GLSP Projections and Findings – 2017 Introduction
The regular season is almost here! Before you know it, you’ll be setting lineups and agonizing over start/sit decisions. Fortunately, we have a number of tools to help you with this process. My favorite are the GLSP Apps. GLSP, or “Gillespie,” stands for Game Level Similarity Projections. For further background and an overview of how to incorporate these projections into your weekly process, I’ve included some links at the end of this post.
Tyrod Taylor faces the Jets defense to start the season. There are a number of ways that we could predict his fantasy point total for this contest. One would be to look at historical results. Taylor has faced the Jets three times as a Bill, recording a wide range of outcomes.
|Year||Week||Opp||Comp||Att||Pct||Yds||Yds/Att||TD||Int||QB Rating||Att||Yds||Yds/Att||TD||Fantasy Points|
Given the small sample size and varying results, it’s hard to draw many conclusions from these games. This is where GLSP projections – yes, that’s redundant – come into play. Instead of focusing solely on Taylor, we could look at similar quarterbacks and how they’ve fared against the Jets. However, Taylor posts unique stat lines on a week to week basis, so the number of comparable passers that faced the Jets within a relevant time frame is low. To account for this, we can expand our search to include all games in which Taylor or comparable players faced defenses similar to the Jets. The results of these matchups provide a proxy for his range of outcomes.
To generate a list of comparable matchups, the apps use average stat lines from a user specified number of games, the default being the last 16 for the player in question. If we roll with the defaults and plug Taylor into the QB GLSP App, it will find quarterbacks that produced similar averages over a course of 16 games. The App will also search for defenses that produced stat lines similar to the Jets average. Next, the App compiles the 20 most comparable matchups and provides a low, median, and high projection. The low projection is equivalent to the 25th percentile point total from the comparable matchups. The median projection provides a benchmark, with even odds of the player producing more or less. The high projection is equivalent to the 75th percentile point total from the comparable matchups.
There’s a variety of different ways that you can use this information when setting your lineups. The important thing to remember is that these projections are based solely on historical matchups. As such, they are agnostic of player specifics such as offensive situation and health. Again, the low and high are based on percentiles only. So only 50 percent of the time (e.g. 75th percentile minus 25th percentile) is the player expected to produce a score between his low and high projection.
A New Wrinkle
The GLSP Apps also factor the game’s point spread and over/under into the comparable matchup search. As a result, we need to wait until the middle of the week to update the data and refresh the apps for the given week. Additionally, the apps allow users to exclude specific games from the average stat lines used to generate the comps. This provides an awesome level of customization but makes the process of sorting through every player time-consuming. To remedy this, I built GLSP models in Excel and will be publishing the results in a table format on Tuesday nights. When the site version of the apps are ready, I’ll compare and contrast with my results, search for surprising projections, and outline my findings in a Wednesday article.
I employed a similar methodology to the site apps while building mine. For each position, I started by identifying the statistics that were the most explanatory of historical scoring. From here, I reviewed the predictive nature of each and considered whether or not they carried from week to week. I used my findings to assign weights to the statistics included in the average stat lines. This allowed the searches to place greater emphasis on the more relevant stats. For example, QBR is significantly better at both explaining historical performance and predicting future performance of quarterbacks than interceptions. Within my QB GLSP, QBR is a significant factor in identifying comparable passers, whereas interceptions are not considered.
I spent a great deal of time back testing each positional GLSP, attempting to bring the absolute differences between the averages calculated by my GLSP projections and actual results to the lowest levels possible. In the end, I was able to shrink the differences down to approximately five points for each position. This means that, on average, a player’s average GLSP projection will have a difference of approximately five points when compared to actual results.
That might sound underwhelming, but I was really pleased with this outcome. When compared to more conventional projection sets that included subjective inputs and could account for player health, offensive specifics, weather, etc., these results were equivalent or better. Also, pinpoint accuracy wasn’t the goal of the exercise. Building a range of outcomes was. The results, however, did give me confidence that my process made sense and that the projections provided were reasonable.
Here’s the average stat line of the comparable matchups identified for Taylor versus the Jets. Range refers to the distance between the low and high projection; the greater the distance the wider the range of outcomes. “AVG RK” refers to the average weekly rank of the QBs within the identified comparable matchups.
|Player||Team||Opp||Range||Low||Med||High||AVG||Comp||Att||Yds||TD||Int||Att||Yds||TD||AVG FP||AVG RK|
I’m looking forward to sharing my results, leveraging the site versions of the apps, and making outcome based start/sit decisions this season!