Intro
The following is an assessment of the agreement between 2021 regular season play and 2021 season win totals. The basic unit of analysis the time average lead for each game played in 2021.
Using 2011-2020 data, I made a simple model estimating the expected win percentage for a given time average lead (and also account for whether game is home or away game). As noted in a previous post, the time average lead for a game is a metric that summarizes its overall competitiveness.
Ten Least Likely Wins
| Wk | Team | Opponent | Time Avg Lead (Points) | Avg Win Percentage (%) |
|---|---|---|---|---|
| 5 | BAL | vs IND | -9.0 | 6.5 |
| 16 | CHI | @ SEA | -6.3 | 12.2 |
| 17 | TB | @ NYJ | -5.6 | 14.5 |
| 2 | TEN | @ SEA | -5.6 | 14.6 |
| 9 | BAL | vs MIN | -5.9 | 15.2 |
| 5 | PHI | @ CAR | -5.4 | 15.6 |
| 18 | SF | @ LA | -5.4 | 15.6 |
| 2 | BAL | vs KC | -5.7 | 16.2 |
| 5 | NE | @ HOU | -5.1 | 16.5 |
| 1 | KC | vs CLE | -5.3 | 18.0 |
Graphical Summary
Table
| Tm | Wins | Pythag Wins | Time Avg Wins | Pythag Win Luck | Time Avg Win Luck |
|---|---|---|---|---|---|
| ARI | 11 | 10.5 | 10.8 | 0.2 | 0.5 |
| ATL | 7 | 4.9 | 7.1 | -0.1 | 2.1 |
| BAL | 8 | 8.4 | 8.1 | -0.1 | -0.4 |
| BUF | 11 | 13.1 | 12.3 | -1.3 | -2.1 |
| CAR | 5 | 5.7 | 7.3 | -2.3 | -0.7 |
| CHI | 6 | 5.9 | 6.9 | -0.9 | 0.1 |
| CIN | 10 | 10.5 | 8.7 | 1.3 | -0.5 |
| CLE | 8 | 7.9 | 9.5 | -1.5 | 0.1 |
| DAL | 12 | 12.2 | 10.3 | 1.7 | -0.2 |
| DEN | 7 | 8.9 | 7.9 | -0.9 | -1.9 |
| DET | 3 | 5.1 | 5.3 | -2.3 | -2.1 |
| GB | 13 | 10.4 | 10.6 | 2.4 | 2.6 |
| HOU | 4 | 4.1 | 6.3 | -2.3 | -0.1 |
| IND | 9 | 10.6 | 11.1 | -2.1 | -1.6 |
| JAX | 3 | 3.4 | 4.0 | -1.0 | -0.4 |
| KC | 12 | 11.2 | 11.2 | 0.8 | 0.8 |
| LA | 12 | 10.6 | 10.3 | 1.7 | 1.4 |
| LAC | 9 | 8.8 | 9.3 | -0.3 | 0.2 |
| LV | 10 | 6.9 | 8.2 | 1.8 | 3.1 |
| MIA | 9 | 7.6 | 9.0 | 0.0 | 1.4 |
| MIN | 8 | 8.5 | 9.8 | -1.8 | -0.5 |
| NE | 10 | 12.4 | 9.7 | 0.3 | -2.4 |
| NO | 9 | 9.3 | 8.1 | 0.9 | -0.3 |
| NYG | 4 | 4.1 | 5.2 | -1.2 | -0.1 |
| NYJ | 4 | 4.1 | 4.5 | -0.5 | -0.1 |
| PHI | 9 | 9.9 | 7.9 | 1.1 | -0.9 |
| PIT | 9 | 7.0 | 6.4 | 2.6 | 2.0 |
| SEA | 7 | 9.3 | 9.8 | -2.8 | -2.3 |
| SF | 10 | 10.1 | 9.7 | 0.3 | -0.1 |
| TB | 13 | 12.0 | 10.8 | 2.2 | 1.0 |
| TEN | 12 | 10.2 | 9.9 | 2.1 | 1.8 |
| WAS | 7 | 6.0 | 6.8 | 0.2 | 1.0 |
Method
I put together a data set for the 2011 through 2020 regular seasons. Doing some EDA, the following histogram stood out.
I fit an extremely simple logistic regression to this data set: for all home teams: win probability as a function of time average lead. The model summary looked acceptable.
I binned the time average leads into 0.5 point bins, and I plotted both the average winning percentage for each bin. It looked visually like logistic regression would be appropriate, so I fit an extremely simple logistic regression to this data set: for all home teams: win probability as a function of time average lead. For the away perspective, I fit a separate model to the same data set for away games.
The “bins + model fit” plot for the away model is provided below.
The home model summary is provided below.
Call:
glm(formula = win ~ mean_point_diff, family = "binomial", data = results_df %>%
filter(home_away == "home"))
Deviance Residuals:
Min 1Q Median 3Q Max
-2.8364 -0.5750 0.1410 0.5969 2.6841
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.08979 0.05768 1.557 0.12
mean_point_diff 0.30574 0.01248 24.495 <2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 3286.4 on 2399 degrees of freedom
Residual deviance: 1875.6 on 2398 degrees of freedom
AIC: 1879.6
Number of Fisher Scoring iterations: 6