Predictive Metrics - Outline

1. Intro
2. Sticky Stats
3. Defining the Algorithm
4. Testing the Algorithm
5. Adjusting for Age
6. Testing the Model

In my previous post Intro to Predictive Metrics Intro to Predictive Metrics, I demonstrated that certain metrics are actually highly correlated to fantasy scoring, others, not so much.

But correlation to fantasy scoring isn’t all that is necessary to establish which metrics are most predictive. Critically you need to know how likely a player is to recreate a given performance for a specific metric, or how “sticky” that stat is year-to-year. In order to establish which receiving stats are most “sticky”, I compiled player data from past ten years as mentioned in the previous post. I then sorted the data to include only players who had at least a 9-year career.

I then performed an analysis to determine the standard deviation of the year over year change in end-of-year-finish at each metric. Recall that standard deviation is a measure of volatility. For example, I tested a player’s standard deviation in the number of targets per game they received one season as compared to their finish the following season. The same with receptions per game, targets, and so on.

What I found is that some metrics were absolutely sticky, both year over year as well as across a player’s career; while others were too volatile to be reliable as a predictive stat. After all, it doesn’t matter how well correlated that metric may be to fantasy scoring if a player can be expected to vary to a significant degree in that metric on a year-to-year basis.

In this sample set receiving yards per game as well seasonal reception totals and receptions for first downs, are all metrics that a that are simply too volatile to be of any particular predictive utility. That relative degree of volatility year over year is represented in the chart below as defined by the stat's year over year Standard Deviation.

However, efficiency metrics tend to be more consistent year-over-year, as you can see from the relatively low degree of volatility for first down receptions per game (1D/G), targets per game, and receptions per game.

The stickiest stats of the group: receptions per game, catch percentage, and first down receptions per game.

If you combine this "stickiness" data with the correlations to fantasy scoring that we determined in the previous post, this then gives us two metrics upon which to compare our stats. Again in terms of their predictive value:

• Year over year volatility (stickiness)
• Strength of correlation with standard fantasy scoring (on per game basis)

The following chart shows clearly how much value we should be giving to each individual stat, in terms of how predictive of fantasy scoring it may be.

In this chart you want to see values with a high correlation to standard fantasy scoring, or data positioned higher vertically on the chart. You also want to see data that has a low year over year volatility, or data that is positioned furthest to the left on this chart. Thus data points positioned furthest to the upper left of the chart above have the lowest degree of year-over-year fluctuation and are the most highly correlated to fantasy scoring.

You can see from this chart that not only are first down receptions per game highly correlated to fantasy scoring, but of this group it has the least volatility year to year.

Targets per game were next most correlated, but demonstrated quite a bit more volatility than that of first down receptions per game. Receptions per game, of which first down receptions per game are obviously a subset, are also very correlated and show less volatility than targets per game.

From this presentation we can conclude that first down receptions per game is going to be our best bet in terms of a predictive metric. Targets and receptions per game are two efficiency metrics that are also of some predictive utility. While highly predictive, touchdowns scored per game are too volatile to use as a predictive metric.

Next we need to see if there is a similar degree of correlation between a player's final efficiency metric for a given sseason and their fantasy points for the following season, which is where the real predictive strength occurs.