Signals, Concepts, and Laws: Toward Universal, Explainable Time-Series Forecasting
Hongwei Ma, Junbin Gao, Minh-Ngoc Tran
TL;DR
This paper introduces DORIC, a novel transformer-based model for multivariate time-series forecasting that emphasizes explainability, physical credibility, and domain universality by leveraging self-supervised concepts and physics-based constraints.
Contribution
The paper presents DORIC, a new model that combines domain-agnostic concepts with differential equation regularization to improve interpretability and physical consistency in time-series forecasting.
Findings
DORIC achieves superior accuracy across multiple domains.
The model provides interpretable insights through its concept-based approach.
Enforces physical constraints to enhance forecast credibility.
Abstract
Accurate, explainable and physically credible forecasting remains a persistent challenge for multivariate time-series whose statistical properties vary across domains. We propose DORIC, a Domain-Universal, ODE-Regularized, Interpretable-Concept Transformer for Time-Series Forecasting that generates predictions through five self-supervised, domain-agnostic concepts while enforcing differentiable residuals grounded in first-principles constraints.
Peer Reviews
Decision·Submitted to ICLR 2026
The paper’s main strength is its innovative integration of concept bottlenecks and physics-informed residuals within a Transformer, effectively combining interpretability with physical plausibility. It presents a comprehensive evaluation across six diverse datasets and strong baselines, supported by detailed ablation studies that clearly show each component’s contribution. The authors emphasize interpretability through five structured latent concepts and provide theoretical grounding via express
1. The paper’s central claim of interpretability is not convincingly demonstrated. While the architecture enforces a five-concept bottleneck, there is no theoretical proof or empirical validation that these learned concepts retain their intended meanings. Evidence is limited to internal correlations, without visual, human, or domain-level verification of interpretability. 2. The paper introduces a driven–damped ODE as the core of its “physics-informed” design, but the justification is mostly heu
-interpretability is done through five fundamental concepts: they can be computed directly from the data and used in a soft-supervised fashion, avoiding the need for labeled concept datasets. - The model demonstrates cross-domain generalization across diverse datasets (electricity, traffic, weather, epidemiology, finance). - There are theoretical results showing doric will eventually yield good enough result (despite the interpretability), with physical plausibility.
-My main problem is that interpretability is claimed but not really studied or demonstrated. The paper does not present any empirical or qualitative evidence that the learned concepts are interpretable or used meaningfully by the model. Overall it remains unclear how these scores are used by the model and how much the final forecast was determined due to the concept scores.There are no visualisations, case studies to show, or attribution analyses showing how forecasts depend on concept activat
1. The motivation in this paper is great. 2. Physics constraints are important for time series forecasting.
1. The quality of the paper writing is weak. The subscripts and asterisk in Eqs. (10-14) are not common in time series publications, but without clear explanations. For example, it would be great to explain 1t, ..., 5t. 2. In your third contribution, you mentioned adaptive positional encoding for a domain-universal architecture, but I did not find the details on that. Moreover, shared concept heads seem to be used to train the model with all time series together, which implies the universal fea
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTime Series Analysis and Forecasting · Forecasting Techniques and Applications