Advanced Computer Lab 9th Sem

I built this in Advanced Computer Lab, 9th semester, to score, after qualifying, whether a driver will finish in the top three. The model never sees the race it is scoring. Training is 1950–2019, the choice between models is 2020–2021, and the test is the 68 Grands Prix of 2022–2024.

What it is allowed to know

A current-race result, lap, pit stop, or points total cannot be a predictor for that same race. The site labels a model score, a historical result, and a what-if separately, and it treats the score as a ranking signal rather than a probability. On the 2024 season the model's top three and the first three on the grid both found 2.00 of the 3 podium drivers, on average. Qualifying already carries most of that. The extra output is a score for every starter, from the grid, recent form, and the championship standings.

Completed Work

Leakage-safe pipeline

One Python pipeline cleans the Ergast tables, builds the driver-race rows, and trains a logistic regression, a random forest, and a shallow network. Logistic regression had the best validation PR-AUC, 0.730 against 0.717 for the other two, so that is the model the site uses. Tests check the grain, the chronological split, and that a race result cannot leak into its own predictors. The pipeline also writes permutation importance and SHAP explanations.

Built with
Python, pandas, scikit-learn, SHAP
Contribution
I wrote the pipeline, the tests, and the notebook that runs the same analysis.

Race site

A site for those 68 held-out races. Each race has the circuit from the F1 timing map, the model's top three, a what-if on the exported weights, and the real classification, which stays hidden until the prediction is already on screen.

Contribution
I exported the fitted model and the race data, and built the site around that cutoff.

Course skills

Data work

Data cleaning, Feature engineering, Leakage control

Modeling

Classification, Chronological split, Model comparison, Explainability

Held-out scores

Test PR-AUC
0.702
Test ROC-AUC
0.935
Validation PR-AUC, logistic regression
0.730
Held-out races, 2022–2024
68

Looking back

  • Tying the starting grid on the top three is not a weak result. Qualifying already names most of the podium. The useful output is the score for the drivers behind that, and not presenting that score as a chance.