Sport Analytics: American Football

A more quantitative approach to football

Julius Grocholski: What was the pattern behind his playcalling system?

24. December 2017, 2.46 p.m.

Kozly J11 have enjoyed an unprecedented and somehow unexpected success this season. For years the team has been much more effective defensively, sometimes scoring more defensive than offensive TDs. The 2017J season marked a successful transition to more balanced football. The transition would not have happened without the QB prospect Guenther Krotsch (class of 2018, 6’6’’, 2016 highlights available here ) and the OC Julius Grocholski. Let’s take a closer look at Grocholski’s playcalling system.

A Wisconsin-native, Julius Grocholski is not new to Polish football. I had had the opportunity to work with him during my tenure with Kozly in 2013, when he’d served as the O-Line coach. Earlier this year I welcomed his promotion to the OC of the senior team. Even more important was to put Grocholski in charge of the junior team offence. It’s always a good idea to put an ex-offensive lineman in charge of the junior team offence: don’t spoil them with air-raid offence, educate them and propagate a smash-mouth-football completed with quick passing game.

The first dashboard breaks down the frequency of plays called. The orange-blue slices visualise the frequencies of running (65) and passing plays (111). Individual percentages of running plays called are depicted in the body of the charts. The season opener against the Angels showed a pretty balanced offence, but then a zigzag pattern was put into appearance. The Archers game started with a stunning 92-yard run for a touchdown, and yet Kozly utilised their running backs only every third play. The trend reversed in the semi-final match against the Lowlanders, when Kozly ran the football in nearly half of the plays. The lowest running percentage was recorded in the most important game of the season. That's surprising, isn't it?

If you still wonder why Kozly J failed to win the championship, the black bar in the second dashboard offers an explanation. Incidentally, it also reveals the reason why Grocholski had selected a running play only 11 times (21%) during that game. Kozly J were unable to establish the running game, finishing with a total rushing gain of 1 yard. Is the OC to be blamed for that? Hardly. Imagine what Kozly J could have achieved with a decent practice attendance. But that’s for something completely different.

The third dashboard offers a more detailed analysis of yardage gained. Blue and white bubbles depict passing and running games, respectively. Their size is determined by the yards gained in an individual play. What’s very interesting is the time-series properties of the passing and running series. Let’s inspect them separately. The blue passing series are driven by level shifts. Take for instance the championship game – 103 total passing yards place the gain slightly above the average, but if you hover over the graph’s surface you’ll find out that the largest bubble contributed greatly to the total yardage. In fact, two plays gave Kozly J 59 yards. Play number 14 shifted the passing gain to a 50-yard level, which hasn’t been passed for the next 19 plays! You can find similar pattern in other games, too. I’m not saying it’s unique to the Kozly J offence. I’m rather stating the obvious, pointing out that dropped and incomplete passes as well as sacks ruin a smooth, upward-moving trend. But the amount of zero-yard (or even negative) gains is overwhelmingly large.

With the exception of the unfortunate championship game, the rushing yardage has been gained in a much more consistent manner. Offensive co-ordinators love to observe a linearly growing rushing yardage. That’s the essence of smash-mouth football. No matter what the DC calls, whether he blitzes, changes fronts, puts additional defenders to the box, the offence consistently gains yards by running the football. The Lowlanders game, particularly plays between 25 to 44, illustrate this point best.

A few months ago a new supernova exploded in the football internet universe. There’s this young gentleman based somewhere in the States, who constructed a forecasting model able to perfectly predict each and every play in the NFL. Kudos to him, that’s what they say in the States, right? I too have constructed predictive models. Some of them were of great help in 2011, when The Crew won the championship. Let’s feed the data to the model* and check how predictable Julius Grocholski's play selection has been.

I’ve regressed Grocholski’s choice (run or pass) against various sets of explanatory variables, but only 3 of them have turned out to be significant. Not surprisingly, the variables a DC should have considered whilst scouting Julius were down, distance, and number of offensive players in the box. Yard line, hashmark, coverage shown prior to snap (i.e. single-high or two-high safeties), number of defenders in the box or number of blitzing defenders (yes, it’s possible to observe it prior to the snap, as the juniors can hardly contain their emotions, especially when they are about to blitz) – all these variables had had insignificant bearings on Grocholski’s decisions. Let’s make it crystal clear. Game plan preparations consist of scouting safeties, but if a DC sticks to Cover 3 Sky all day long, the number of high-safeties becomes irrelevant, as the OC calls plays to his advantage, utilising the weaknesses of Cover 3 Sky. The data, however, won’t show any statistically significant relationship between the coverage and selected plays.

The estimated probabilities are presented as red (incorrect predictions) and green (correct predictions) dots. You can hover over the graph’s surface to get more information on a selected play. For instance, for the 7th play of the Archers game, the model predicted a pass with 78.3% probability. Fill the shoes of the Archers’ DC for a moment. You see 2nd. down, the distance is 10, and the offence is aligned in trips formation (check the hover information: one player in the box, not including the QB). But Julius Grocholski called a run. A run was less likely but not impossible, as the model predicted 20% chances for a run. More importantly, wouldn’t you expect the same? Obviously, the model is not perfect. Nothing is perfectly certain if probability is considered. The model correctly predicted 91% of passes and 48% of runs. These numbers again prove the importance of running game. Scouting an OC leaves a larger margin for error if the offence runs the football effectively.

Julius Grocholski is less recognised as a successful football coach than he deserves to be. Whatever side of the pond he will be on, I wish him best of luck. I hope we’ll hear a great deal of his coaching skills in the nearest future. Better still, he’ll become a millionaire and establish a football league of his own.

* The estimated model was a binary-choice logit. The McFadden R^2 was equal to 0.202, log-likelihood to -92.458, and the BIC to 200.427.

For more information on 2017 PLFA Junior season visit PLFA website

Play Selection Frequencies

Yardage Gained: Rushing vs. Passing Yards

Play Selection and Yardage Gained

Play Selection Probability