Introducing the nfl-dbt repo

Did you know you can now sign up for weekly-ish updates to our blog via email?

In my last post, “Bayesian Modeling of NFL Football Fourth Down Attempts with PyMC3”, I referenced the nflscrapR-data repo where I had sourced the relevant play-by-play for NFL games from 2009 through 2019. After initially loading this data to BigQuery and querying the raw tables, I quickly realized that this calls for a clean and repeatable data loading and transformation process if this dataset is to be used for model building or teaching purposes.

So, I’m happy to release v1 of the nfl-dbt repo!

This repo contains dbt models to transform NFL Play-by-Play (pbp) data sourced from nflscrapR-data into analytical models.

The repo currently assumes that raw data is loaded to and transformed on a local or remote PostgreSQL instance. The load script outlined below and the dbt models could be easily modified to work with other databases supported by dbt such as Snowflake, BigQuery or Redshift. PRs welcome!

Update Frequency

The nflscrapR-data repo is updated with some regularity, but since this is a voluntary and free resource (thanks to Ron Yurko!), we can’t rely on play data being updated weekly. So, this dataset and the analytical models are best used for teaching and model building purposes, and perhaps less so for weekly decisions on sports bets etc.

Models

• dates: list of all game dates by season and season type (PRE, REG, POST)
• games: game id, dates, teams and final scores by game
• players: player id and name for every player
• plays: combines play data from all available seasons (2009 to 2019) into a single table for easier analysis
• teams: team code and consolidated code, in case of team moves of renames
• teams_players: team rosters by season, showing player and (primary) position for the season

Notes

• a few missing player_id values in the players and teams_player models have been (at least attempted to be) fixed
• any duplicate plays (likely a result of the scraping process) are removed from plays

Data Load

The repo assumes that the raw scraped data has been loaded to a PostgreSQL database, with one raw file corresponding to a single table in a database called raw.

The included Python script extract_load.py is intended to do the following:

• Clone and/or locally refresh the nflscrapR-data repo
• Create empty tables in a local Postgres instance
• Load raw data files to Postgres using a dbt run-operation to load each file using Postgres’ copy command

The extract_load.py file can be easily configured to work with a remote Postgres server, e.g. hosted on AWS RDS. With a little bit of extra work this can also be modified to work with Snowflake, BigQuery or Redshift.

Future Work

The following items would make great natural extensions and improvements to the repo:

• Update extract_load.py to use connection info from ~/.dbt/profiles.yml
• Add support for Snowflake, BigQuery and Redshift
• Add report models to more easily enable analytical models:
  • Player stats
  • Game stats
  • Season stats
• Remove dependency on nflscrapR-data and include R scripts to scrape the data independently

As always, let me know what you think and I’m looking forward to hearing what you do with the data!