CombinationTS: A Modular Framework for Understanding Time-Series Forecasting Models

TL;DR

CombinationTS introduces a modular evaluation framework for time-series models, enabling robust component attribution and revealing that simpler structures often outperform complex ones when data views are well-designed.

Contribution

It proposes a probabilistic, modular evaluation framework that decomposes models into orthogonal components for better attribution and understanding of architectural effectiveness.

Findings

A well-designed data view (Embedding) can make simple models competitive with complex ones.

Explicit structural priors improve performance-stability trade-offs over increasing encoder complexity.

The Identity Paradox shows simple, parameter-free encoders often match or outperform complex backbones.

Abstract

Recent progress in time-series forecasting has led to rapidly increasing architectural complexity, yet many reported State-of-the-Art gains are statistically fragile or misattributed. We argue that progress requires a shift from model selection to modular attribution, identifying which components truly drive performance. We propose CombinationTS, a self-contained probabilistic evaluation framework that decomposes forecasting models into orthogonal modules--Input Transformation, Embedding, Encoder, Decoder, and Output Transformation--and evaluates them under a shared evaluation condition space. By quantifying each component via marginalized performance ($\mu$) and stability ($\sigma$), CombinationTS enables robust attribution beyond fragile point estimates. Through large-scale paired evaluation, we uncover the Identity Paradox: once the data view (Embedding) is well-designed, a parameter-free Identity Encoder often matches or outperforms complex backbones. We further show that explicit structural priors introduced via Input Transformations yield a more favorable performance-stability trade-off than increasing Encoder complexity, establishing a principled baseline for architectural necessity.