Formula 1 Race Results Predictor

No single source had what we needed, so the data was assembled from several. The Ergast Data repository carried comprehensive historical Formula 1 data, but it had real gaps, and the rest had to be scraped. Six individual dataframes were collected and merged into one final dataset, keeping the features we believed influenced race outcome and stripping the redundant columns.

Before modeling anything we went deep on the data itself — circuit analysis, driver nationality, championship wins, and the number of races won by each driver and constructor.

Clear trends emerged: the dominance of teams like Ferrari and Mercedes by both race wins and championships, and a meaningful split in DNF ratio between driver error and constructor error, which sharpened how much reliability matters in this sport.

We also investigated how drivers and constructors perform in their home races. The plots made it obvious — home ground carries a large advantage — and that observation went on to shape the entire feature engineering approach.

We trained six classification models to predict whether a driver would finish on the podium, finish in the points, or record a DNF: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, Gaussian Naive Bayes, and K-Nearest Neighbors. They were chosen for their standing in the literature and their fit to classification problems.

Parameters were selected through research and experimentation rather than defaults. The Decision Tree used the entropy criterion to measure split quality with a bounded max_depth to prevent overfitting; the Random Forest tuned n_estimators for the number of trees and max_features to limit the features considered at each split.

Every model was assessed with k-fold cross-validation — partitioning the data into k equal subsets and training and evaluating k times, each time holding out a different subset. That gave a far more reliable estimate of generalization than a single train/test split, and let us tune hyperparameters without overfitting to one particular partition.

Random Forest was ultimately selected as the final model. Although SVC scored marginally higher on raw accuracy, Random Forest provided feature importance scores, which let us identify which features were actually driving the predictions — the inference half of the project's goal.

Accuracy was the headline metric — the proportion of correctly classified instances over the total. Alongside it we tracked precision, recall, and F1 score, which account for the trade-offs between true positive, false positive, true negative, and false negative rates, and used cross-validation throughout to test robustness rather than a single lucky split.

The first pass was disappointing. Training on the raw combined dataset, four algorithms — logistic regression, a neural network regressor, random forest, and SVC — landed between 0.50 and 0.68 accuracy. Not useless, but nowhere near what we expected.

The cause was our own assumptions. We had thrown in a large number of variables we believed would influence race outcome. They did influence it, technically — just insignificantly, and collectively they drowned the signal that mattered.

Applying the engineered features — driver and constructor confidence, home team advantage, DNF rate — plus feature selection and hyperparameter tuning produced a substantial jump across every model. SVC reached 0.95, Random Forest 0.94, Logistic Regression and K-Nearest Neighbors both 0.93, and Gaussian Naive Bayes 0.87.

The lesson was unambiguous and is the one I've carried into everything since: the choice of model mattered far less than what the model was allowed to see.

Motorsport analysis leans heavily on human intuition and subjective opinion. Taking a strictly data-driven approach to predicting outcomes was itself part of the contribution, but the novel features were the real addition — driver and constructor confidence, home team effects, and DNF ratios and indices for both drivers and constructors. Together they captured nuances of motorsport performance that the standard variables miss entirely.

The models predict podium and points bands accurately, but predicting exact finishing positions at the same confidence remains out of reach. The other clear gap is that everything here is static — incorporating real-time race data would allow dynamic, in-race prediction rather than a single pre-race call.

With better data access, both are tractable. The broader approach isn't specific to motorsport either; the same methodology transfers to other sports and to domains like finance and marketing.