Lucas Reymond

I competed in the QRT (Qube Research & Technologies) Data Challenge hosted by ENS, where the goal was to predict the outcome of a football match (home win, draw, or away win) from pre-match team and player statistics. The data was fully anonymized: no team, player, or league names, every feature rank-scaled, and a test set drawn from different seasons than the training data, which made the problem both noisy and genuinely difficult. I built features that aggregate individual player histories up to the team level and contrast the two sides of each match, then trained a diverse ensemble combining gradient boosting (LightGBM, XGBoost, CatBoost) with neural networks, including a set-based model that learns directly from a team’s raw squad of players. Most of my work went into validation: I relied on nested cross-validation to choose between models, so I was never misled by configurations that only looked strong in-sample. The final blended model placed 19th out of 960 participants (top 2%). Beyond the ranking, the project taught me how to extract signal from low-signal data.