Flower
Data Science & Machine Learning Intern
Jun 2026 – Aug 2026
- Built a backtest measuring the strength and independence of each forecast provider's updates. Showed that an expensive provider, though accurate on RMSE and MAE, added little to the cheaper ensemble, especially during large imbalances. The team decided against the investment, avoiding several thousand euros per month.
- Traced a recurring under-forecasting pattern to mislabeled turbine downtime data, present in both live inference and the training set, and flagged it to the team.
- Diagnosed miscalibrated probabilistic forecasts, tracing the problem partly to power-based reweighting of the training data. Prototyped adaptive conformal quantile prediction and showed it restored the target coverage.
