Cognitive Performance Prediction (Datathon 2026)

Regression models I developed for Datathon 2026 (YZTA) to predict a cognitive performance score from variables such as sleep, stress, physical activity, working hours and demographics. Submissions were evaluated on Kaggle using RMSE; the repository contains the best-performing model versions and the final blending script.

Approach

• Missing-value indicator flags (treating missingness as informative) and IterativeImputer for imputation
• Outlier clipping on variables such as body mass index
• Engineered features such as sleep irregularity, sleep efficiency, stress × workload and a combined health score
• CatBoost regressors trained with 5-fold cross-validation, plus a pseudo-labelling experiment
• Final prediction from a weighted blend of the three best models (0.40 / 0.30 / 0.30)

Each model version (v14, v21, v22, v23) is kept as a separate script, so the effect of each technique can be compared step by step.

Technologies

Python, pandas, NumPy, CatBoost, scikit-learn.