Overview

Every NFL play is driven by context. Down, distance, field position, score, and time remaining all influence whether an offense chooses to pass, run, punt, or attempt a field goal. Some teams follow clear situational tendencies, while others are harder to anticipate.

This project uses 2025 NFL play by play data to build a multinomial regression model that predicts play type from game context. The goal is not just to classify plays, but to measure how predictable each team is using an entropy based score derived from model probabilities.

Results

Rank Team Plays Predictability Accuracy
1 MIN 112 0.638 0.786
2 TEN 106 0.636 0.726
3 CLE 136 0.636 0.772
4 DAL 147 0.635 0.782
5 NE 128 0.635 0.820
6 PIT 116 0.634 0.741
7 CIN 124 0.633 0.782
8 HOU 136 0.633 0.846
9 TB 137 0.627 0.737
10 DEN 154 0.625 0.812
11 LAC 160 0.625 0.850
12 CHI 137 0.624 0.730
13 ARI 119 0.622 0.773
14 JAX 151 0.622 0.762
15 LV 110 0.613 0.727
16 MIA 117 0.613 0.692
17 LA 146 0.607 0.685
18 PHI 113 0.606 0.788
19 NO 121 0.606 0.760
20 NYJ 121 0.605 0.744
21 SEA 100 0.603 0.730
22 BUF 116 0.602 0.750
23 BAL 118 0.602 0.686
24 SF 125 0.598 0.752
25 GB 128 0.596 0.641
26 CAR 141 0.594 0.738
27 NYG 141 0.591 0.730
28 DET 122 0.589 0.730
29 IND 119 0.584 0.782
30 KC 145 0.583 0.710
31 ATL 118 0.582 0.686
32 WAS 134 0.576 0.769

Predictability by Team

Interpretation

Teams near the top are easier for the model to read because their decisions align more consistently with game situation. Teams lower on the chart show more uncertainty in predicted probabilities, indicating less predictable behavior.

This does not imply one approach is better. High predictability reflects structure, while lower predictability reflects variation. The value is in measuring how clearly a team’s tendencies show up in the data.

Conclusion

This project moves beyond simple classification by using predicted probabilities to evaluate team behavior. Instead of only asking whether the model can predict the next play, it asks how readable that decision is in the first place. That creates a more meaningful view of offensive tendencies through the lens of data and probability.