ashton dirks
all work
live2026·sports analytics

NFL Picks

NFL betting analytics + a walk-forward prediction model trained on every game since 2019

stack

Next.js · TypeScript · Supabase · Postgres · Python · XGBoost · nflverse

links

The full case study is being written. The product is live — sign-in gates the model's predictions, but the rest of the site is browseable.

What it is

A private analytics workspace for NFL betting. Six seasons of schedules, scores, depth charts, injury reports, futures odds, and play-by-play data — all queryable through a fast Next.js UI. On top of the data sits a prediction model: XGBoost classifier for win probability, a margin regressor, and a Monte Carlo simulator that turns model output into cover probabilities and Strong / Lean / Pass recommendations.

A /model page tracks how the model has actually performed against the market: ATS record, ROI at flat $100 / -110, calibration plot, season-by-season breakdown. Walk-forward backtesting across 2020–2025 keeps the evaluation honest — no peeking at future games.

Why I built it

Public betting tools are either flashy and shallow or accurate but rented. I wanted one stack I owned end-to-end — data ingestion, modeling, evaluation, and a UI I'd actually use during a Sunday slate. Friends-and-family access only; not a betting service.

What's next

A full write-up of the build — feature engineering decisions, what the calibration plot revealed, where the model bleeds money, what I'd do differently — is coming. For now: visit the live product →

all workvisit nfl picks