#### Data

To test this, I'm using 247Sports composite scores to determine recruit quality and S&P+ ratings to determine team performance. The composite scores include all recruits instead of just the last year's, and S&P+ is a rating system that analyzes all plays and determines each team's opponent-adjusted effectiveness so it theoretically has less bias and noise than something like the AP rankings.

I'm using the top-50 recruiting classes from 2015 - 2018 throughout.

I'm using the top-50 recruiting classes from 2015 - 2018 throughout.

#### First test - S&P+ vs recruiting score

Below is a plot of S&P+ rating for a team/season combo vs the recruiting composite score for that same team/season combo along with a linear fit to the data:

The r^2 there is 0.39. You can hover over each point to see info on it. An obvious question is 'which teams are the biggest outliers?'. The teams that did the worst relative to their recruiting scores in that time frame are:

Team | Year | S&P+ vs expected |
---|---|---|

Florida State | 2018 | -20.2 |

Rutgers | 2016 | -19.3 |

Louisville | 2018 | -16.9 |

Tennessee | 2017 | -15.8 |

Rutgers | 2015 | -15.4 |

The teams that did the best relative to their recruiting scores in that time frame are:

Team | Year | S&P+ vs expected |
---|---|---|

Wisconsin | 2017 | 19.4 |

Washington | 2016 | 16.1 |

Louisville | 2016 | 14.8 |

Missouri | 2018 | 14.7 |

Ole Miss | 2015 | 12.9 |

There are a lot of obvious limitations here. Good coaches get good recruits and also have successful teams, so this is obviously not the only factor. Still kind of cool. What about other obvious metrics?

#### Next test - S&P+ vs last year's S&P+

Below is a plot of S&P+ rating for a team/season combo vs the previous season's S&P+ rating for the same team.

The r^2 there is 0.68. That's much better than the recruiting one. Do those two together give you an even better fit?

#### Recruiting score and last year's S&P+ to produce current S&P+

We can use linear regression to see how much each term helps. Using a simple model of last year's S&P+ rating and recruiting score as explanatory variables for this year's S&P+ rating yields the following summary:

'recruits' is the recruiting score and 'old_sps' is last year's S&P+ score. It's pretty interesting. The simple interpretation of this limited data set is that if you know the previous year's S&P+ rating, adding the recruiting info to the model doesn't actually improve its accuracy (if you recall, 0.68 for the r^2 value was what we got for just last year's S&P+ values earlier). The p value is also quite high for the recruits term.

#### Summary

You can get a better than nothing estimate of team performance simply from knowing the team's roster. You can get a better estimate than that simple from knowing the prior year's performance. The most important takeaway though is that Tennessee greatly underperformed their recruiting classes in 2017 and 2018.

## 0 comments:

## Post a Comment