Next Gen Stats powered by AWS is excited to present our latest series of advanced stats for the 2021 NFL season. The Next Gen Stats toolbox is growing. Since partnering with AWS in 2018, the Next Gen Stats team has made significant advancements in deriving predictive models from play-tracking technology. Models like Completion Probability, which tells you how likely it is that a pass will be completed, and Expected Rushing Yards, which predicts how many yards a rusher will gain on a run play, create analytical insights and add context to the game.
This season is no different. Here's a breakdown of some of our new metrics to watch for this season:
Expected Rushing Yards for QBs
In 2020, we debuted Expected Rushing Yards, a metric based on the winning submission to the 2020 Big Data Bowl competition. The goal of the model was to estimate how many rushing yards a ball-carrier will gain from the moment of the handoff. The model produces a full probability distribution of outcomes in terms of yards gained, which allows us to also derive the likelihood the ball-carrier gains a first down or scores a touchdown. However, by building the model around the handoff, we were missing out on an ever-growing aspect of football: the quarterback run.
This offseason, we replicated the neural network modeling architecture built for handoffs and trained it on quarterback run plays using the AWS SageMaker platform with the help of our new quarterback dropback classification model (which we'll discuss later in this article). For plays with a handoff, the relative speed, direction and location of all 22 players is taken at the moment of the handoff. For plays without a handoff, the model uses a snapshot when the quarterback makes clear his intention to run.
As it becomes more common to see quarterbacks making plays with their legs, we can better quantify the impact these quarterback runs have with our expanded expected rushing yards model. Instead of being limited to basic max speed metrics to highlight breakaway scrambles like Lamar Jackson's 48-yard TD in the Ravens' Super Wild Card Weekend win last season, we can open our toolbox to better exemplify what makes Jackson so special.
With tight coverage down the field and a collapsing pocket, Jackson attempted to scramble for a first down. The average quarterback would only be expected to gain 6 yards in the same situation. Not Jackson. Using his 20-plus mph speed to find space in the secondary, Jackson scored on a 48-yard TD run, gaining 42 yards over expectation.
Jackson has gained +1,307 rushing yards over expected since entering the league in 2018, comfortably ahead of not just every quarterback but every player in the NFL over that span (Derrick Henry ranks second at +971 RYOE).
Quarterback Dropback Type
The design of an offensive play is predicated on the immediate movement of the quarterback. Does the signal-caller take a straight drop? Does he roll out of the pocket? Is he forced to scramble? Is it a designed run? We've built a model to classify quarterback dropback type in real time with the help of player-tracking data.
The classification model leverages the combination of a convolutional neural network (CNN) and long short-term memory (LSTM) to predict dropback type using player-tracking data as inputs. In essence, we look at a quarterback's location, speed and direction, powering three separate models to determine:
- Whether the quarterback took the path of a straight drop or designed rollout.
- If the quarterback was scrambling during a pass dropback.
- Whether a quarterback run was a result of a scramble or designed run.
Using dropback classification logic derived entirely from raw player-tracking data, we can create new splits to better understand individual and league-wide passing performance controlling for a quarterback's movement in the pocket.
Here are a few insights from the 2020 season via the results of the new models:
-- Quarterbacks averaged +0.12 expected points added (EPA) on designed rollout dropbacks compared with +0.02 on straight drops.
-- Designed rollouts are run to the right side of the formation (most quarterbacks' strong side) more than twice as often (71%) as designed rollouts to the left (29%).
-- Four out of every five designed rollout pass plays also featured play-action (82%).
-- 2020 pass EPA leaders by dropback type:
Going deeper into the splits, using our new quarterback dropback logic, we can more accurately classify a play-call's intended play type. A play in which the quarterback drops back to pass, scrambles and runs can now be credited as a called pass play. This distinction will allow for more robust analysis of run-pass play-calling tendencies.
Next Gen Stats Big Play Score
The top plays of an NFL season are the ones worth watching again and again. By combining several Next Gen Stats machine-learning models, the Next Gen Stats Big Play Score grades every play on a score from 0 to 100, driven by three primary components:
- Win Probability Effect
- Points Added
- Play Improbability
Win Probability Effect is derived from our win probability model, while Points Added comes from our expected points model. The Play Improbability factor is a more play-type specific, complex combination of completion probability, expected rushing yards, and expected yards after the catch.
The top three plays of the Next Gen Stats era (since 2016) based on the new Big Play Score:
Keep an eye out this season for highlight reels of the biggest plays utilizing this metric.
Expected Fantasy Points
We understand that modern fandom can be driven just as much by one's fantasy team as it is to allegiance to one's favorite NFL team. As such, the Next Gen Stats team is making it priority to deliver actionable insights and metrics to help you win your fantasy league (or even more importantly in some leagues, avoid the dreaded last-place punishment).
In fantasy football, opportunity is king. Snap counts, touches and targets are used as proxies for a player's involvement in the offense. More volume equals more fantasy points, as the old adage goes. But not all opportunities carry the same weight in most fantasy scoring formats. A carry at midfield yields considerably less value than a rush attempt from the opponent's goal line. Expected Fantasy Points, and derivative stats like Fantasy Points Over Expected (FPOE), are two metrics you'll be able to use to make in-season adjustments to help you win it all this season.
So how does the Next Gen Stats Expected Fantasy Points metric work? Our new all-encompassing fantasy metric is calculated using a combination of the outputs of several NGS machine learning models:
Take the difference between a player's actual fantasy points scored and expected fantasy points, and you get the aforementioned FPOE. A high FPOE value indicates a player who has significantly outperformed expectations in the past, but could regress if he is unable to maintain the same level of efficiency. Likewise, a player who is underperforming in expected fantasy points is a candidate to improve so long as the volume and value of the opportunity does not change.
Traditionally, quantifying fantasy opportunity has required multi-faceted analysis of metrics across a series of splits and situations. Expected Fantasy Points distills the value of those opportunities into a single metric that is even more actionable, especially given its scale is comparable across fantasy positions.
The NGS team is not done yet when it comes to unveiling new stats and features for the 2021 campaign. We have more advanced metrics to be announced in the coming days, weeks and months.