About the project

Why this platform exists

As a fan, this project came from the same questions fans ask all season: which teams are actually better, why certain matchups swing the way they do, and how differently the playoff picture might have looked if a few moments changed. The goal is to turn those questions into something you can explore with real context through rankings, team-by-team history, and matchup tools that make the season easier to break down.

Overview

System intent

This NFL Elo dashboard was created by me, a student at West Chester University of Pennsylvania, for the purpose of practicing and demonstrating full-stack development. The project loads season data, stores and computes Elo ratings in the backend, and generates a website that keeps charts and rankings easy to browse.

The goal is not just to display records, but to show the full path from raw game data to analytics output and a polished presentation.

Architecture

Tech stack

Data

Python 3.11 SQLite httpx requests

API

FastAPI Uvicorn Pydantic

Build

Jinja2 Python build scripts Docker

Frontend

Vanilla JS (ES6+) Custom CSS Google Fonts MathJax

Statistical models

Full methodology

Win Probability

Expected Score (Elo)

The probability that Team A beats Team B, derived from the standard Elo formula. A 200-point Elo gap gives the stronger team roughly 76% win probability.

\[ E_A = \frac{1}{1 + 10^{\frac{R_B - R_A}{400}}} \]

\(R_A, R_B\): Elo ratings of teams A and B
400: scale constant (standard Elo)

Home-Field Advantage

HFA Adjustment

Before computing win probability, the home team receives a temporary Elo point boost that is not stored in the rating. Two calibrated values are used: +55 pts for regular-season Elo computation (Glickman & Jones 1999), and +65 pts in the playoff bracket simulator (FiveThirtyEight NFL playoff calibration). The Super Bowl uses no HFA — it is played at a neutral site.

\[ E_{home} = \frac{1}{1 + 10^{\frac{R_{away} - (R_{home} + \delta)}{400}}} \]

\(\delta = 55\): regular-season Elo rating computation
\(\delta = 65\): playoff bracket simulation (home = higher seed)
\(\delta = 0\): Super Bowl (neutral site)
Source: Glickman & Jones (1999), FiveThirtyEight NFL Elo

Elo Update Rule

Rating Adjustment

After each game, both teams' ratings shift by the K-factor (25) scaled by the surprise of the result and a margin-of-victory multiplier. Blowout wins are worth more than narrow ones.

\[ \Delta_{home} = K \cdot M \cdot (S_{home} - E_{home}) \] \[ S_{home} \in \{1,\; 0.5,\; 0\}, \quad K = 25 \]

\(S\): actual result (win=1, draw=0.5, loss=0)
\(M\): margin-of-victory multiplier (see below)

Margin of Victory

MOV Multiplier

Straight Elo treats a 1-point win the same as a 40-point blowout. The FiveThirtyEight MOV multiplier scales the update by point differential while correcting for the fact that dominant teams already have high Elo ratings.

\[ M = \frac{\ln(|PD| + 1) \cdot 2.2}{0.001 \cdot (\text{ELO}_W - \text{ELO}_L) + 2.2} \]

\(PD\): absolute point differential
\(\text{ELO}_W, \text{ELO}_L\): pre-game ratings of winner and loser
Denominator shrinks the multiplier when the favored team wins big (autocorrelation correction)

Score Prediction

OLS Margin Model

After computing all season Elo ratings, an ordinary least-squares regression is fit over every game, predicting point differential from pre-game Elo difference. This gives interpretable, calibrated margin estimates for any matchup.

\[ \hat{m} = \alpha + \beta \cdot (\text{ELO}_A - \text{ELO}_B) \]

\(\alpha\): intercept (home-field bias in margins)
\(\beta\): slope ≈ points per 100 Elo difference
Fit on all 2024 regular-season games (n ≈ 272)

Game Simulation

Score Simulation

To simulate a final score, the expected margin is used as the mean of a Normal distribution (σ = 13.45, the historical NFL game margin std dev). Total points are drawn from a separate Normal (μ = 43.5, σ = 10). Random draws use the Box-Muller transform for true Normal sampling.

\[ \text{margin} \sim \mathcal{N}(\hat{m},\; \sigma_m = 13.45) \] \[ \text{total} \sim \mathcal{N}(43.5,\; \sigma_t = 10) \] \[ Z = \sqrt{-2\ln u_1}\,\cos(2\pi u_2), \quad u_1, u_2 \sim U(0,1) \]

\(\sigma_m = 13.45\): historical NFL margin standard deviation
43.5 pts total: 2024 season average combined score
Box-Muller: transforms two uniform random numbers into one Normal variate

Model Validation

Monte Carlo Confirmation

To validate the playoff simulator, 100,000 full bracket simulations were run using the 2024 seedings and Week 18 Elo ratings. Super Bowl win rates by seed were compared against historical NFL outcomes (1975–present). The simulator matches historical benchmarks closely — confirming the Elo engine, margin model, and HFA are well-calibrated.

Seed	Simulated win %	Historical range	Result
1	47.0%	~35–40%	High — 2024 #1 seeds (Detroit 1711, KC 1631) were historically strong
2	25.3%	~20–25%	✓ Match
3	14.0%	~10–15%	✓ Match
4	3.0%	~8–10%	Low — Houston Texans (1511) were a weak 4-seed
5	3.6%	~5–8%	✓ Match
6	4.2%	~3–5%	✓ Match
7	2.9%	~1–2%	Slightly high — Denver Broncos (1616) were a strong 7-seed

N = 100,000 full bracket simulations
Deviations at seeds 1, 4, and 7 reflect this specific 2024 bracket, not model error — higher/lower Elos than typical for those seed positions
Adding HFA (+65) moved seeds 6 and 7 from significantly over-estimated into their historical bands

System architecture

How it works

Layer 1 · Ingestion

Data Store

Fetch game data from ESPN API
Normalize scores & metadata
Persist games to SQLite
Serve via /games endpoint

FastAPISQLitePython

Layer 2 · Analytics

Elo Pipeline

Load all played games
Compute weekly Elo ratings
Apply MOV multiplier & HFA
Fit OLS margin model
Build win-probability matrix
Export elo_2024.json

PythonEloOLSJSON

Layer 3 · Delivery

Static Site

Read Elo JSON artifact
Render Jinja2 templates
Team & leaderboard pages
Matchup simulator
Playoff bracket tool
Output to ui/dist/

Jinja2HTMLCSSJS