Identifiable Variational Dynamic Factor Model (iVDFM)

• iVDFM is a probabilistic generative model for multivariate time series that guarantees identifiable latent innovations by conditioning on the innovation process, uniquely defining factor trajectories up to permutation and affine transformations.
• It employs diagonal dynamics, exponential-family innovation priors, and an injective decoder to restrict the equivalence class of latent factors, enhancing interpretability and stability.
• The model exhibits competitive forecasting performance and supports intervention analysis, making it well suited for dynamic factor extraction, causal inference, and structural reasoning.

Searching arXiv for the cited papers and closely related work on identifiable variational dynamic factor models. The Identifiable Variational Dynamic Factor Model (iVDFM) is a probabilistic generative model for multivariate time series in which latent factors are learned with explicit identifiability guarantees. Its defining move is to apply iVAE-style conditioning to the innovation process driving the dynamics rather than to the latent states themselves, so that the innovations, and hence the factor trajectories generated from them, are identifiable up to permutation and component-wise affine transformations, or, under stronger regularity, up to permutation and component-wise monotone invertible transformations. In its core form, iVDFM combines conditional exponential-family innovation priors, linear diagonal or per-component companion-form dynamics, an injective decoder, and variational inference with scalable companion-matrix and Krylov methods (Chang et al., 24 Mar 2026).

1. Historical and methodological setting

The immediate background to iVDFM is the long-standing identification problem in dynamic factor models (DFMs). In linear macroeconometric DFMs, the impulse-response function can be written as $k(z)=d(z)c(z)^{-1}$, but the same $k(z)$ can be generated by many different pairs $(d(z),c(z))$ through observationally equivalent postmultiplications. The right matrix fraction description in reversed echelon form, RMFD-E, was introduced precisely to select a unique representative of that equivalence class, nesting the Bai–Wang normalization as a special case and allowing overidentifying restrictions through Kronecker indices (Koistinen et al., 2022).

A parallel line of work pursued practical identifiability through Bayesian regularization rather than exact structural uniqueness. Sparse dynamic factor models with loading selection by variational inference augment the measurement equation with binary inclusion indicators $Z$, yielding $y_t=(Z\circ\Lambda)F_t+\epsilon_t$, and use spike-and-slab-like priors so that factors become associated with distinctive subsets of variables or lags. That approach improves practical identifiability by favoring sparse, interpretable loading patterns, but it does not by itself provide formal uniqueness of the factor space (Spånberg, 2022).

A third precursor comes from identifiable variational autoencoders for spatio-temporal blind source separation. In that literature, observations are generated as $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$, while the latent components have exponential-family priors conditioned on auxiliary variables $\mathbf{u}$. The crucial insight is that sufficiently rich variation in the natural parameters $\boldsymbol{\lambda}(\mathbf{u})$ can break the generic non-identifiability of nonlinear latent-variable models. This framework was first adapted to spatio-temporal settings without explicit Markov state dynamics (Sipilä et al., 2024), and then to autoregressive latent processes with nonstationary AR parameters, where latent components become identifiable up to the familiar permutation-scaling-translation class in the Gaussian AR case (Sipilä et al., 15 Sep 2025).

This suggests that iVDFM occupies a precise intersection: it inherits the DFM concern with factor dynamics and impulse-response structure, while importing the iVAE mechanism for identifiability into the dynamic setting by relocating the conditioning structure to the innovations.

2. Generative specification

In iVDFM, the observed series are $\boldsymbol{y}_t\in\mathbb{R}^N$, the latent factors are $\boldsymbol{f}_t\in\mathbb{R}^r$, and the innovations are $k(z)$0. There is also an observed auxiliary variable $k(z)$1 and a deterministic regime embedding $k(z)$2, computed from $k(z)$3 by a neural network. The innovation process is specified componentwise as an exponential family,

$k(z)$4

The regime embedding is deterministic:

$k(z)$5

There is no sampled regime variable; the regime weights are functions of the observed context (Chang et al., 24 Mar 2026).

The factor dynamics are linear and diagonal:

$k(z)$6

with $k(z)$7 diagonal. In the multi-regime formulation,

$k(z)$8

where each $k(z)$9 is diagonal. Within each training window, $(d(z),c(z))$0 is frozen so that the dynamics are constant in that window (Chang et al., 24 Mar 2026).

The observation model is

$(d(z),c(z))$1

where $(d(z),c(z))$2 is an injective decoder, typically a neural network, and the noise distribution has full support. The full joint model can therefore be written as

$(d(z),c(z))$3

A distinctive structural feature is that the only stochastic latent variables in the generative model are the innovations; the factors are deterministic functions of past innovations and parameters (Chang et al., 24 Mar 2026).

Higher-order temporal dependence is handled through AR($(d(z),c(z))$4) companion form. With

$(d(z),c(z))$5

the dynamics become

$(d(z),c(z))$6

and unrolling yields

$(d(z),c(z))$7

By construction, the companion blocks remain block-diagonal across factor components (Chang et al., 24 Mar 2026).

3. Identifiability mechanism

The central theoretical claim is that identifiability is established first at the innovation level and then propagated through the dynamics to the factors. The innovation prior is a product exponential family with linearly independent sufficient statistics, and the natural-parameter map $(d(z),c(z))$8 must vary sufficiently across contexts. More precisely, the model assumes that the conditional innovation prior has exponential-family structure, the regime weights are deterministic given $(d(z),c(z))$9, the decoder is injective on the support of the factors, the observation noise has full support and admits deconvolution, and the natural-parameter map spans a full-rank subspace for sufficiently many distinct values of $Z$0 (Chang et al., 24 Mar 2026).

Under these conditions, the innovation sequence $Z$1 is identifiable from $Z$2 up to permutation and component-wise affine transformations, or, with stronger regularity, up to permutation and component-wise monotone invertible transformations. Because the factor dynamics are diagonal or block-diagonal by component, each factor depends only on its own past and its own innovation component. Consequently, the factor trajectories $Z$3 inherit the same equivalence class: no cross-factor rotations are introduced by the dynamics (Chang et al., 24 Mar 2026).

This is the decisive difference from classical Gaussian DFMs. In a standard Gaussian factor model, arbitrary invertible linear transformations of the latent factors can often be absorbed by the observation and transition operators. iVDFM excludes that symmetry in two ways. First, the innovations are modeled as non-Gaussian exponential-family variables rather than Gaussian shocks. Second, the dynamics are restricted to diagonal or per-component companion structure. The paper’s Gaussian-degeneracy argument shows that if the innovations are Gaussian, then the log-density remains quadratic under arbitrary invertible linear transformations, and the model collapses back to the classical rotational indeterminacy of Gaussian factor models (Chang et al., 24 Mar 2026).

The role of diagonal dynamics is therefore structural rather than cosmetic. If one replaced the diagonal $Z$4 and $Z$5 by full matrices, arbitrary linear mixing of innovations could be reabsorbed into the state dynamics, reintroducing rotational ambiguity. The identifiability guarantee is specific to diagonal or block-diagonal per-component dynamics (Chang et al., 24 Mar 2026).

A common misunderstanding is that “identifiable” means fully unique in an absolute sense. The stated equivalence class is narrower than in classical DFMs, but it is not trivial: permutation, component-wise affine maps, and, in the stronger formulation, component-wise monotone invertible maps remain admissible. In earlier nonlinear autoregressive identifiable VAE formulations, the Gaussian AR case similarly yielded identifiability only up to permutation, diagonal scaling, and translation (Sipilä et al., 15 Sep 2025).

4. Variational learning and numerical structure

Because the decoder is nonlinear, the exact posterior over innovations is intractable. iVDFM therefore uses a variational approximation of the form

$Z$6

with each factor modeled as a diagonal Gaussian,

$Z$7

The encoder may use the entire observation sequence $Z$8 to parameterize the local posterior moments. Sampling uses the usual reparameterization trick,

$Z$9

The evidence lower bound is

$y_t=(Z\circ\Lambda)F_t+\epsilon_t$0

The first term is a reconstruction log-likelihood; the second aligns the approximate posterior over innovations with the context-conditioned innovation prior (Chang et al., 24 Mar 2026).

Scalability follows from the fact that only the innovations are sampled. The latent dimensionality of the stochastic part is $y_t=(Z\circ\Lambda)F_t+\epsilon_t$1, typically much smaller than the observation dimension $y_t=(Z\circ\Lambda)F_t+\epsilon_t$2, while factors and observations are deterministic given innovations. For AR($y_t=(Z\circ\Lambda)F_t+\epsilon_t$3) dynamics, the companion-form representation leads to linear filtering structure, and the paper explicitly exploits companion-matrix and Krylov methods, together with FFT-based convolution, to evaluate long-horizon state evolution efficiently. This places iVDFM near structured state-space approaches in its numerical organization, even though its purpose is latent-variable identification rather than purely deterministic sequence modeling (Chang et al., 24 Mar 2026).

Earlier variational DFM work took rather different routes. Sparse dynamic factor models with loading selection use a structural mean-field family over stacked factors, factor dynamics, loadings, variances, and Bernoulli selectors, with closed-form updates and a modified Kalman smoother to handle missing data (Spånberg, 2022). Linear DFM identification through RMFD-E was estimated by EM in a Gaussian state-space model, where identification constraints were enforced directly in the M-step through selector matrices encoding structural zeros and ones (Koistinen et al., 2022). Relative to those approaches, iVDFM moves the identifiable object from loading matrices or static sparse structure to the innovation process itself.

5. Empirical behavior and use cases

The empirical evaluation of iVDFM covers synthetic factor recovery, intervention analysis, and probabilistic forecasting. On synthetic data, two data-generating processes are studied: a dynamic DGP in which latent factors follow AR dynamics and observations are nonlinear mixes of those factors, and a static DGP with conditionally independent latents and nonlinear mixing. With $y_t=(Z\circ\Lambda)F_t+\epsilon_t$4, $y_t=(Z\circ\Lambda)F_t+\epsilon_t$5, and $y_t=(Z\circ\Lambda)F_t+\epsilon_t$6, the reported metrics are MCC, subspace distance, smoothness, and Trace $y_t=(Z\circ\Lambda)F_t+\epsilon_t$7. On the dynamic DGP, iVDFM achieves the best or near-best MCC, subspace distance, and Trace $y_t=(Z\circ\Lambda)F_t+\epsilon_t$8, whereas the classical DFM performs worst. On the static DGP, the VAE leads on most metrics, and iVDFM is competitive but not best, which is consistent with the model being designed primarily for dynamic structure (Chang et al., 24 Mar 2026).

The intervention experiments address whether the learned representation supports model-based impulse-response reasoning. In synthetic structural causal models, the authors intervene on a shock component at time $y_t=(Z\circ\Lambda)F_t+\epsilon_t$9 and compare the resulting model-implied impulse responses with the true SCM responses. Across Base, Regime, and Chain SCM variants, the reported sign accuracy remains around $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$0 to $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$1, while IRF-MSE rises with structural complexity. The stated interpretation is that the learned innovations can serve as estimated shocks whose interventions yield reasonably consistent directional effects (Chang et al., 24 Mar 2026).

For real-world probabilistic forecasting, iVDFM is evaluated on ETTh1, ETTh2, ETTm1, ETTm2, and Weather, for forecast horizons $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$2, $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$3, $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$4, and $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$5, against baselines including iTransformer, TimeMixer, TimeXer, and Deep DFM. The overall pattern is not that iVDFM dominates forecasting metrics; rather, TimeMixer often obtains the best CRPS and MSE, while iVDFM remains in the same general performance range and offers identifiable, interpretable latent structure that the forecasting-specialized baselines do not target (Chang et al., 24 Mar 2026).

The broader literature provides related empirical support for the practical importance of identification. In the linear DFM setting, RMFD-E-based estimation of a monetary policy shock on a standard high-dimensional macroeconomic dataset produced strong and intuitive reactions, whereas principal-components-based static DFMs gave qualitatively counterintuitive results (Koistinen et al., 2022). In sparse variational DFMs, simulation experiments showed well identified sparsity patterns, precise loading and factor estimation, and robust handling of missing data (Spånberg, 2022). These results do not establish the same identifiability class as iVDFM, but they reinforce the substantive point that how factors are identified materially affects factor recovery, impulse responses, and downstream inference.

6. Relations, interpretation, and limitations

iVDFM should be distinguished from three adjacent model classes. First, it differs from classical DFMs by replacing Gaussian rotationally indeterminate latent shocks with non-Gaussian conditional exponential-family innovations and by using diagonal dynamics rather than generic VAR dynamics. Second, it differs from static iVAE by conditioning the innovation process rather than the latent state itself, thereby preserving a dynamic latent-factor interpretation (Chang et al., 24 Mar 2026). Third, it differs from earlier spatio-temporal iVAE formulations in which time and space enter primarily as auxiliary covariates modulating $$\mathbf{x}=\mathbf{f}(\mathbf{z})+\boldsymbol{\epsilon}$$6, without an explicit Markov or state-space prior over latent trajectories (Sipilä et al., 2024).

A second common misunderstanding is to equate sparsity or nonstationarity with identifiability in the strict theorem-proving sense. Sparse loading selection produces a posterior that concentrates on simple loading structures and often empties out superfluous factors, but sparsity alone does not guarantee formal uniqueness of the factor space (Spånberg, 2022). Likewise, nonstationarity in latent means, variances, or AR parameters is a powerful identifiability resource, but it only yields a theorem when combined with specific exponential-family structure, injective mixing, and sufficient rank variation in the natural-parameter map (Sipilä et al., 15 Sep 2025).

The principal limitations follow directly from the assumptions that make the theorem work. Identifiability is only as strong as the approximation of the theoretical regime: innovations must be genuinely non-Gaussian, the conditioning variables must vary enough to avoid a degenerate prior, the decoder must be injective on the relevant support, and the dynamics must remain diagonal or block-diagonal. These restrictions improve semantic stability of the learned factors, but they also reduce dynamic expressiveness by ruling out cross-factor state interactions (Chang et al., 24 Mar 2026).

Forecasting performance also illustrates the trade-off. iVDFM is competitive on standard benchmarks, but it is not presented as a state-of-the-art forecasting model in CRPS or MSE. Its comparative advantage lies in identifiable dynamic factors, stable intervention behavior on synthetic SCMs, and a representation whose equivalence class is much narrower than the arbitrary rotations admitted by Gaussian DFMs (Chang et al., 24 Mar 2026). This suggests that iVDFM is most appropriate when dynamic latent structure is meant to support interpretation, intervention analysis, or structural reasoning, rather than when forecasting accuracy is the sole objective.