Bloomberg Sports Analytics projects Spurs to finish 5th

Bloomberg Sports Analytics is a consulting company that produces proprietary sports analysis which they sell to people who want to buy things like that. Presumably professional sports teams and the occasional news outlet. As a marketing strategy, they like to occasionally release interesting top-line numbers based on the underlying analysis they want to sell to people. This week, they released the results of their league projections for the English Premier League, the Bundesliga, and Ligue 1.

They are willing to say that their methods are advanced, but they say very little about what those methods are. I grant them their business plan, but I have to admit to some skepticism about the methods. They do say that they make use of some proprietary individual player value numbers in putting together the projections. Those latter adjustments are surely complicated, and BSA's individual player value methodology certainly incorporates a lot of Opta data. There's nothing on the site to clearly demonstrate the link between individual player Z-Score and team goals / team points, but the results are at least plausible. Since I have stuck mostly to the team level in my analysis, that's what I'm going to focus on here.

The table below is the Bloomberg projected table by points and goal difference. For clubs that were in the EPL last year, I have also included their 2012-2013 goal difference. For the three promoted clubs, I have listed their G/GA/GD from the championship, marked with a C.

As you can see, the table is arranged nearly perfectly in terms of goal difference. The two exceptions—City/United and Stoke/Norwich—actually demonstrate the method Bloomberg is using. Because Stoke scored and allowed so many fewer goals than Norwich, their goals ratio is actually worse even though their arithmetic difference is better. The points numbers are nearly identical to those produced by Howard Hamilton's Soccer Pythagorean method. So I hypothesize that Bloomberg is using projected goals, run through something very similar to Hamilton's Pyth method, as the basis for their points projections.

Update: As has been pointed out in the comments, the BSA Projected Tables FAQ notes that they are using a Monte Carlo simulation to produce projected points. I will be using a similar method, so this will be fun to track over the season.

I also think the similarities between the goal difference numbers, projected and from last year, are notably strong. The correlation coefficient is 0.95. The vast majority of a club's performance in Bloomberg's projections, then, can be explained by their 2012-2013 goal difference.

This is particularly notable to me because the work I've done this summer has suggested consistently that simple goals scored / goals allowed numbers are too dependent on random variation to be used as a primary basis for football team analysis. Using data like shots in the box on target and big chances, I have argued that Tottenham, Everton, and Southampton were a good bit better than their G/GA numbers, while Arsenal, Manchester United, and West Brom were a good bit worse. I see nothing in this table that correlates with any of the underlying stats.

Bloomberg seems to have basically used G/GA (or perhaps some more complex numbers that ended up summing to G/GA?) to project Spurs, Arsenal, Man United and Southampton. I don't get what happened with Everton and West Brom. West Brom's -4 GD reflected some good fortune, and Bloomberg has docked them. But they didn't dock United at all, and they've got Everton as a roughly .500 club. I would be interested to know what has caused the variations from G/GA projections that you see with a few clubs in the table—also Newcastle—but these are a minority of the clubs. For the most part, it's all about last year's goal difference.

It may be that Bloomberg have some great new methods and analytics. I don't know because I don't work for them. The player analysis that surely went into boosting the stats for Chelsea and Man City is at least interesting, though again all we have are the broadest of strokes. But if I had to speculate, I would hypothesize that their Premier League projections are based mostly on a relatively superficial approach focused mostly on goals scored and goals allowed from the previous season.

An Announcement

The summer of MCofA learns to code because he doesn't have to write his stupid dissertation anymore has ended with some real success. I have written a Monte Carlo simulator in Python which will allow me to run 1,000,000 simulations of the EPL season each week. I have renoobulated the method, as I will discuss in open-sourced detail later, and I believe I have solved the "too few draws" problem that vexed me last year. I will have EPL projections and power rankings every week based on these numbers.

However, the best simulator is only as good as the inputs you give it. I believe that after about ten weeks, the season data will be useful for projections. But what to do for the first ten? I have tried not to criticize Bloomberg here too much for their methodology, because I don't have a better one. I am extremely skeptical of player value analysis in football, but without player value stats you can't reasonably project a club like Chelsea or Man City, or really anyone at all who's made some changes to the roster.

So this is my plan. I will use two parts objective statistical analysis and one part crowd-sourced subjective analysis. I will ask, before the season starts, for projected EPL tables. It's too early yet to ask, since we don't know some very important information about club rosters, but please do get ready to rank. I'll be asking next week.