MAX 33 · F1 2026 Oracle

A machine-learning CLI that predicts Max Verstappen's 2026 Formula 1 season — race by race.

VER5TAPPEN TRUTHER · in my books he is either #1 or #33 · i do not acknowledge max3

---

What it does

• Pulls race results (2022–2026) for Max and top competitors from the Jolpica/Ergast API
• Trains two Random Forest models on historical race data (position predictor + win probability classifier)
• Predicts finishing position + win/podium probability for every remaining 2026 race
• Runs 10,000 Monte Carlo simulations to project Max's final championship points range
• Auto-refreshes the model when new race results appear
• Correctly handles DNFs — a "Retired" status never counts as a classified finish

---

Quick start

# 1. Clone
git clone https://github.com/yourusername/max33-oracle
cd max33-oracle

# 2. Install deps
pip install -r requirements.txt

# 3. Run
python f1_dashboard.py

# Force a full data refresh (clears cache)
python f1_dashboard.py --force-refresh

First load takes 5–15 minutes on a cold cache (fetching years of race data from the API + rate-limit backoffs). After that, cached runs complete in seconds.

---

File structure

max33-oracle/
│
├── f1_dashboard.py          # CLI entrypoint — argparse, output formatting, state
├── f1_api.py                # All API fetching, caching, and data pipeline logic
├── f1_features.py           # Feature engineering — merges DataFrames, computes metrics
├── f1_models.py             # Random Forest training, predictions, Monte Carlo sim
├── f1_weather_competitors.py# Historical race date lookup table for weather fetching
│
├── requirements.txt         # Python dependencies
└── f1_cache/                # Auto-created — cached API responses + trained model pickle
    ├── *.json               # Cached API responses (race results, lap times, quali, etc.)
    └── model_*.pkl          # Pickled trained RandomForest models

---

How the predictions work

1. Data collection (f1_api.py)

Fetches the following from the Jolpica/Ergast F1 API (free, no API key):

• Race results (2022–2026): finishing position, grid, status, points, lap count
• Qualifying times: Q1/Q2/Q3 times, gap to pole position
• Lap-by-lap times: average/best lap, pace drop-off (tyre degradation proxy), position recovery
• Rival lap comparisons: Max's lap times vs top 5 competitors per race
• Pit stop data: number of stops, average stop duration
• Sprint results: sprint grid/finish for sprint weekend races
• Driver standings: championship points history (2023–2026)
• Constructor standings: Red Bull's car performance ranking per season
• Weather data: historical race-day conditions from Open-Meteo Archive API (free, no API key)

All responses are cached locally to f1_cache/ with smart TTLs:

• Race results/standings → 15 minutes (refreshes frequently during live season)
• Historical/weather data → 30 days (never changes)
• Corrupt or empty cache files are automatically deleted and re-fetched

2. Feature engineering (f1_features.py)

Transforms raw API data into model-ready features:

Feature	Description
grid	Starting grid position
quali_pos	Qualifying position
roll_win_5	Rolling win rate (last 5 races, exponentially weighted)
roll_pts_5	Rolling points average
roll_podium_5	Rolling podium rate
roll_pos_5	Rolling average finish position
roll_gap_5	Rolling gap to pole average
circ_win_rate	Max's historic win rate at this circuit
circ_avg_gap_pole	Max's average quali gap at this circuit
circ_mech_dnf_rate	Mechanical DNF probability at this circuit
circ_crash_rate	Crash/collision DNF probability
circ_avg_stops	Average pit stop count at this circuit
car_perf	Red Bull's constructor standing (normalised)
car_trend	Compound lap pace trend vs 2024/2025
champ_gap	Points gap to championship leader
is_street	Whether it's a street circuit
is_high_speed	High-speed track flag
is_high_df	High downforce track flag
overtake_hard	Low overtaking opportunity flag
wx_temp_c	Race-day temperature (°C)
wx_rain_mm	Race-day rainfall (mm)
wx_wind_kmh	Race-day wind speed (km/h)

Recency weighting: 2026 races → 4.0×, 2025 → 2.5×, 2024 → 1.5×, 2022–2023 → 1.0×

3. Model training (f1_models.py)

Two separate scikit-learn Random Forest models:

• Position predictor (regressor): predicts finishing position (1–20)
• Win predictor (classifier): predicts probability of winning

Both use StandardScaler in a Pipeline. Models are pickled to f1_cache/ and only retrained when new 2026 race results appear (detected via a data hash).

4. Prediction output

For each remaining 2026 race:

• Predicted finishing position
• Win probability %
• Podium probability %
• Historical win rate at that circuit

DNFs are displayed as DNF (not a classified position) in completed races.

5. Championship simulation

Monte Carlo approach (10,000 iterations):

• For each remaining race, randomly sample a finish outcome using the model's predicted probabilities
• Sum the championship points across all remaining races
• Report: mean projection, 10th percentile (worst-case), 50th percentile (median), 90th percentile (best-case)

---

Requirements

pandas>=2.0.0
numpy>=1.24.0
requests>=2.31.0
scikit-learn>=1.3.0

Python 3.9+ recommended.

---

Notes

• Rate limits: The Jolpica API rate-limits aggressive parallel requests. The app automatically backs off with exponential retry (1s → 2s → 4s). First cold-cache run will be slow — this is normal.
• Cache TTL: Race results refresh every 15 minutes. Force a full re-fetch anytime with --force-refresh.
• 2026 calendar: Hardcoded in f1_api.py (CALENDAR_2026). Update each season as the FIA publishes the official schedule.

---

Built for personal use. Not affiliated with Red Bull Racing, Formula 1, or anyone who actually knows what they're doing.