Work

Better Imputation Does Not Guarantee Better Forecasting: Transmission Efficiency in Two-Step Time Series Pipelines

Year: 2026
Type: preprint
Abstract: Reconstruction accuracy often guides imputer selection for time series forecasting. We examine this practice on four datasets under a historical-window protocol. Inputs exclude observations after the forecast origin; learned imputers are fitted only on training data and frozen during validation and testing. The experiment comprises $2{,}880$ unique configurations and $8{,}640$ training runs across three forecasting seeds ($42$, $123$, $2026$), five imputers and three forecasters: long short-term memory (LSTM), a decomposition-based linear model (DLinear), and a patch-based time series transformer (PatchTST). Forecasting errors are averaged across seeds within each configuration before matched comparisons with Mean imputation. Among $2{,}127$ valid comparisons, median relative changes are $-2.90\%$ for forecasting mean absolute error (MAE) and $-41.22\%$ for reconstruction MAE. The median transmission ratio is $0.0449$. Reconstruction improves while forecasting deteriorates in $19.51\%$ of valid comparisons, with model-specific rates of $34.41\%$ for LSTM, $6.91\%$ for DLinear and $17.21\%$ for PatchTST. Across all $2{,}304$ non-Mean configurations, $74.65\%$ retain the same forecasting-error-change sign across the three seeds. Reconstruction accuracy alone is therefore insufficient to assess downstream forecasting performance under these settings.
Language: English
Open Access status: green
OpenAlex ID: W7235337255
Cites: 24
Cited by:
Related to: 10