An interesting article on fivethirtyeight.com reveals a startling and consistent improvement in the efficiency of NFL field goal kickers. Field goal success rates have steadily improved throughout the history of the league with some of the most dramatic upgrades occurring in the 50+ yard range.

It now appears that the New York Times’ “4th Down Bot” may have been criticizing the conservative play-calling of NFL coaches unfairly. As revealed in the article, the field goal accuracy assumptions used by the Bot haven’t been updated since 2004. Many of the fourth and short decisions assessed by the Bot in the “dead zone” and tagged as errors may, in fact, be incorrect. Perhaps these conservative NFL play-callers knew what they were doing on fourth downs after all. Not so fast…

As part of a team that developed the ZEUS play-calling model, we too have often criticized the conservative actions of NFL coaches on fourth downs during the past decade. Fortunately, our model has always incorporated the customized characteristics of individual teams in rushing, passing and kicking (on both offense and defense). These customizations are updated each week of each season and accurately reflect the expected field goal accuracy at the time of the game. The good news for NFL coaches is that field goal accuracy has indeed improved. The bad news is that they are still costing themselves about as much on poor fourth down decision making as they always have.

The majority of cost on fourth downs does not occur in the dead zone but rather the red zone. Taking the “guaranteed” points on a fourth and one inside the twenty yard line is generally very costly. While we still never condone kicking on a fourth and one, with few exceptions, the costs of those decisions can get magnified, in terms of win probability, when you approach the end zone. Green Bay gave up nearly 13% game-winning chance in the recent NFC championship game for this very reason.

In 2013 the average NFL team gave up between 0.6 & 0.7 expected wins on fourth down errors alone, according to our model. While this figure is a slight improvement over 2004, very little has changed in the past ten years.