The Granularity Paradox: How Temporal Disaggregation Inflates In-Sample Fit and Compounds Out-of-Sample Error

2026-07-05 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

The "Granularity Paradox" in time-series forecasting reveals that finer temporal disaggregation, like moving from Monthly to Daily data, improves in-sample diagnostics and dataset size but degrades out-of-sample accuracy due to recursive error compounding over longer horizons. This trade-off was formalized and benchmarked across 10 models (naïve, statistical, ML, deep learning) using a 13-year public procurement dataset at six granularities. Empirical results show recursive models, such as Holt-Winters, degrade substantially at high frequencies (Daily grain: Test R-squared -151, TPFE 425.85%). The LSTM, however, traces a U-shaped error curve, improving at Daily (TPFE 4.35%, R-squared 0.66) after worsening from Monthly (19.66%) to Bi-Weekly (35.94%). Linear Regression remained stable (16.3-17.0% TPFE), confirming the paradox is driven by recursive feedback, not model complexity. Standard pointwise metrics (RMSE, MAE) systematically mask cumulative error, necessitating goal-dependent cumulative metrics. A consensus-dissensus diagnostic is introduced to identify models masking systematic error propagation.

Key takeaway

For Machine Learning Engineers evaluating time-series models, you must prioritize cumulative error metrics like TPFE over standard pointwise metrics (RMSE, MAE) to avoid misleading assessments of model adequacy. Understand that recursive feedback topology, not just model complexity, drives performance degradation at finer granularities. When disaggregating data, carefully assess how your chosen model's architecture handles error propagation, as some, like LSTM, can adapt better at daily frequencies.

Key insights

Finer temporal disaggregation improves in-sample fit but degrades out-of-sample accuracy due to recursive error compounding.

Principles

• Recursive models degrade substantially under high-frequency forecasting.
• Pointwise metrics can systematically mask cumulative error propagation.
• The granularity paradox is driven by recursive feedback topology.

Method

A consensus-dissensus diagnostic compares pointwise metrics against cumulative TPFE across granularities, identifying models whose standard diagnostics mask systematic error.

In practice

• Evaluate forecasts using goal-dependent cumulative metrics.
• Consider model feedback topology when choosing data granularity.
• LSTM can overcome error propagation at daily granularity.

Topics

• Time-series Forecasting
• Granularity Paradox
• Error Propagation
• Forecasting Metrics
• LSTM
• Holt-Winters

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Discovering Entity-Conditioned Lag Heterogeneity: A Lag-Gated Neural Audit Framework for Panel Time Series — AC-GATE provides a verifiable framework to expose entity-conditioned lag heterogeneity as a structured, inspectable model output.
• QuantSightBench: Evaluating LLM Quantitative Forecasting with Prediction Intervals — LLMs consistently exhibit overconfidence in quantitative forecasting, failing to achieve target prediction interval coverage.
• Causal Semantic Alignment for LLM-based Time Series Forecasting — CVAformer disentangles time series variables into invariant and dynamic components to improve LLM-based forecasting accuracy by mitigating confounding effects.
• Filtered ANN as a Phase Transition: When Selectivity-Estimation Error Causes Plan Regret — Selectivity-estimation error causes significant plan regret in filtered ANN queries, especially near strategy phase boundaries.
• Granularity-Regulated Adaptive Computational Efficiency for Optimal Verification in Test-Time Scaling — Optimal LLM verification granularity depends on problem difficulty and compute budget, exhibiting a phase transition.
• When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis — Fragility, a new metric, reveals LLM pre-training dynamics invisible to standard probing accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.