Forecasting Multivariate Time Series under Predictive Heterogeneity: A Validation-Driven Clustering Framework
Forecasting Multivariate Time Series under Predictive Heterogeneity: A Validation-Driven Clustering Framework
Authors: Ziling Ma, Ángel López Oriona, Hernando Ombao, Ying Sun Date: 2026-04-15 Paper ID: openalex:2604.13748
Summary
This paper addresses the problem of predictive heterogeneity in high-dimensional multivariate time series forecasting by proposing a validation-driven adaptive pooling framework. Unlike methods that group series based on representation similarity, this approach partitions data based on out-of-sample predictive risk estimated through validation error. A leakage-free fallback mechanism ensures that specialization is only employed when it demonstrably outperforms a global model, providing a robust safeguard against negative transfer. The framework supports both point and probabilistic forecasting and maintains efficiency in high-dimensional settings.
Key Contributions
- Formulates adaptive pooling as a statistical decision problem, prioritizing predictive risk over representation similarity.
- Introduces a leakage-free fallback mechanism that dynamically switches to a global model to prevent performance degradation from poor specialization.
- Demonstrates robust performance on large-scale traffic datasets using Huber and pinball loss functions for point and probabilistic forecasting.
Open Questions & Future Work
Key Concepts
- validation-driven-adaptive-pooling: A framework for adaptive pooling in multivariate time series that uses validation-based predictive risk to determine optimal model specialization.
Archivist Review
I have approved the core methodological contribution of validation-driven adaptive pooling, as it offers a principled, performance-oriented alternative to traditional feature-based clustering in time series. I also approved the open question regarding soft vs. hard specialization, as it directly addresses a structural limitation of existing clustering approaches in non-stationary or overlapping forecasting regimes. Other candidates were rejected for being generic or lacking sufficient specificity for the vault.
Approved Concepts
- Validation-driven Adaptive Pooling: It shifts the focus of time series clustering from feature-space similarity to predictive-performance-based partitioning.
Approved Open Questions
- Soft vs Hard Specialization: Addressing hard assignment restrictions is crucial for improving forecasting robustness in settings where time series do not belong clearly to a single regime, potentially enabling smoother and more accurate predictive performance.
Rejected Candidates
- [dataset] Traffic datasets (
traffic-datasets) - generic: Generic reference to a broad category of data rather than a specific, named benchmark dataset.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.13748
- paper_id
- 2604.13748
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- validation-driven-adaptive-pooling
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-18T05:34:17Z