The beginnings of a framework to run ML models on large amounts of fantasy football data, for the purpose of evaluating player scoring potential.

Tested with Python 3.5 and up. The recommended way to run this is in your own script using a Python virtual environment:

pip install virtualenv
virtualenv --python=/path/to/python/interpreter env
source env/bin/activate
pip install -r requirements.txt
{script call}

NOTE: Project is a work-in-progress and not guaranteed to run smoothly at this point

See ff_data/data.py.

The framework currently supports raw scraping from Pro Football Reference, adhering to all PFR guidelines. It includes player profile data as well as season logs and individual game logs going back an arbitrary number of years. The scraper focuses only on fantasy-relevant players for each given year to boost efficiency.

More data sources and features coming soon.

See ff_data/models.py.

There are basic sklearn-centric implementations for Decision Tree Regression as well as XGBoost gradient-boosted Decision Tree Regression ensembles. XGBoost hyperparamater optimization can be run via GridSearch with repeated k-fold cross-validation. The framework also supports train-test splits.

More models coming soon.

• The NFL is one of the most unpredictable sports that we know. The FF community would be well served to start thinking about player outcomes in terms of probability distributions and confidence intervals, rather than static projections or linear positional rankings.
• After experimenting with several kinds of models, it seems decision tree is the most effective base learner when looking at season-level data (e.g. projecting player scoring potential over the course of a full season). Naturally, random forest and gradient-boosted models are much stronger than individual trees.
• Tree structures allow for grouping players together based on feature values. This can be valuable for making pure regressive scoring predictions, but also for providing other forms of statistical context to a player. For example, we can use a gradient-boosted tree sequence to generate a large number of trees. For each node/leaf in each tree, a player is grouped with other similar players - in the aggregate, these groupings can give us input into a player's range of scoring outcomes (e.g. worst-case scenario or best-case scenario).
• Decision trees also might be useful for creating "factor" models that measure statistically significant feature values. Tree splits inherently identify meaningful feature values. This can help us answer questions like "What is the optimal height/weight for a running back?" There are conventional answers to this question, but tree structures give us the tools to quantify things at a deeper level that can interact with other features.
• Feature engineering is paramount for projecting real-world NFL football outcomes. It is incredibly difficult to create an accurate learning model without some very creative features. This reality is difficult to overcome for 2 reasons. First, it requires deep understanding of the game as well as how the game has changed over time. Second, it creates a burden to unearth data that is often non-public, expensive, fragmented, unstructured, or otherwise hard to source, clean, or model. We simply cannot acquire all the data we need by scraping an online source like PFR (although PFR gives us a very good starting point).
• On-field NFL player performance can typically be assessed in 3 dimensions:
  • Individual player profile - attributes specific to the player, like 40-yard dash time
  • Team roster - attributes related to the team context the player plays in, like offensive line quality
  • Coaching/system - attributes related to the football system the player plays in, like offensive pass %
  • It is challening to isolate these dimensions - many attributes operate on multiple levels. For example, a player's draft capital may indicate a player's individual profile, but it also represents a team's investment in the player and relates to the team roster context. Each aspect can impact the model in different ways - consider an instance in which a player moves on to a new team after being drafted elsewhere.
• There is some promise in predicting "opportunity" and/or underlying statistics instead of predicting direct fantasy scoring. Rather than modeling total fantasy points for a season, we can model something like opportunities per game, points per opportunity, or expected number of games played. Granularized predictions may be able to produce more resilient models.
• Injuries are not easy to account for because of the vast spectrum of injury type and severity, coupled with the lack of standardized data. Individual players can also respond differently to the same injury. But injuries can have massive impact on a player's fantasy value from season to season, so ignoring them is not a good option.
  • Maybe it's possible to engineer a proxy feature like "missed game probability" that can be incorporated into the model.
• Rookie players are particularly hard to model because of the lack of NFL statistical history. While we can leave rookie stats as undefined and split them with XGBoost, we probably need a separate model entirely with rookie-specific features.