of these team level stats were more repeatable than others.

A quick recap:

Relationship Between Any Given Stat In Games 1-8 and Games 9-16

R2
Sot For 0.68
SoT +/- 0.59
Goals For 0.56
Shots +/- 0.55
Shots Against 0.53
SoT For % 0.42
Shots for 0.37
Points 0.35
Goal Difference 0.31
SoT Against 0.21
Goals Against 0.13
Scoring% 0.01
SoT Rating% -0.001
PDO -0.05
SoT Prevention% -0.14
Save% -0.22

So we know which stats were repeatable from the first chunk of 8 games to the second 8 games of the season. That is all well and good, but what we really want to know is which stats are strongest at predicting future Points Won or future Goal Difference (GD).

To find out which stats are the most predictive of future performance we simply take the first 8 weeks (games) of any given stat and see how well that stat predicts future (games 9-16) Points and Goal Difference.

The higher the r2 correlation coefficient the better predictor the stat is of future performance.


Predictive Value

Scoring%

Pv_sc__medium

Well, weeks 1-8 of scoring% have virtually zero predictive value toward Points and Goal Difference from weeks 9-16. Remember a low r2 is bad.

Save%

Pv_sv__medium

A fractionally higher r2. But still, there is no predictive value here.

PDO

Pv_pdo_medium

PDO is a combination of save% and scoring% so it comes as no surprise that its predictive value is close to nil.

Shots On Target Rating

Pv_sot_rating_medium

Well, well.

Shots on Target Rating is explained here:

Week_10__alternative_pl_tables___bitter_and_blue_medium

So a teams first 8 games of shots on target rating is ever so slightly predictive of a teams points(0.17) and Goal Difference (0.19) in the next 8 games. But quite frankly the r2 is so weak that it is likely just noise.

Shots +/-

Pv_shots_medium

Shots +/- in weeks 1-8 is a pretty decent predictor of Points (r2=0.44) and Goal Difference (r2=0.32). But shots +/- isn't doesn't have the strongest predictive value of all the stats we look at.

This stat does....

Shots On Target +/-

Pv_sot_medium

Shots On Target +/- in weeks 1-8 is a fine predictor of Points in weeks 9-16 (r2=0.6) and even better predictor of Goal Difference (r2=0.76). I'm not particularly sure why there is such a large discrepancy between the predictive values for points and Goal Difference, someone smarter than I will have to figure it out. Anyhow, Shots On Target was a really strong predictor of future performance in that second batch of 8 games.

As stated in the intro to this piece, 8 game samples are a quite inadequate sample sizes when searching for concrete evidence of which stats have the highest predictive value. Some of these high, or low, predictive values may well be caused by early season noise and variance.

The point of this article was to point out in simple layman's terms which stat, at week 8 of the season, was the best predictor of future performance in terms of points and goal difference. And that stat is Shots on Target +/-.

I'll return to this article at the seasons end and re-evaluate the predictive values.

Full Chart

r2 Points GD
SoT +/- 0.6 0.75
SoT For 0.62 0.47
Shots +/- 0.44 0.32
Shots For 0.56 0.3
SoT For % 0.3 0.3
Shots Against 0.26 0.25
SoT Rating% 0.17 0.19
SoT Against 0.12 0.14
SoT Prevention% 0.03 0.02
Save% 0.01 0.02
Scoring% 0.005 0.008
PDO 0.02 0.006