Lineup Optimizer
Find the best F1 Fantasy lineups within your budget, or get transfer recommendations for your current team.
How does the Lineup Optimizer work?
Checks all 1.4 million possible team combinations to find the best lineup within your budget. Uses ML predictions for qualifying and race positions.
- Lock picks: Left-click a driver/constructor card to lock them in your lineup
- Exclude picks: Right-click to exclude a player from consideration
- Chips: Select your active chip to see optimized lineups with boost applied
Lock / Exclude Picks
Left-click to lock (force into lineup). Right-click to exclude (remove from consideration). Constructors shown in ALL CAPS.
Data Analysis
No FP analysis data available. Run pipeline/10_fp_analysis.py first.
No post-race analysis data available.
2026 Season
Race Calendar & Results
Championship Standings
Official F1 World Championship points based on race and sprint results.
Fantasy Points Standings
Driver Price Tracker
Constructor Price Tracker
Head-to-Head Matchup
Compare any two drivers side-by-side. See who to pick based on predictions, value, and historical performance.
Model Accuracy
Track how well our predictions match reality. Updated after each race weekend.
Loading accuracy data...
Race Deep Dive
Detailed post-race telemetry analysis with fuel-corrected pace, sector breakdowns, and tyre degradation.
Select a round to view deep dive analysis...
Latest Videos
Race weekend analysis, predictions, and commentary from the BoxBox F1 Fantasy YouTube channel.
Subscribe on YouTubeLoading videos...
Race Weekend Articles
Analysis, insights, and commentary from each race weekend.
Loading articles...
How It Works
BoxBoxF1Fantasy uses a multi-stage machine learning pipeline to predict F1 race outcomes and convert them into fantasy point projections. Here's the full process from raw data to your lineup recommendations.
1. Data Collection
We pull from two sources: FastF1 for live telemetry (lap times, sectors, tyre compounds, weather, track status) and the Jolpica API for historical results, qualifying, and championship standings from 2020-2026.
2. Feature Engineering
Two feature layers are built. Layer 1: Jolpica historical priors — rolling 3/5/10-race averages, team strategy ratings, driver skill ratings, and track classification features (always available). Layer 2: Free practice telemetry — 27+ engineered features including pace, consistency, degradation, long-run performance, and sector analysis (when FP data is available).
3. ML Predictions
Three XGBoost models are trained with walk-forward temporal validation on 2,600+ historical rows. The qualifying model predicts grid positions, which feed into the race model for finish predictions. A third model scores confidence based on data completeness and model agreement.
4. Monte Carlo Simulation
10,000 race simulations sample from prediction distributions with calibrated noise — stochastically modelling DNFs, overtakes, fastest laps, and DOTD. Because fantasy scoring is nonlinear, the expected points of a distribution differ from points of the expected position.
5. Fantasy Points
Monte Carlo outputs are aggregated into mean, median, P5-P95 percentile ranges. Combined with driver prices to compute PPM (Points Per Million) value scores. The lineup optimizer then finds the best team combinations within your budget.
6. Post-Race Analysis
After each race, the deep dive pipeline processes full telemetry: fuel-corrected lap times, sector breakdowns, tyre degradation curves, dirty-air effects, race momentum phases, and speed trap data — all viewable in the Race Deep Dive tab.
Model Details
- Qualifying model: XGBoost (1200 trees, lr=0.025, depth=3) — predicts qualifying position from historical priors + FP telemetry
- Race model: XGBoost (650 trees, lr=0.03, depth=5) — predicts finish position using predicted quali as input feature
- Confidence model: ExtraTrees classifier trained on FP-available rows, scoring prediction reliability 0-100
- Monte Carlo: 10,000 simulations with calibrated noise, DNF sampling (rolling 5-race probability, capped at 25%), stochastic overtake/fastest-lap/DOTD assignment
- Walk-forward validation: Models are only ever trained on past data — no future leakage. Jolpica features use shift(1) before rolling to prevent look-ahead bias
- Feature count: 136 features total across both layers, with XGBoost handling missing FP data natively via histogram-based tree method
Pipeline Automation
- Weekend orchestration: Automated phases — pre-FP (download + train), post-FP (features + predict), post-quali (re-predict with quali data), post-race (actuals + analysis)
- Weather updates: Open-Meteo forecasts every 6 hours via GitHub Actions, showing rain probability for Friday/Saturday/Sunday
- Deep dive generation: 7-step data cleaning (null removal, accuracy filter, pit/lap-1/safety-car exclusion, fuel correction, outlier removal) then 10+ analysis modules
Disclaimer
These predictions are for entertainment purposes only. F1 races are inherently unpredictable — weather, crashes, strategy calls, and reliability can dramatically alter outcomes. Use these predictions as one input for your fantasy decisions, not the only one.