./docs

Documentation

Technical reference for the F1 prediction platform

01 / overview

What is this?

F1 Intelligence is a Formula 1 race winner prediction platform. It ingests real session data from every grand prix weekend using the FastF1 library, computes a 12-feature weighted score for each driver, and applies a softmax function to produce win probabilities.

Sprint weekends run a separate 8-feature model tuned for the shorter format — heavier weight on grid position and car pace, no tyre degradation or long-run pace signals since there are no pit stops in a sprint.

What data is used?

▸

FastF1 — Lap times, sector splits, tyre data, qualifying times, race results, weather — all from the official F1 timing feed

▸

Historical data — 2018–present with full lap telemetry; 2000–2017 with race/qualifying results only

▸

Static circuit data — Overtake rate and safety car probability seeded once per track — never changes at runtime

▸

No external APIs — FastF1 is the sole data source. Everything else is computed or static

02 / how it works

FastF1

live F1 session data

ETL Jobs

Python / Render

Neon DB

PostgreSQL

Hono API

Cloudflare Workers

Astro UI

Cloudflare Pages

After qualifying

FastF1 pulls Q1/Q2/Q3 lap times, sector splits, and grid positions. FP2 long-run stint data is also ingested for race pace estimation.

Feature computation

compute_features scores each driver across 12 dimensions: car pace, tyre deg, reliability, qualifying delta vs teammate, win rate, luck factor, sector strength, and more.

Prediction

Scores are passed through softmax at temperature T=0.3. Lower temperature = more decisive — small score gaps produce clear probability differences.

After the race

compute_season_stats re-aggregates driver and team season totals. DNF counts, win rates, and car performance scores update for the next prediction.

Sprint weekends

Run an extra ingest + prediction cycle on Friday for the sprint session. The sprint model uses SQ qualifying times and grid position as the dominant signals.

Historical accuracy

Past predictions are stored alongside actual race results so you can see how the model performed over the season.

03 / reference

⬡ architecture.md

Architecture

Monorepo layout, four-layer service topology, and how the pieces connect

→ read

⤷ data-pipeline.md

Data Pipeline

ETL jobs, FastF1 ingestion, cron schedule, and race status flow

→ read

⊞ database-schema.md