Finer time-series data looks better until you actually try to forecast with it.
Researchers behind arXiv:2607.05450 (Bilal et al., 2026-07-08) tested 10 models across six temporal granularities — from annual down to daily — using 13 years of public procurement data. The headline finding: disaggregating to finer intervals inflates in-sample diagnostics while degrading out-of-sample accuracy, because recursive autoregressive models accumulate errors over longer horizons. Holt-Winters illustrated the extreme end of this, reaching a Test R-squared of -151 and a Total Percentage Forecast Error (TPFE) of 425.85% at the daily grain. Meanwhile, Linear Regression held steady across all granularities, posting 16.3-17.0% TPFE regardless of how fine the data got — suggesting the paradox is a function of recursive feedback topology, not model sophistication.
The more interesting result is the LSTM's U-shaped error curve. It degraded from a monthly TPFE of 19.66% through a bi-weekly peak of 35.94% before recovering sharply to 4.35% at the daily grain — outperforming Linear Regression at that resolution and posting an R-squared of 0.66. That pattern implies LSTMs can eventually absorb enough signal from dense data to overcome error propagation, but only after passing through a dangerous middle zone where neither granularity helps. Practitioners who benchmark only at one or two intervals would miss this entirely.
The paper also argues that standard pointwise metrics like RMSE and MAE systematically hide cumulative error drift, and introduces a consensus-dissensus diagnostic to flag models whose tidy-looking scores mask systematic propagation failures. That critique lands in a forecasting culture that still defaults to RMSE as a proxy for adequacy — a habit this research suggests is worth breaking.