During April 2025, Thomas Blume developed a dynamic basis function feature for the NBEATS model in the Nixtla/neuralforecast repository, focusing on enhancing the model’s adaptability to diverse time series data. He architected updates that allow the integration of multiple basis types, including Legendre, polynomial, and changepoint, within the NBEATS architecture. Using Python, PyTorch, and Ray Tune, Thomas designed and validated the new dynamic basis integration, enabling more flexible and accurate forecasting pipelines. This work addressed the challenge of static basis limitations, supporting improved inventory and demand planning by reducing manual tuning and increasing the model’s responsiveness to changing data patterns.