Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting
The study investigates how assembling fixed-length synthetic time series from differentially private generators affects downstream forecasting performance. By testing various overlap rates and weighting schemes across different forecasting models, the research found that optimal assembly depends heavily on the specific forecaster used. This suggests sequence assembly is a forecaster-aware design choice, guiding practitioners to select configurations based on their model's characteristics.