BTK Datathon 2026 — Career Success Score Prediction

This is my project for BTK Akademi''s Datathon 2026 competition, developed in collaboration with Google and the Turkish Entrepreneurship Foundation. The goal was to predict a continuous career_success_score (0-100) for students based on academic, technical, project, portfolio, interview, and social profile data - evaluated using Mean Squared Error (MSE), where lower is better. Rather than building one large model, I approached this iteratively: six versions, each adding exactly one idea on top of the last, each submitted to Kaggle to measure its real impact. The public MSE dropped steadily from 99.09 (baseline XGBoost) to 88.94 (final model) across these six iterations. The dataset combined numeric features (skill scores), categorical features (department, target role, university tier), and a natural language field - a short Turkish mentor feedback text - requiring genuine NLP work rather than just tabular modeling. Key techniques used: smart missing-value imputation based on domain logic (rather than blind median filling), feature engineering (weighted skill averages, interaction terms, log transforms), a three-layer NLP strategy on the mentor feedback (sentiment word counting, sentiment ratio, and TF-IDF) that reached a 0.40 correlation with the target, Optuna-tuned XGBoost + LightGBM + CatBoost combined via Ridge-regression stacking, and two final refinements in the last version: out-of-fold target encoding for categorical features, and a soft-blend correction for a ceiling effect in the target variable (773 students scored a perfect 100). The final model is a stacking ensemble across 167 engineered features, reaching 88.94 public / 89.88 private MSE on Kaggle - a 10.15-point improvement over the baseline.