level: research
this paper introduces the granularity paradox in time-series forecasting. finer temporal disaggregation, like moving from monthly to daily data, increases the number of observations and improves in-sample diagnostics. however, it also leads to worse out-of-sample accuracy because errors compound over longer forecast horizons. coarse aggregation, such as annual data, avoids recursive error propagation but reduces the data available for model estimation.
the study benchmarks ten models across six granularities using a 13-year public procurement dataset. models include naive, statistical, machine learning, and deep learning approaches. results show a non-monotonic threshold structure. recursive autoregressive and seasonal models degrade sharply at high frequencies. for example, holt-winters achieves a test r-squared of -151 and a total percentage forecast error of 425.85% at the daily grain. in contrast, lstm exhibits a u-shaped error curve, worsening from monthly to weekly but improving at daily granularity.
the findings highlight a trade-off between data quantity and forecast horizon length. practitioners must balance the benefits of more data points against the costs of longer recursive error chains. the paradox suggests that optimal granularity depends on the model type and the forecasting task. simple models may fail at high frequencies, while some neural networks can recover performance at very fine grains.
why it matters: choosing the right time granularity is crucial for building accurate forecasting models, as finer data can mislead model selection and hurt real-world predictions.