This post is adapted from the project README and focuses on engineering decisions, modeling assumptions, and system design tradeoffs, not football takes.

NFL Impact Forecasting (V1)

Why I built this

I love football — and I love arguing about it.

But most NFL debates collapse immediately because people skip the hardest part:

What exactly are we measuring, and what assumptions are we making?

Player impact is inherently subjective. Football is interdependent, contextual, and noisy. Pretending otherwise doesn’t make models better — it just hides the assumptions.

This project is my attempt to do the opposite:

  • make assumptions explicit
  • encode them into the system
  • version them
  • test them
  • and leave room for disagreement

The goal is not to claim truth.
The goal is to build a system where truth can evolve without breaking everything downstream.

The problem in one paragraph

Using NFL play-by-play data, we want to attribute value to individual players in a way that is:

  • consistent
  • reproducible
  • defensible
  • extensible

Given a play with an Expected Points Added (EPA) value, the system must decide who gets credit, how much, and why — while supporting reruns, metric revisions, and multi-year analysis.

Correctness and clarity matter more than sophistication.

Constraints first, architecture second

Before writing any attribution logic, I locked in a few non-negotiable constraints:

  • Assumptions must be declared, not implied
  • Raw data must be immutable
  • Attribution must be deterministic
  • Metrics must be versioned
  • Re-runs must be idempotent

If you can’t rerun a season and get the same result, the system isn’t trustworthy — no matter how clever the model is.

High-level architecture

Raw NFL PBP (nflverse)
↓
Parquet Artifacts (immutable)
↓
Canonical Plays (SQLite)
↓
Play-Level Impact Attribution (versioned)
↓
Season Aggregations / Reporting

The pipeline is intentionally linear and layered.

Each stage produces new derived data rather than mutating existing records, which keeps recomputation explicit and auditable.

Core design principles

1. Immutable raw data

Play-by-play data is ingested once and stored as immutable Parquet artifacts.

This provides:

  • a permanent audit trail
  • safe historical backfills
  • insulation from upstream data changes

If attribution logic changes, the past is recomputed — not patched.

2. Canonical plays as the foundation

Before attribution, all plays are normalized into a single canonical table.

This includes:

  • play type normalization (PASS / RUSH / OTHER)
  • filtering invalid or non-actionable plays
  • consistent identifiers for games, plays, and players

Every downstream metric depends on this table.
If something is wrong here, everything downstream is wrong — so this layer is deliberately boring and conservative.

3. Attribution is assumption-driven, not “correct”

V1 uses a deliberately simple attribution rule:

Play TypeAttribution
PASS70% QB, 30% Receiver
RUSH100% Rusher
OTHERIgnored

This is stored explicitly as:

impact_version = "epa_v1_qb70_rec30"

This is not a claim of truth.

It is a claim of:

  • consistency
  • testability
  • debuggability

Future versions can change the weights, add offensive line credit, or introduce defensive attribution — without overwriting history.

4. Versioned metrics are non-negotiable

Every impact row is keyed by:

(game_id, play_id, player_id, impact_version)

This guarantees:

  • no silent overwrites
  • safe re-runs
  • side-by-side metric comparisons

If V2 disagrees with V1, that’s a feature — not a bug.

Example schema: play_impact

Each row represents a single attribution decision:

ColumnDescription
game_idNFL game identifier
play_idPlay identifier
player_idPlayer credited
roleQB / WR / RB (V1)
impact_metricEPA
impact_valueAttributed EPA
impact_versionAttribution ruleset
batch_idLineage back to raw data
computed_atComputation timestamp

Nothing is aggregated prematurely.
Season-level insights are derived later, intentionally.

Testing philosophy

Tests focus on invariants, not exact values:

  • total impact per play equals EPA
  • missing players don’t break attribution
  • re-running attribution doesn’t duplicate rows
  • metric versions don’t overwrite each other

If those invariants hold, the system is doing its job — even if the assumptions evolve.

Why this is position-agnostic (eventually)

V1 still labels QB / WR / RB roles explicitly.

But the long-term goal is position-agnostic impact:

  • value attributed based on involvement, not title
  • aggregation driven by contribution, not roster slot

This is harder — and more honest — than position-specific metrics.

What I intentionally didn’t build

This project deliberately excludes:

  • predictive models
  • fantasy scoring
  • betting logic
  • dashboards or visualizations

Those are downstream consumers.

This project exists to answer a more fundamental question:

Can we build an analytics system where assumptions are first-class citizens?

V1 Example Results

As of Week 17 of the 2025 season, the V1 pipeline produces the following results.

Top 30 players by total impact (EPA attribution):
rank player_name          role         total_epa    plays   epa/play
   1 Drake Maye           QB           +107.09    540   +0.198
   2 Jordan Love          QB            +83.66    469   +0.178
   3 Matthew Stafford     QB            +83.07    579   +0.143
   4 Dak Prescott         QB            +75.74    627   +0.121
   5 Josh Allen           QB            +72.87    548   +0.133
   6 Jared Goff           QB            +67.59    573   +0.118
   7 Brock Purdy          QB            +61.93    279   +0.222
   8 Daniel Jones         QB            +51.96    423   +0.123
   9 Patrick Mahomes      QB            +51.19    540   +0.095
  10 Sam Darnold          QB            +41.06    480   +0.086
  11 Bo Nix               QB            +40.91    642   +0.064
  12 Puka Nacua           WR            +39.10    163   +0.240
  13 Joe Burrow           QB            +33.83    240   +0.141
  14 C.J. Stroud          QB            +32.52    433   +0.075
  15 Mac Jones            QB            +32.16    318   +0.101
  16 George Pickens       WR            +27.76    135   +0.206
  17 Trevor Lawrence      QB            +27.24    599   +0.045
  18 Jalen Hurts          QB            +27.19    538   +0.051
  19 Malik Willis         RUSHER        +26.93     45   +0.598
  20 Caleb Williams       QB            +26.35    579   +0.046
  21 Jonathan Taylor      RUSHER        +26.19    363   +0.072
  22 Jaxon Smith-Njigba   WR            +25.49    162   +0.157
  23 Trey McBride         WR            +21.35    162   +0.132
  24 Stefon Diggs         WR            +20.39     99   +0.206
  25 AJ Barner            WR            +15.83     75   +0.211
  26 DeVonta Smith        WR            +15.81    110   +0.144
  27 Amon-Ra St. Brown    WR            +15.46    160   +0.097
  28 D'Andre Swift        WR            +14.35    259   +0.055
  29 Nico Collins         WR            +13.79    123   +0.112
  30 Jameson Williams     RUSHER        +13.74    100   +0.137

Interpreting These Results

These results strongly favor quarterbacks. This is expected behavior, not a flaw.

Under V1 assumptions:

  • Pass plays attribute 70% of EPA to the QB
  • Rushing EPA is attributed solely to the rusher
  • Offensive line, defensive pressure, coverage quality, and situational leverage are not yet modeled

Because quarterbacks touch the ball on nearly every offensive play, the system naturally concentrates impact at the QB position.

This outcome serves as a validation of the attribution rules, not a claim about true player value.

Why This Is Important

The purpose of V1 is not to produce “correct” rankings — it is to:

  • Prove the pipeline is internally consistent
  • Verify that impact attribution sums correctly
  • Expose systemic bias introduced by assumptions
  • Create a baseline for future iterations

Seeing quarterbacks dominate the rankings confirms that the system is behaving deterministically and transparently under its declared rules.

What This Unlocks Next

These results directly motivate future work:

  • Offensive line attribution on pass and rush plays
  • Position-agnostic impact normalization
  • Defensive player involvement
  • Snap-weighted and context-adjusted impact
  • Multi-year trajectory modeling

Crucially, all of these improvements can be introduced as new impact versions without invalidating or overwriting V1.

That is the core design goal.

What comes next

Planned evolutions include:

  • offensive line attribution
  • defensive impact modeling
  • snap-weighted contributions
  • multi-season forecasting (5–10 year horizon)
  • capital-aware team projections

Each extension builds on the same foundation rather than replacing it.

Closing thoughts

Football analytics debates rarely fail because of math.

They fail because assumptions are hidden, pipelines are brittle, and history gets quietly rewritten.

This project is my attempt to build the opposite:

  • explicit assumptions
  • deterministic pipelines
  • versioned disagreement

If people argue with the results — good.
At least now we’re arguing about something real.

Source code:
https://github.com/tkalp/nfl-analytics