Batched Robust iHMM for Streaming Time Series

• The paper introduces a batched robust iHMM that uses an HDP prior, WoLF emission updates, and batched latent-state inference to limit outlier influence.
• BR-iHMM is defined by a rigorous probabilistic framework with linear-Gaussian emissions and bounded generalized Bayes updates, ensuring stable regime discovery.
• Empirical results demonstrate that BR-iHMM significantly reduces forecasting errors and spurious state creation compared to conventional online HMM approaches.

Searching arXiv for the specified papers to ground the article in the cited work. Batched Robust iHMM (BR-iHMM) denotes an online infinite hidden Markov model for streaming, nonstationary time series in which robustness is enforced simultaneously in observation space and latent state space. In its explicit formulation, BR-iHMM combines a hierarchical Dirichlet process (HDP) prior over an unbounded regime set, robust generalized-Bayes emission updates, and batched latent-state inference that restricts regime changes to batch boundaries. The resulting model is “doubly outlier-robust” in the sense that both the posterior over emission parameters and the posterior over latent states have bounded posterior influence under arbitrary contamination, while retaining online prediction and interpretable regime discovery (Yiu et al., 15 Apr 2026).

1. Conceptual placement and lineage

The immediate substrate of BR-iHMM is the infinite hidden Markov model, or HDP-HMM, in which the hidden state space is countably infinite and the number of active states is inferred rather than fixed in advance. In the canonical construction, a shared global measure over states is drawn by stick breaking and each transition row is a Dirichlet-process draw from that shared base, yielding an HMM whose complexity is controlled by the data and the hyperparameters rather than by a preset state count (Sgouralis et al., 2016).

The term BR-iHMM appears explicitly in the online robust formulation of “Doubly Outlier-Robust Online Infinite Hidden Markov Model” (Yiu et al., 15 Apr 2026). Several earlier strands supplied constituent ideas later associated with batched and robust iHMM design. ICON couples an iHMM to a continuous drift process so that data with drift from one or many traces can be analyzed jointly without introducing artifact states (Sgouralis et al., 2016). Prototype-defined RBF-iHMMs replace linear emissions by nonlinear autoregressive radial-basis-function emissions and emphasize few-shot segmentation of non-stationary sequences (Qarout et al., 2021). Restricted Collapsed Draw samplers make simultaneous restricted draws feasible in collapsed HDP-HMM representations, enabling blocked, beam, and split–merge updates (Makino et al., 2011).

A persistent source of ambiguity is that the acronym “iHMM” is not uniform across subfields. In safe runtime monitoring, “iHMM” can denote an interval Hidden Markov Model with lower and upper bounds on initial and transition probabilities (Skurka et al., 16 Feb 2026). In imprecise-probability work, “iHMM” denotes an imprecise hidden Markov model based on coherent lower previsions and epistemic irrelevance (Bock et al., 2012). BR-iHMM, in the sense used here, refers to the infinite-state HDP-HMM lineage rather than those interval or imprecise formulations.

2. Probabilistic structure

In the published BR-iHMM, the latent regime sequence $s_t \in \{1,2,\ldots\}$ follows an HDP prior. The global weights are drawn by stick breaking,

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$

and each state-specific transition distribution is

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$

At finite time $t$, the truncated global vector is written $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$ (Yiu et al., 15 Apr 2026).

The emission model is linear-Gaussian with exogenous covariates. Observations satisfy

$$P(y_t \mid x_t, s_t, \theta_{1:t}) = \mathcal{N}\!\bigl(y_t \mid F_t\,\theta_{s_t},\; R_t\bigr),$$

where $y_t \in \mathbb{R}^d$, $F_t = f(x_t)\in \mathbb{R}^{d\times m}$, regime parameters $\theta_\ell \in \mathbb{R}^m$, and $R_t$ is a known observation covariance. Each $(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$0 has a Gaussian prior $(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$1. Conditional on a state path, the posterior factorizes as

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$2

with sufficient statistics $(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$3 (Yiu et al., 15 Apr 2026).

In the non-robust online iHMM, the active state is updated by Kalman-style recursions,

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$4

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$5

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$6

The predictive transition law obtained from Dirichlet–multinomial conjugacy is

$(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$7

where $(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$8 is the transition count matrix and $(\beta_1,\beta_2,\ldots) \sim \mathrm{SB}(\gamma),$9 are structural HDP sufficient statistics (Yiu et al., 15 Apr 2026).

This linear-Gaussian specification is not the only possible emission family within the broader iHMM literature. A nearby nonlinear variant replaces emissions by order-$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$0 autoregressive RBF networks with state-specific prototype-defined centers, making the emission map

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$1

for segmentation of complex non-stationary signals (Qarout et al., 2021). This suggests that BR-iHMM is best understood as a robust state-inference scheme attached to an HDP-HMM core, rather than as a single fixed emission architecture.

3. Doubly outlier-robust mechanism

The central claim of BR-iHMM is that robustness must hold for both emission-parameter learning and latent-state inference. The paper formalizes this by the Posterior Influence Function (PIF). For a contaminated observation $$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$2, the state-path PIF is

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$3

and the emission-parameter PIF is

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$4

The joint PIF decomposes as

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$5

Hence bounded joint influence requires bounded influence in both spaces (Yiu et al., 15 Apr 2026).

The paper proves that observation-only robustness is insufficient. Even if emission parameters are updated by a robust generalized-Bayes rule with bounded $$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$6, the state PIF remains unbounded in a standard online iHMM, because an arbitrarily large outlier can still force posterior mass toward a newly created state. This yields spurious regimes and unstable switching (Yiu et al., 15 Apr 2026).

BR-iHMM addresses the emission side by adopting the Weighted Observation Likelihood Filter (WoLF). The ordinary likelihood contribution is replaced by a weighted likelihood

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$7

with inverse multiquadratic weight

$$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$8

where $$\boldsymbol\pi_\ell \sim \mathrm{Dir}(\alpha\,\boldsymbol\beta_t), \qquad P(s_t = k \mid s_{t-1} = \ell) = \pi_{\ell,k}.$$9 is a robustness scale. Under the required boundedness conditions,

$t$0

large residuals produce very small weights and therefore very cautious parameter updates. Operationally this modifies the predictive covariance to

$t$1

This robust generalized-Bayes perspective aligns with broader generalized filtering work in which standard likelihood factors are replaced by bounded-influence potentials under misspecification (Boustati et al., 2020).

State-space robustness is enforced by batching. For a batch size $t$2, BR-iHMM evaluates candidate batchwise regime assignments through

$t$3

$t$4

so the state is held constant within the batch. The no-switching constraint is encoded by a degenerate sticky HDP prior,

$t$5

Thus regime changes are permitted only at batch boundaries (Yiu et al., 15 Apr 2026).

The two tuning parameters are therefore $t$6 and $t$7. Larger $t$8 pools more evidence before a switch and increases robustness to transient anomalies, but induces adaptation lag for genuine regime changes. Smaller $t$9 makes residuals “large” sooner and downweights them more aggressively, increasing observation-space robustness at the cost of conservatism (Yiu et al., 15 Apr 2026).

4. Inference and batching regimes

In its published implementation, BR-iHMM is learned online by Particle Learning. Each particle carries a sampled state trajectory $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$0, HDP structural statistics $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$1, emission statistics $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$2, and a weight $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$3. For each batch, particles compute multi-step predictives for candidate states, update particle weights from batch likelihood scores, resample when ESS falls below threshold, sample a boundary state from the batch score, set intra-batch states equal to that sampled state, update HDP counts and hyperparameters, perform WoLF emission updates inside the batch, and prune rarely used stale states when the active-state cap MAX_STATES is exceeded (Yiu et al., 15 Apr 2026).

The batch score used for a particle and candidate state $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$4 is

$$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$5

after which the particle samples $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$6 and sets

$$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$7

This converts a time-pointwise online iHMM into a genuinely batched online filter (Yiu et al., 15 Apr 2026).

Other iHMM literatures instantiate batching differently. ICON is multi-trace rather than boundary-batched: multiple traces share a hidden state sequence while each channel has its own emission parameters and its own drift process, allowing all traces to be treated on an equal footing without manual detrending or trace selection (Sgouralis et al., 2016). Prototype-defined RBF-iHMMs admit a natural batched extension across multiple sequences by running forward–backward steps per sequence in parallel and aggregating transition counts and emission sufficient statistics, although the published method itself uses batch Gibbs sampling rather than stochastic variational inference (Qarout et al., 2021). In collapsed HCRP-HMMs, the Restricted Collapsed Draw sampler provides exact simultaneous restricted draws for blocked Gibbs, beam sampling, and split–merge moves, which is relevant when “batched” refers to large coupled latent-state updates rather than to online mini-batches (Makino et al., 2011).

These variants indicate that batching in the iHMM literature has at least three distinct meanings: non-overlapping time batches in online filtering, multiple traces analyzed jointly under shared latent structure, and blocked or simultaneous MCMC updates of many latent variables. BR-iHMM, in the strict sense of (Yiu et al., 15 Apr 2026), uses the first of these.

5. Empirical behavior

Across limit order book data, hourly electricity demand, and a synthetic high-dimensional linear system, BR-iHMM reduces one-step-ahead forecasting error by up to 67% relative to competing online Bayesian methods (Yiu et al., 15 Apr 2026). The reported empirical behavior is not uniform across domains; it depends on whether batching helps suppress transient contamination without masking genuine fast regime changes.

Setting	BR-iHMM observation	Comparative note
Synthetic high-dimensional linear system	RMSE 46.1; near-correct number of regimes recovered	WoLF-only 103.8; standard online iHMM 101.7; BOCD 123.1; offline beam 2.9
Hourly electricity demand	RMSE 0.47; lowest rolling RMSE during volatile periods	WoLF-only 0.63; standard online iHMM 0.57; BOCD 0.80
Limit order book OFI	Hyperparameter optimization chose $$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$8; RMSE 0.616	Robust batching not helpful; BR-iHMM effectively reduced to WoLF-only
Well-log segmentation	More resistant to spikes; fewer artefact states; high PPV/TPR with moderate detection delay	Student-$$\boldsymbol\beta_t = (\beta_1,\dots,\beta_{t+1}) \in \Delta^t$$9 iHMM still exhibited outlier-driven state creation

In the synthetic linear-regression experiment, non-batched online methods created more than 30 spurious regimes and exhibited rapid switching, whereas BR-iHMM stabilized quickly and was particularly helpful in the high-dimensional setting. In hourly electricity demand, BR-iHMM detected transitions around March 2020, while the non-robust online iHMM tended to remain in a single state. In order-flow imbalance, the optimum $$P(y_t \mid x_t, s_t, \theta_{1:t}) = \mathcal{N}\!\bigl(y_t \mid F_t\,\theta_{s_t},\; R_t\bigr),$$0 shows that the method does not impose batching when the data favor rapid adaptation. In well-log segmentation, the method produced interpretable segments aligned with geological layers and resisted spike-driven over-segmentation (Yiu et al., 15 Apr 2026).

Adjacent robust iHMM families report analogous phenomena under different mechanisms. RBF-iHMM achieved more than 80% balanced accuracy with only 5% of the EEG training data, while a linear AR-iHMM remained near random and a VAE-LSTM required about 40% of the data to match it; on synthetic switching nonlinear data, the same model reached latent-state label accuracy 0.95 with transition-matrix MSE $$P(y_t \mid x_t, s_t, \theta_{1:t}) = \mathcal{N}\!\bigl(y_t \mid F_t\,\theta_{s_t},\; R_t\bigr),$$1 (Qarout et al., 2021). ICON showed that with correct drift modeling the posterior over state number peaks sharply at the true value, whereas a plain iHMM shifts toward higher state counts and introduces artifact states (Sgouralis et al., 2016). These results support the broader view that robust emission modeling and robust nuisance-process handling materially affect regime recovery.

6. Limitations, neighboring formulations, and scope

The published BR-iHMM comes with explicit trade-offs and limits. Robust batching induces detection delay for change points of up to roughly one batch. The current theory is proved for linear-Gaussian emissions with bounded WoLF weights. Batch size $$P(y_t \mid x_t, s_t, \theta_{1:t}) = \mathcal{N}\!\bigl(y_t \mid F_t\,\theta_{s_t},\; R_t\bigr),$$2 is fixed rather than adaptive, no online adaptive batching is provided, and the principal theoretical guarantees are bounded-influence results rather than full asymptotic consistency statements. Computationally, the method also relies on a heuristic active-state cap and pruning scheme, even though the underlying prior is nonparametric (Yiu et al., 15 Apr 2026).

Outside that specific formulation, “BR-iHMM” is also used descriptively for a broader design pattern. In the prototype-defined RBF-iHMM work, a batched robust iHMM would typically mean inference over many sequences jointly and/or in mini-batches, together with robustness mechanisms such as heavy-tailed emission noise, shrinkage priors on RBF weights, and hierarchical prototype pooling across batches or patient subgroups (Qarout et al., 2021). ICON supplies another nearby template: an iHMM plus an additive continuous control process for drift, shared states across traces, and separate per-trace drifts, thereby avoiding artifact states caused by nuisance trends (Sgouralis et al., 2016). These are not the same algorithm as BR-iHMM in (Yiu et al., 15 Apr 2026), but they occupy the same design space.

Two further neighboring formulations underscore the importance of terminology. Interval Hidden Markov Models for runtime monitoring use interval-valued initial and transition probabilities and compute cautious risk bounds by maximizing over all HMMs consistent with those intervals; robustness there means worst-case monitoring under model uncertainty with convergence of interval refinements (Skurka et al., 16 Feb 2026). Imprecise hidden Markov models based on coherent lower previsions instead return Walley–Sen maximal state sequences rather than a single MAP path, robustifying decoding by explicitly preserving ambiguity when a unique sequence is not credally justified (Bock et al., 2012). A plausible implication is that BR-iHMM should be read as one member of a larger robust-sequential-modeling family in which “robustness” may refer to bounded posterior influence, interval conservatism, nuisance-process separation, or credal set-valued decoding, depending on the modeling objective.