Granularity Paradox Reveals Forecasting Errors in Time Series Disaggregation.

Time-series forecasting often involves a critical trade-off known as the "Granularity Paradox." This phenomenon describes how disaggregating data into finer temporal units, such as moving from monthly to daily, can misleadingly improve in-sample model fit and increase dataset size. However, this apparent improvement often leads to a significant degradation in out-of-sample accuracy, primarily because errors compound recursively over longer forecasting horizons. Conversely, using coarser aggregations, like annual data, can eliminate recursive error propagation but reduces the amount of data available for model training. The research formalizes this trade-off and evaluates ten different forecasting models—spanning naive, statistical, machine learning, and deep learning architectures—across six granularities using a 13-year public procurement dataset. The findings show a non-monotonic error structure: recursive models like Holt-Winters perform poorly at high frequencies, while LSTMs exhibit a U-shaped error curve, eventually overcoming propagation penalties at daily granularity. Linear Regression remained stable, indicating that the paradox is driven by recursive feedback topology rather than model complexity. The study emphasizes that standard pointwise metrics (RMSE, MAE) systematically mask cumulative error propagation, advocating for goal-dependent cumulative metrics to accurately assess model adequacy.