nfelo DCM

nfelo DCM is an abstraction layer for loading and caching NFL related CSVs stored on the web. DCM stands for Dataframe-CSV Mapping. The goal of the DCM is to get pandas dataframes of fresh data loaded in a way that balances simplicity, efficiency, and performance.

import nfelodcm

## Load 2 dataframes
db = nfelodcm.load(['pbp', 'games'])

## access the PBP dataframe
db['pbp']

Maps

Maps are config files that tell the DCM where data CSVs are located, how they should be retrieved, and what fields to pull. Each CSV has its own config in Maps/{table}.json, where parameters can be set for things like freshness SLAs, CSV parsing engines, iteration strategy, and assignments (mutations).

An important characteristic of these maps is that all fields must be 1) specified in the map and 2) typed. Fields not listed in the map will not be loaded. Untyped fields will throw an error.

Here is a sample config:

{
  "name": "games",
  "description": "nflgamedata games",
  "download_url": "https://raw.githubusercontent.com/nflverse/nfldata/master/data/games.csv",
  "compression": null,
  "engine": "c",
  "freshness": {
    "type": "gh_commit",
    "gh_api_endpoint": "https://api.github.com/repos/nflverse/nfldata/commits",
    "gh_release_tag": null,
    "sla_seconds": 500
  },
  "iter": {
    "type": null,
    "start": null
  },
  "assignments": [
    "fastr_team_id_repl",
    "score_clean"
  ],
  "map": {
    "game_id": "object",
    "season": "int32",
    "week": "int32",
    ...
  }
}

Config Fields

Freshness

The DCM uses a two-tier freshness strategy:

1. SLA Check: If the last freshness check was within sla_seconds, skip the remote check entirely
2. Remote Check: Query GitHub API to compare remote timestamps against local state

For gh_release tables, freshness is determined by the updated_at timestamp of release assets. For gh_commit tables, freshness is based on the latest commit date.

Per-File Freshness (v0.2.1+)

For season-iterated tables (pbp, rosters, player_stats, etc.), the DCM tracks freshness per-season. When an update is needed, only stale seasons are re-downloaded - cached seasons are read from Data/Parts/{table}/. This significantly reduces bandwidth for incremental updates.

Data Storage

Data/
  games.csv
  pbp.csv                # Combined table CSV
  Parts/
    pbp/
      1999.csv           # Per-season cache (iter tables only)
      2000.csv
      ...
State/
  Tables/
    games.json           # Per-table state (last_local_update, last_freshness_check)
    pbp.json
  Parts/
    pbp.json             # Per-season timestamps (iter tables only)
  Global/
    season_state.json    # Current NFL season state

Assignments

Assignments are DataFrame transformations applied after data is pulled. They take a DataFrame as input and return a mutated DataFrame. Assignments are defined in Engine/Assignments/ and referenced by name in config files.

Common assignments include:

• fastr_team_id_repl - Standardize team abbreviations
• score_clean - Fix known data errors in game scores
• penalty_formatting - Parse penalty descriptions

GitHub Token (Optional)

To increase GitHub API rate limits from 60/hr to 5,000/hr, create a .env file in your working directory:

GITHUB_TOKEN=ghp_your_token_here

The token is used for freshness checks only, not for downloading CSVs. A token is only relevant/needed when pulling many tables with extremely fast processing times. In most use cases, the default rate limit is sufficient.

API

import nfelodcm

# Load tables
db = nfelodcm.load(['pbp', 'games'])

# Get a single DataFrame
df = nfelodcm.get_df('pbp')

# Get table config
config = nfelodcm.get_map('games')

# List available tables
tables = nfelodcm.list_tables()

# Get current season state
season, week = nfelodcm.get_season_state('last_full_week')

Further Detailed Documentation