Compose the Horizon: Parameter-Shared Blockwise Forecasting under a Sequential-Compute Trade-off

Key points

• Most long-horizon forecasters emit the complete future in one pass.1
• We ask whether a compact operator can instead be reused to emit contiguous future blocks, trading stored parameters for sequential computation.1
• The blocks occupy disjoint future intervals and are concatenated, not added as trend, seasonal, or residual components, and each emitted block joins the rolling context that conditions the next application.1
• In controlled three-seed experiments, removing forward feedback hurts all four evaluated datasets while stopping gradients through it changes little, so the useful signal is the evolving forecast state itself; reapplying the shared operator helps consistently, but by less than widening it.1

Sources (1)

• [1]Compose the Horizon: Parameter-Shared Blockwise Forecasting under a Sequential-Compute Trade-off
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 07:19 PM

  Recursive Functional Decomposition (RFD) factorizes the horizon mapping into three parameter-shared block predictions.

  Most long-horizon forecasters emit the complete future in one pass.

Extractive summary: sentences quoted from the sources.