Paper: arXiv 2603.22886

Authors: Minkey Chang, Jae-Young Kim

Abstract

We propose the Identifiable Variational Dynamic Factor Model (iVDFM), which learns latent factors from multivariate time series with identifiability guarantees. By applying iVAE-style conditioning to the innovation process driving the dynamics rather than to the latent states, we show that factors are identifiable up to permutation and component-wise affine (or monotone invertible) transformations. Linear diagonal dynamics preserve this identifiability and admit scalable computation via companion-matrix and Krylov methods. We demonstrate improved factor recovery on synthetic data, stable intervention accuracy on synthetic SCMs, and competitive probabilistic forecasting on real-world benchmarks.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents strong theoretical rigorousness with identifiability proofs for variational dynamic factor models, addressing a major gap in latent variable representations for time series. Empirical validation is solid across synthetic benchmarks and probabilistic forecasting tasks, though market-specific backtests are absent.

Research Flowchart

  flowchart TD
    A[Research Goal: Identifiable Latent Factors<br>in Multivariate Time Series] --> B[Input Data: Synthetic SCMs &<br>Real-World Benchmarks]
    B --> C[Methodology: Apply iVAE Conditioning<br>to Innovation Process in VDFM]
    C --> D[Computation: Linear Diagonal Dynamics via<br>Companion-Matrix & Krylov Methods]
    D --> E[Key Outcome 1: Theoretical Identifiability<br>Up to Permutation & Affine Scale]
    D --> F[Key Outcome 2: Improved Factor Recovery &<br>Competitive Probabilistic Forecasting]

Cited References (25)

  1. On lines and planes of closest fit to systems of points in space (1901)
  2. Independent component analysis: Algorithms and applications (2000)
  3. Auto-Encoding Variational Bayes (2014)
  4. Variational Autoencoders and Nonlinear ICA: A Unifying Framework (2020) — link
  5. Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning (2019) — link
  6. On the Identifiability of Switching Dynamical Systems (2024) — link
  7. Forecasting Using Principal Components From a Large Number of Predictors (2002)
  8. Deep Dynamic Factor Models (2020) — link
and 17 more
  1. Macroeconomics and Reality (1980)
  2. Identification and Estimation of Dynamic Causal Effects in Macroeconomics Using External Instruments (2018)
  3. Causality: Models, Reasoning, and Inference (2009)
  4. Elements of Causal Inference: Foundations and Learning Algorithms (2017)
  5. Causation, Prediction, and Search (2000)
  6. Toward Causal Representation Learning (2021)
  7. Efficiently Modeling Long Sequences with Structured State Spaces (2022) — link
  8. Effectively Modeling Time Series with Simple Discrete State Spaces (2023)
  9. Mamba: Linear-Time Sequence Modeling with Selective State Spaces (2023) — link
  10. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2023) — link
  11. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (2024) — link
  12. TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables (2024)
  13. Deep Kalman Filters (2015) — link
  14. Strictly Proper Scoring Rules, Prediction, and Estimation (2007)
  15. Scoring Rules for Continuous Probability Distributions (1976)
  16. New Introduction to Multiple Time Series Analysis (2005)
  17. Time Series Analysis (1994)

0 of 25 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.