Predicting Saves? Good Luck!
– Bob Lemon (SP Cleveland Indians, 1946-58)
The volatility of major league bullpens is an issue that will forever haunt fantasy baseball players. It takes approximately 100 saves to win the category in your average fantasy baseball league. This means that you will need 3 full-time closers to even have a chance of winning the category. While drafting three closers will have a negative effect on your strikeout and win totals, closers will help in the other three pitching categories (Saves, ERA, WHIP). When drafting closers one must take into account all factors, not just the number of saves they amass. However, closers most significant contribution is to the saves category. The question becomes, can we accurately predict how many saves a closer will have?
It is widely agreed by most experts that saves are the hardest category to predict. This has led to the exclusion of saves in some prediction systems, such as ZiPS. Other systems that do predict saves, simply take a three year weighted average. From 2006 to 2007, the correlation between saves (using the top 32 closers from both years) was approximately 45%. In other words, if you were to base your saves predictions off of 2006, on average, you could predict the number of saves a player will earn with 45% accuracy. Clearly this is not nearly accurate enough.
Because predicting saves with methods like the one above is so inaccurate, players often attempt to over compensate by finding patterns that really do not exist. The most common of these is assuming that a closer on a good team will have more save opportunities. The chart below highlights the records and save opportunities of the 30 ball clubs from 2007.
As you can see, there is little correlation between team record and saves opportunities. If we take a look farther back (I’ll spare you the graphs) this non-existent pattern remains, well -- non-existent. One theory that seems to make slightly more sense concerns teams run differentials. A team that loses or wins by more than three runs can not, because of the definition of a save, earn a save during that game. For this reason, one might conclude that a team with a low (positive) run differential would earn more save opportunities. I’ve now incorporated run differential into the chart you saw above:
Pretty much the only correlation we see in the above chart is between run differential and record, except for the Arizona Diamondbacks which is just a stunning anomaly. Regardless, there is no correlation between the amount of runs a team scores and the amount of save opportunities a team gets.
However, save opportunities isn’t a completely useless statistic. In fact, there is a lot that can be done with save opportunities. Because save opportunities inexplicably vary so much from year to year, we can conclude two things:
1 Teams that have more save opportunities than average can be expected to regress towards the mean the following year.
2) Teams that have less save opportunities than average can be expected to progress towards the mean the following year.
The following chart contains SvOp data from 2006 and 2007 in order to demonstrate this phenomenon:
The fourth column is the most important one. As you can see, only four teams that had an above average amount of save opportunities in 2006 did not show some sort of regression in 2007. Likewise, only five teams that had a below average amount of save opportunities in 2006 did not show some sort of progression in 2007. While predicting the amount of progression/regression is still quite hard, understanding the concept is incredibly important.
With this information in hand, the most important data to look at is not saves, but rather save %. In other words, not how many saves a player earns, but rather what percentage of save opportunities the player converts. This is calculated using a very simple formula:
I will be discussing the various uses of Sv% over the coming weeks. However, what is important to glean from this article is that SvOp and likewise SVs, while not connected to the overall performance of a team are not random. The regression and progression shown in the above chart demonstrate that there is a pattern, even if it exists only in the arbitrariness of the data.
To conclude, it is important to realize the significance that this information has on projecting closers. Closers that amass a large amount of saves with a large amount of save opportunities are likely to show regression the following year. On the other hand, closers that convert a large amount of their saves, but have a limited amount of save opportunities, are likely to show progression the following year. The save category, like many traditional baseball statistics, is an incredibly poor indicator of performance. With this information in hand, Sv% will become a much more efficient way to project closers.
