High-Frequency Data Can Mask Massive Forecasting Errors
New research reveals a 'Granularity Paradox' where finer time-series data improves in-sample fit while compounding out-of-sample failure.
Increasing the frequency of time-series data—moving from monthly to daily grains, for example—often creates a deceptive illusion of model accuracy. While finer granularity increases the available dataset size and improves in-sample diagnostics, it frequently degrades out-of-sample performance because recursive errors compound over longer horizons.
Testing 10 models across a 13-year procurement dataset, researchers found that recursive autoregressive and seasonal models collapse under high-frequency forecasting. The Holt-Winters model, for instance, reached a Test R-squared of -151 and a Total Percentage Forecast Error (TPFE) of 425.85% at the daily grain. Linear Regression remained stable across all granularities, proving the failure is driven by recursive feedback topology rather than model complexity.
https://arxiv.org/abs/2607.05450
Deep learning models exhibit a different, non-monotonic behavior. An LSTM traced a U-shaped error curve, with TPFE worsening from 19.66% at the monthly grain to 35.94% at the bi-weekly grain, before eventually overcoming the propagation penalty to hit a TPFE of 4.35% at the daily grain.
Standard pointwise metrics like RMSE and MAE systematically mask this cumulative error propagation. To solve this, the authors introduce a consensus-dissensus diagnostic that compares pointwise metrics against cumulative TPFE to identify models whose standard diagnostics hide systematic failure.