acceptodds
Under review as a conference paper at ICLR 2027

RegimeUnfold: Continual Regime-Space Learning for Deep Hidden Markov Process Forecasting

Abstract

Deep hidden Markov process forecasting aims to predict future observations while explicitly identifying the latent regimes governing their dynamics. Existing approaches typically combine deep learning with Hidden Markov Models (HMMs) or their variants, seeking to retain the strong nonlinear predictive capability of deep models while preserving explicit latent regime representations. However, these methods generally operate over a fixed regime set, making their performance sensitive to a prespecified regime-space size and limiting adaptation after deployment. To address this limitation, we propose RegimeUnfold, a continual regime-space learning framework with offline regime discovery and online regime evolution for deep hidden Markov process forecasting. During the offline stage, RegimeUnfold jointly infers discrete regime assignments and learns predictive dynamics while progressively constructing and refining the regime space without predefining its size. During the online stage, it performs hard regime decoding and uses predictive-insufficiency and structural-novelty evidence to retain applicable regimes, update drifting regimes, or expand the regime space under stability constraints. Experiments on five synthetic and four real-world datasets demonstrate accurate forecasting, reliable regime inference, and effective adaptation to evolving predictive dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
Deep Hidden Markov Process Forecasting, Latent Regime Inference, Continual Regime-Space Learning, Regime-Aware Forecasting, Online Regime Evolution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EK1oOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RegimeUnfold, a continual regime-space learning framework for deep hidden Markov process forecasting that does not assume a fixed regime set. Its offline regime discovery stage constructs the regime space without predefining its size. Its online regime evolution stage updates existing regimes or expands the regime space under stability constraints. Experiments on five synthetic and four real-world datasets are reported to show accurate forecasting, reliable regime inference, and effective adaptation.

Critical strengths
  1. Well-motivated problem. The paper targets a practical limitation of existing DL-HMM hybrids that assume a fixed regime space specified before training.
  2. Clear benefit when unseen dynamics emerge. On the Progressive dataset the full model raises regime inference accuracy from 51.56% to 91.49% over RegimeUnfold w/o Evolution.
  3. Detailed apendix with techinical details is provided.
Critical weaknesses
  1. Inconsistent baseline reporting. Tables 1 and 2 mix Adaptive and Non-adaptive rows across baselines and omit the Large-window variants, which overstates the gains reported in the main text.
  2. Regime-number sensitivity is shifted to thresholds. The method still requires a maximum allowed regime-space size along with many thresholds and loss weights, yet neither their values nor a sensitivity analysis is provided.
  3. Key claims and components are not directly validated. No experiment isolates the benefit of hard regime decoding or the ability to update drifting regimes, and the two ablations cannot attribute gains to individual components.
  4. Experimental setting does not match the stated motivation. Linear AR(1) synthetic data and 3- to 5-dimensional real-world datasets cannot test the claimed advantage on complex nonlinear predictive dependencies in high-dimensional multivariate time series, and online regime evolution is only shown qualitatively.
  5. Limited discussion of novelty and related work. The offline split and merge procedure closely resembles the birth and merge moves of [1] without discussion, and comparisons with deep switching dynamical systems and online forecasting methods are missing.

[1] Scalable adaptation of state complexity for nonparametric hidden Markov models.

Questions for authors

See weakness

8 comments

Thank you very much for your valuable suggestions. We have provided point-by-point responses to address all the raised questions and concerns.

Weakness 1: Inconsistent baseline reporting. Tables 1 and 2 mix Adaptive and Non-adaptive rows across baselines and omit the Large-window variants, which overstates the gains reported in the main text.

Response to Weakness 1: Thank you for the suggestion. Due to space limitations, the original Tables 1 and 2 provided a compact summary of representative baseline results, while the complete results were reported in Appendix F. In addition, the Large-window variants use substantially longer historical inputs than the default configurations, making the comparisons less directly comparable under the same input settings. Therefore, these variants were initially presented only in the appendix.

We understand that a more complete presentation would help clarify the differences between experimental settings. To address this concern, we have revised Tables 1 and 2 to explicitly distinguish Adaptive and Non-adaptive settings and include the Large-window variants. We have also revised the corresponding comparisons and conclusions in the main text to provide a more comprehensive and fair evaluation.

Weakness 2: Regime-number sensitivity is shifted to thresholds. The method still requires a maximum allowed regime-space size along with many thresholds and loss weights, yet neither their values nor a sensitivity analysis is provided.

Response to Weakness 2: Thank you for the comment. The key distinction is that a fixed regime number KK directly determines the size of the regime space, whereas the maximum allowed regime number Kmax⁡K_{\max} only limits the search capacity without prescribing the final number of regimes. RegimeUnfold instead determines the regime-space size through evidence-based structural adaptation. During offline discovery, structural changes are evaluated based on predictive improvement, structural reliability, and regime-transition evidence. During online evolution, UPDATE and CREATE decisions are evaluated based on accumulated evolution evidence, predictive improvement, and stability requirements. Unlike a fixed regime number, these criteria allow the regime space to adapt when previously unseen dynamics emerge. In our experiments, Kmax⁡K_{\max} was set above the number of regimes ultimately identified, and the limit was never reached.

To further examine parameter sensitivity, we conducted additional experiments covering eleven hyperparameters or groups of related parameters for offline regime discovery and four parameters for online regime evolution. The results show relatively limited sensitivity to the tested structural selection settings. On Fixed-K3, varying the split acceptance threshold γsplit\gamma_{\mathrm{split}} from 0.015 to 0.06 yields mean regime inference accuracies between 99.07% and 99.11%. Similarly, varying Kmax⁡K_{\max} over {4,8,12}\{4,8,12\} yields accuracies of 97.32%, 99.09%, and 98.99%, respectively. These results indicate that regime discovery performance remains relatively stable across the examined ranges of these two structural controls, although other hyperparameters have a greater influence.

The complete sensitivity analysis, including loss weights, structural selection thresholds, and online adaptation parameters, is presented in our response to Reviewer Ay5y, Question 3, with the corresponding experimental results added to Appendix G of the revised manuscript. The relevant parameter settings are provided in Appendix F, while the detailed offline and online procedures are described in Appendices A and B, respectively.

Weakness 3: Key claims and components are not directly validated. No experiment isolates the benefit of hard regime decoding or the ability to update drifting regimes, and the two ablations cannot attribute gains to individual components.

Response to Weakness 3: Thank you for the comment. To further validate the key mechanisms of RegimeUnfold, we conducted additional component-wise ablation studies covering hard regime decoding, adaptation of established regimes, and stability-aware control. Specifically, we consider the following ablations:

  • Soft Forecast Aggregation: Replaces hard single-expert forecasting with a probability-weighted average of regime-specific predictions. Both approaches retain argmax-based regime identification, allowing us to examine the contribution of hard regime decoding to forecasting.
  • w/o UPDATE: Prevents the incorporation of updates to established regimes to evaluate the contribution of adapting existing predictive dynamics.
  • w/o Established-Regime Stability Checks at Admission: Removes the protection checks for established regimes during candidate acceptance to evaluate their role in preserving previously learned dynamics.

We evaluate these ablations on Fixed-K2, Fixed-K3, and Progressive. The results are reported below.

Table 1: Effects of hard regime decoding and online adaptation mechanisms on Fixed-K2.

ModelAccuracy (%)F1 (%)MAEMSE
Soft Forecast Aggregation95.78 ± 2.6096.88 ± 2.270.0094 ± 0.00020.00022 ± 0.00001
w/o UPDATE95.53 ± 2.4596.84 ± 2.250.0092 ± 0.00020.00021 ± 0.00001
w/o Established-Regime Stability Checks at Admission95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
RegimeUnfold97.85 ± 2.2297.68 ± 2.080.0093 ± 0.00020.00022 ± 0.00001

Table 2: Effects of hard regime decoding and online adaptation mechanisms on Fixed-K3.

ModelAccuracy (%)F1 (%)MAEMSE
Soft Forecast Aggregation95.23 ± 1.1295.32 ± 0.480.0098 ± 0.00020.00041 ± 0.00003
w/o UPDATE94.71 ± 1.3795.16 ± 0.470.0097 ± 0.00020.00041 ± 0.00003
w/o Established-Regime Stability Checks at Admission95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
RegimeUnfold96.31 ± 0.2395.88 ± 0.280.0093 ± 0.00010.00026 ± 0.00001

Table 3: Effects of hard regime decoding and online adaptation mechanisms on Progressive.

ModelAccuracy (%)F1 (%)MAEMSE
Soft Forecast Aggregation86.13 ± 4.0186.23 ± 5.490.0146 ± 0.00100.00089 ± 0.00008
w/o UPDATE86.90 ± 5.0786.87 ± 6.260.0145 ± 0.00130.00094 ± 0.00011
w/o Established-Regime Stability Checks at Admission66.25 ± 16.7955.10 ± 25.170.0260 ± 0.01100.00243 ± 0.00157
RegimeUnfold91.49 ± 0.6191.19 ± 0.770.0134 ± 0.00000.00081 ± 0.00001

The full RegimeUnfold achieves higher regime inference accuracy than Soft Forecast Aggregation across all three datasets, while maintaining lower or comparable forecasting errors. On Progressive, replacing hard single-expert forecasting with soft aggregation reduces accuracy from 91.49% to 86.13% and increases MSE from 0.00081 to 0.00089. These results support the effectiveness of hard regime decoding within our regime-aware forecasting framework.

Disabling UPDATE also reduces regime inference accuracy across all three datasets. On Progressive, accuracy decreases from 91.49% to 86.90%, while MSE increases from 0.00081 to 0.00094, supporting the contribution of adapting established regimes. Furthermore, removing established-regime stability checks substantially degrades performance on Progressive, reducing accuracy to 66.25% and increasing MSE to 0.00243. These results demonstrate the importance of stability-aware control during online regime evolution.

The complete component-wise ablation studies, including additional training objectives, regime decoding mechanisms, regime selection components, and evolution assessment criteria, are provided in our response to Reviewer Ay5y, Weakness 1, and in Appendix G of the revised manuscript.

Weakness 4: Experimental setting does not match the stated motivation. Linear AR(1) synthetic data and 3- to 5-dimensional real-world datasets cannot test the claimed advantage on complex nonlinear predictive dependencies in high-dimensional multivariate time series, and online regime evolution is only shown qualitatively.

Response to Weakness 4: Thank you for the comment. To further evaluate RegimeUnfold under more complex predictive dynamics, we constructed an additional 64-dimensional nonlinear synthetic dataset with an evolving latent regime space.

The dataset contains 20,000 time steps and 64 dynamically coupled variables. Unlike the original linear autoregressive settings, the observations are generated using nonlinear cross-variable interactions and regime-dependent dynamics with four-step historical dependencies. Specifically, let yt∈R64\mathbf{y}_t\in\mathbb{R}^{64} denote the observation at time tt, and let ztz_t denote its latent regime. We first construct nonlinear cross-variable features:

yˉt=13∑h=24yt−h,ϕt=tanh⁡(2.5Pyt−1),\bar{\mathbf y}_t=\frac{1}{3}\sum_{h=2}^{4}\mathbf y_{t-h},\qquad\boldsymbol{\phi}_t=\tanh(2.5P\mathbf y_{t-1}),

where P∈R32×64P\in\mathbb{R}^{32\times64} is a sparse projection matrix with four nonzero entries per row, producing 32 nonlinear features from the observed variables. Each observation dimension jj is then generated as

y_{t,j}={}&a_{z_t}y_{t-1,j} +0.10\bar y_{t,j}\\ &+0.45s_j^{(0)}\phi_{t,c_j}\\ &+0.30s_j^{(z_t)}\phi_{t,r_{z_t,j}}\\ &+0.05\phi_{t,u_j}\phi_{t,v_j} \tanh(2.5\bar y_{t,w_j}) +\epsilon_{t,j}, \end{aligned}$$ where $a_{z_t}\in\{-0.45,0.45\}$ is a regime-dependent autoregressive coefficient, $s_j^{(0)}$ and $s_j^{(z_t)}$ are shared and regime-specific sign coefficients, and $c_j$, $r_{z_t,j}$, $u_j$, $v_j$, and $w_j$ specify cross-variable feature connections. The regime-specific connections vary across regimes, while $\epsilon_{t,j}\sim\mathcal{N}(0,0.12^2)$ denotes independent Gaussian noise. This construction combines shared nonlinear dynamics, regime-specific predictive mechanisms, and multiplicative nonlinear interactions. All 64 variables participate in the conditional dynamics rather than serving as additional noise dimensions. The dataset also retains the progressive regime evolution setting, with the number of available regimes increasing from 2 to 3 and then 4. New regimes are introduced at 50% and 70% of the sequence, respectively. Regime transitions follow second-order dependencies with state-dependent duration distributions. **Table 4: Regime inference and forecasting results on the 64-dimensional nonlinear Progressive dataset.** | Model | Accuracy (%) | F1 (%) | MAE | MSE | |---|---|---|---|---| | RegimeUnfold | 89.51 ± 2.94 | 84.01 ± 3.92 | 0.1564 ± 0.0117 | 0.04999 ± 0.00959 | RegimeUnfold achieves 89.51% regime inference accuracy and 84.01% F1, with MAE of 0.1564 and MSE of 0.04999, averaged over five runs. These results provide additional evidence that RegimeUnfold can discover evolving latent regimes and perform forecasting under higher-dimensional nonlinear predictive dynamics. Furthermore, we conducted additional quantitative ablation studies on UPDATE, CREATE, and stability-aware control, as detailed in our response to **Reviewer NbVW, Weakness 2** and **Appendix G** of the revised manuscript. These experiments further validate the contributions of the online regime evolution mechanisms.
Reviewer_NbVWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RegimeUnfold is designed for multivariate time-series forecasting with unknown latent regimes that may change after deployment. It combines a shared neural predictor with regime-specific residual experts. During offline training, it jointly learns regime assignments and predictive dynamics, using split, merge, and refinement operations to identify the regime structure. During online prediction, a causal router selects the regime to use. Prediction errors and structural novelty signals prompt the method to compare updating an existing regime with creating a new one. Experiments on five synthetic and four real-world datasets show improvements in forecasting and regime identification.

Critical strengths

The paper proposes a method for discovering an initially unknown regime structure and adapting it after deployment. This could be useful for forecasting systems that need to handle changing dynamics while maintaining explicit representations of different regimes.

Critical weaknesses
  1. The method combines prediction error, prototype separation, and temporal regularization, but it is unclear under what assumptions the learned partitions and number of regimes correspond to the states that generate the data. Accurate forecasts and well-separated prototypes are not sufficient to establish this correspondence. A flexible predictor may model several regimes together, while multiple experts may divide a single mechanism into separate groups. Likewise, choosing CREATE over UPDATE based on predictive performance does not necessarily mean that a new regime has emerged. The synthetic state-recovery results are encouraging, but they do not clearly show when recovery is possible or when these ambiguities occur. This issue matters because regime discovery and evolution are central contributions of the paper. The authors should clarify whether the goal is to learn useful predictive partitions or recover the underlying regimes. If the latter is intended, they should provide supporting assumptions, analysis, or targeted experiments.

  2. Although the experiments include adaptive baselines, the ablations remove the entire evolution pipeline without isolating the effects of parameter adaptation, regime creation, and stability control. The paper also does not directly evaluate whether the method correctly decides when to update an existing regime or create a new one, or whether it preserves previously learned regimes. The overall performance gains therefore support the effectiveness of the full system, but do not clearly demonstrate that the proposed evolution mechanisms provide the claimed capabilities.

Questions for authors
  1. Could you describe the online evolution procedure in more detail? In particular, how is structural discrepancy measured, how is evidence accumulated and decayed over time, and how are UPDATE and CREATE proposals accepted or rejected? How are the corresponding thresholds calibrated?
  2. Could you clarify the chronological evaluation protocol and whether the online evaluation includes genuinely unseen regimes? Please specify the offline training, development, calibration, and online evaluation intervals, and confirm whether newly introduced regimes are excluded from all offline learning stages.
  3. Why is prediction-error minimization preferred over likelihood-based learning in this setting? Are there controlled experiments showing that the training objective contributes to the reported gains independently of the architecture, routing, and adaptation procedures?
  4. What advantages does RegimeUnfold offer over a simpler neural-expert model that can add experts over time? Which proposed mechanisms account for these advantages? A comparison using the same backbone and comparable model capacity, adaptation budgets, and memory budgets would help establish the value of these mechanisms, particularly for distinguishing gradual drift from the emergence of new regimes and maintaining performance when previously seen regimes recur.
7 comments

Thank you very much for your valuable suggestions. We have provided point-by-point responses to address all the raised questions and concerns.

Weakness 1: The method combines prediction error, prototype separation, and temporal regularization, but it is unclear under what assumptions the learned partitions and number of regimes correspond to the states that generate the data. Accurate forecasts and well-separated prototypes are not sufficient to establish this correspondence. A flexible predictor may model several regimes together, while multiple experts may divide a single mechanism into separate groups. Likewise, choosing CREATE over UPDATE based on predictive performance does not necessarily mean that a new regime has emerged. The synthetic state-recovery results are encouraging, but they do not clearly show when recovery is possible or when these ambiguities occur. This issue matters because regime discovery and evolution are central contributions of the paper. The authors should clarify whether the goal is to learn useful predictive partitions or recover the underlying regimes. If the latter is intended, they should provide supporting assumptions, analysis, or targeted experiments.

Response to Weakness 1: Thank you for asking us to clarify the learning objective. Our primary objective is to learn reusable, prediction-oriented regimes that support forecasting and continual adaptation. Each learned regime is associated with a regime-conditioned predictor, while temporal regularization encourages coherent regime assignments over time.

Accurate forecasting alone does not guarantee a one-to-one correspondence between learned and underlying regimes or recovery of the true regime number. Underlying regimes with similar predictive dynamics may be merged into one learned regime, while a single underlying regime may be represented by multiple learned regimes. To mitigate these ambiguities, RegimeUnfold jointly considers predictive improvement, structural reliability, and segment-level support when evaluating regime-space adaptations, rather than relying solely on prediction errors or prototype separation.

Similarly, CREATE adds a learned regime when the candidate satisfies our predictive-improvement, stability, and validity criteria, including persistent support and sufficient distinction from existing regimes. Its acceptance does not by itself establish that a new underlying generative state has appeared.

To further establish the theoretical foundations of regime recovery, we have added a theoretical analysis to the revised manuscript. First, we show that different underlying regimes can be distinguished through their predictive dynamics. Specifically, under squared prediction loss, the expected excess error of using regime jj when the true regime is ii satisfies

E[∥Y−μj,t∥22−∥Y−μi,t∥22∣Ft,Z=i]=∥μi,t−μj,t∥22.\mathbb{E}[\Vert Y-\boldsymbol{\mu}_{j,t}\Vert_2^2- \Vert Y-\boldsymbol{\mu}_{i,t} \Vert_2^2\mid\mathcal{F}_t,Z=i]=\Vert\boldsymbol{\mu}_{i,t}-\boldsymbol{\mu}_{j,t}\Vert_2^2.

where μi,t\boldsymbol{\mu}_{i,t} and μj,t\boldsymbol{\mu}_{j,t} denote the corresponding regime-conditional predictive means. This result establishes a direct connection between distinct predictive dynamics and regime differentiation.

Building on this property, we further analyze regime recovery through offline decoding and cross-KK structural selection. Under sufficient predictive separability, accurate regime-conditioned prediction, controlled assignment uncertainty, and appropriate structural selection conditions, we establish that the learned regime assignments and regime number can converge to their true underlying counterparts:

dH(z^,z∗)N→p0,\frac{d_H(\widehat{\mathbf{z}},\mathbf{z}^*)}{N}\xrightarrow{p}0,

Pr⁡(K^=K∗)⟶1.\Pr(\widehat K=K^*)\longrightarrow1.

where dHd_H measures regime-assignment errors after label alignment, and K∗K^* is the true regime number. These results provide theoretical support for recovering the underlying regime structure through prediction-oriented regime learning.

Finally, our synthetic datasets are constructed with distinct regime-dependent autoregressive dynamics and known ground-truth regime labels, providing controlled settings with the predictive separability examined in our theoretical analysis. These datasets allow us to directly evaluate regime recovery and examine whether newly emerging regimes are identified as the sequence evolves. The results show that the learned regime assignments closely match the ground-truth labels, achieving regime accuracies of 97.85%, 96.31%, and 91.49% on Fixed-K2, Fixed-K3, and Progressive, respectively. The Progressive results further demonstrate the identification and subsequent reuse of newly introduced regimes.

The complete theoretical assumptions, propositions, and proofs have been added to Appendix C of the revised manuscript.

Weakness 2: Although the experiments include adaptive baselines, the ablations remove the entire evolution pipeline without isolating the effects of parameter adaptation, regime creation, and stability control. The paper also does not directly evaluate whether the method correctly decides when to update an existing regime or create a new one, or whether it preserves previously learned regimes. The overall performance gains therefore support the effectiveness of the full system, but do not clearly demonstrate that the proposed evolution mechanisms provide the claimed capabilities.

Response to Weakness 2: Thank you for the comment. To provide theoretical support for the proposed online regime evolution mechanism, we have added analyses of evolution-evidence persistence, validation-based candidate selection, and established-regime stability.

First, we prove that requiring persistent structural-novelty evidence reduces false activations caused by transient deviations. Under the assumption that the conditional probability of a pointwise false activation is bounded by pp, the probability of persistent false activation satisfies

Pr⁡(PersistentNoveltyt)≤min⁡p,2p2.\Pr(\mathrm{PersistentNovelty}_t)\leq\min\\{p,2p^2\\}.

Second, we establish theoretical properties of validation-based UPDATE and CREATE selection. Under uniform validation-error bounds, the selected candidate achieves expected predictive performance close to the best admissible alternative. Moreover, when the established-regime stability constraint is satisfied, the expected prediction loss for each previously learned regime satisfies

\leq (1+\epsilon_{\mathrm{old}})R_{\mathrm{old},k}(a_0) +\Delta_{\mathrm{abs}} +(2+\epsilon_{\mathrm{old}})\beta_{\mathrm{old},k}.$$ These results provide theoretical support for reliable evolution assessment and stability-aware adaptation.

To further validate the individual contributions of the proposed mechanisms, we have also conducted component-wise ablation studies that separately examine regime adaptation, regime creation, and stability-aware control, rather than removing the entire evolution pipeline.

Specifically, we consider the following ablations:

  • w/o UPDATE: Prevents the incorporation of updates to established regimes to evaluate the contribution of adapting existing predictive dynamics.
  • w/o CREATE: Prevents permanent regime-space expansion to assess the importance of incorporating emerging regimes.
  • w/o Established-Regime Stability Checks at Admission: Removes established-regime protection checks during candidate acceptance to evaluate their role in preserving previously learned dynamics.
  • w/o CREATE Validity Requirement: Removes the additional persistence and structural distinction requirements for CREATE to evaluate their role in preventing unnecessary regime creation.
  • w/o Online Evidence Persistence: Removes evidence persistence requirements to evaluate their role in filtering transient changes before adaptation.

We evaluate these ablations on Fixed-K2, Fixed-K3, and Progressive to examine their effects under different regime evolution settings. The results are reported below.

Table 1: Effects of online regime evolution mechanisms on Fixed-K2.

ModelAccuracy (%)F1 (%)MAEMSE
w/o UPDATE95.53 ± 2.4596.84 ± 2.250.0092 ± 0.00020.00021 ± 0.00001
w/o CREATE95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
w/o Established-Regime Stability Checks at Admission95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
w/o CREATE Validity Requirement95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
w/o Online Evidence Persistence95.29 ± 2.5196.67 ± 2.230.0092 ± 0.00020.00022 ± 0.00001
RegimeUnfold97.85 ± 2.2297.68 ± 2.080.0093 ± 0.00020.00022 ± 0.00001

Table 2: Effects of online regime evolution mechanisms on Fixed-K3.

ModelAccuracy (%)F1 (%)MAEMSE
w/o UPDATE94.71 ± 1.3795.16 ± 0.470.0097 ± 0.00020.00041 ± 0.00003
w/o CREATE95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o Established-Regime Stability Checks at Admission95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o CREATE Validity Requirement95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o Online Evidence Persistence93.94 ± 1.4394.76 ± 0.580.0098 ± 0.00010.00046 ± 0.00002
RegimeUnfold96.31 ± 0.2395.88 ± 0.280.0093 ± 0.00010.00026 ± 0.00001

Table 3: Effects of online regime evolution mechanisms on Progressive.

ModelAccuracy (%)F1 (%)MAEMSE
w/o UPDATE86.90 ± 5.0786.87 ± 6.260.0145 ± 0.00130.00094 ± 0.00011
w/o CREATE81.01 ± 1.6581.74 ± 2.720.0151 ± 0.00030.00095 ± 0.00004
w/o Established-Regime Stability Checks at Admission66.25 ± 16.7955.10 ± 25.170.0260 ± 0.01100.00243 ± 0.00157
w/o CREATE Validity Requirement86.43 ± 3.5986.32 ± 5.360.0148 ± 0.00110.00096 ± 0.00008
w/o Online Evidence Persistence90.08 ± 3.9091.76 ± 4.680.0137 ± 0.00060.00087 ± 0.00003
RegimeUnfold91.49 ± 0.6191.19 ± 0.770.0134 ± 0.00000.00081 ± 0.00001

The full RegimeUnfold achieves the highest regime inference accuracy across all three datasets while maintaining strong forecasting performance. Disabling UPDATE or CREATE reduces regime inference accuracy, demonstrating the importance of both adapting established regimes and expanding the regime space. The effects are particularly evident on Progressive, where disabling UPDATE and CREATE reduces accuracy from 91.49% to 86.90% and 81.01%, respectively. Removing established-regime stability checks causes a more substantial degradation, reducing accuracy to 66.25% and increasing MSE from 0.00081 to 0.00243. Removing CREATE validity requirements also degrades performance on Progressive, highlighting the importance of validating emerging regimes before their incorporation. Moreover, removing evidence persistence leads to greater performance variability across runs.

These results provide additional evidence of the individual contributions of UPDATE, CREATE, and stability-aware control to reliable online regime evolution. The complete component-wise ablation studies have been added to Appendix G of the revised manuscript, while the theoretical assumptions, propositions, and proofs are provided in Appendix C.

Reviewer_Ay5yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a regime-aware forecasting framework based on HMMs with variable number of latent regimes that can increase in an online manner. The method has two phases. In offline stage, a shared base predictor with regime specific heads is jointly optimized in sequence chunks that are constructed using a dynamic programming approach. In the online stage, an additional router scores insufficiency and novelty scores. Depending on certain criteria, UPDATE or CREATE commands are executed. The promplem is well motivated and experiment results are promising. Both synthetic and real-world dataset comparisons show that the proposed method is on par or better that the compared baselines.

Critical strengths
  • Most existing work indeed considers fixed number of states in nonstationary time series analysis with HMMs. This work is interesting because, it provides a way to enhance flexibility under heavily nonstationary scenarios where new states may emerge out-of-distribution. Also, distinguishing CREATE and UPDATE operations makes sense .

  • Progressive synthetic dataset indeed shows the benefit of having new states as time series progresses, which demonstrates the impact of the proposed technique.

  • Experiments are extensive, many of previous statistical and machine learning-based techniques are considered in comparisons.

Critical weaknesses
  • Overall, I find the proposed technique a bit heavy-engineered. There are many components, hyper parameters (for example in the loss calculation), thresholds etc. Most of these are not ablated and needs to be manually tuned.

  • No computation cost analysis is provided.

Questions for authors
  • How is regime accuracy matching evaluated if number of regimes can be variable?

  • Can this technique be used in multi-step forecasting?

  • How sensitive are the results to loss hyper parameters, thresholds etc?

10 comments

Thank you very much for your valuable suggestions. We have provided point-by-point responses to address all the raised questions and concerns.

Weakness 1: Overall, I find the proposed technique a bit heavy-engineered. There are many components, hyper parameters (for example in the loss calculation), thresholds etc. Most of these are not ablated and needs to be manually tuned.

Response to Weakness 1: Thank you for the comment. RegimeUnfold consists of two main modules: offline regime discovery and online regime evolution. The former involves predictive dynamics learning, regime inference, and adaptive structural search, while the latter consists of regime selection, evolution assessment, and stability-aware adaptation. Each component plays a necessary role in these two stages. We explain their respective roles below and provide corresponding component-wise ablation studies to validate their contributions.

(1) Offline Regime Discovery

Offline regime discovery involves several complementary mechanisms. The shared-base prediction loss supports the learning of common predictive dynamics, while regime representation and specialization losses encourage the differentiation of regime-specific dynamics. Regime-specific regularization prevents excessive residual corrections, whereas switching and duration penalties reduce fragmented assignments caused by local prediction fluctuations. Moreover, adaptive structural search allows the regime space to be refined when a fixed-regime solution is insufficient, with predictive and structural validation preventing unsupported changes. We consider five ablations to examine the effects of different training objectives and regime decoding mechanisms:

  • w/o Shared-Base Prediction Loss: Removes the auxiliary prediction loss of the shared base predictor to evaluate its contribution to learning common predictive dynamics.
  • w/o Regime Representation and Specialization Losses: Removes prototype commitment, prototype separation, and regime specialization losses to examine their effects on regime differentiation.
  • w/o Regime-Specific Regularization: Removes regularization on regime-specific residuals, biases, and non-assigned residuals to evaluate its contribution to stable dynamics learning.
  • w/o Temporal Regularization in Decoding: Removes both switching and duration penalties to examine the importance of temporal continuity in regime discovery.
  • w/o Soft Duration Penalty in Decoding: Removes only the soft duration penalty while retaining the switching penalty to evaluate the additional contribution of duration regularization.

The results on Fixed-K2 and Fixed-K3 are reported below.

Table 1: Regime discovery and dynamics fitting results on Fixed-K2.

ModelAccuracy (%)F1 (%)MAEMSE
w/o Shared-Base Prediction Loss93.02 ± 15.1687.78 ± 26.790.0121 ± 0.00630.00040 ± 0.00039
w/o Regime Representation and Specialization Losses95.19 ± 10.3995.97 ± 8.580.0095 ± 0.00130.00018 ± 0.00006
w/o Regime-Specific Regularization86.36 ± 18.4683.10 ± 26.090.0127 ± 0.00660.00050 ± 0.00048
w/o Temporal Regularization in Decoding83.23 ± 17.4478.77 ± 24.320.0135 ± 0.00500.00055 ± 0.00035
w/o Soft Duration Penalty in Decoding97.15 ± 5.1997.02 ± 5.590.0095 ± 0.00060.00024 ± 0.00006
RegimeUnfold97.68 ± 4.6897.55 ± 4.870.0097 ± 0.00040.00026 ± 0.00008

Table 2: Regime discovery and dynamics fitting results on Fixed-K3.

ModelAccuracy (%)F1 (%)MAEMSE
w/o Shared-Base Prediction Loss97.04 ± 4.4297.41 ± 3.140.0091 ± 0.00030.00016 ± 0.00004
w/o Regime Representation and Specialization Losses99.01 ± 0.2098.80 ± 0.250.0086 ± 0.00030.00014 ± 0.00002
w/o Regime-Specific Regularization91.53 ± 13.3188.40 ± 20.060.0096 ± 0.00150.00020 ± 0.00010
w/o Temporal Regularization in Decoding50.76 ± 11.2940.17 ± 8.210.0173 ± 0.00480.00079 ± 0.00039
w/o Soft Duration Penalty in Decoding88.21 ± 24.6985.73 ± 29.890.0112 ± 0.00520.00034 ± 0.00043
RegimeUnfold99.09 ± 0.2998.90 ± 0.360.0089 ± 0.00020.00015 ± 0.00001

(2) Online Regime Evolution

Online regime evolution faces three related challenges. First, regime selection must combine predictive compatibility and temporal persistence to identify an appropriate regime for forecasting, while hard single-expert selection preserves regime-specific predictions. Second, predictive insufficiency and structural novelty provide complementary evidence for detecting changes, with evidence persistence reducing responses to transient deviations. Third, UPDATE and CREATE address different forms of adaptation, while established-regime stability checks and CREATE validity requirements prevent unnecessary modifications to previously learned dynamics and premature regime-space expansion. We further consider eleven ablations covering regime selection, evolution assessment, and stability-aware adaptation.

For regime selection, we examine:

  • Predictive Branch Only: Retains only the predictive routing branch to assess the contribution of recent predictive evidence.
  • Temporal Branch Only: Retains only the temporal routing branch to evaluate the contribution of regime persistence and temporal dependencies.
  • Fixed Score Fusion: Assigns equal weights to predictive and temporal routing scores when reliable previous-regime context is available, instead of using adaptive fusion.
  • Soft Forecast Aggregation: Replaces single-expert forecasting with a weighted average of regime-specific predictions while retaining hard regime assignments.

For evolution assessment, we examine:

  • Predictive Insufficiency Only: Uses only predictive-insufficiency evidence to trigger adaptation.
  • Structural Novelty Only: Uses only structural-novelty evidence to trigger adaptation.
  • w/o Online Evidence Persistence: Removes evidence persistence requirements to evaluate their role in filtering transient changes.

For stability-aware adaptation, we examine:

  • w/o UPDATE: Prevents the incorporation of updates to established regimes to evaluate the contribution of adapting existing dynamics.
  • w/o CREATE: Prevents permanent regime-space expansion to assess the importance of incorporating emerging regimes.
  • w/o Established-Regime Stability Checks at Admission: Removes established-regime protection checks during candidate acceptance to evaluate their role in preserving previously learned dynamics.
  • w/o CREATE Validity Requirement: Removes the additional persistence and structural distinction requirements for CREATE to evaluate their role in preventing unnecessary regime creation.

The results on Fixed-K2, Fixed-K3, and Progressive are reported below.

Table 3: Online regime inference and forecasting results on Fixed-K2.

ModelAccuracy (%)F1 (%)MAEMSE
Predictive Branch Only95.98 ± 2.1097.08 ± 1.750.0093 ± 0.00020.00022 ± 0.00001
Temporal Branch Only95.13 ± 3.3095.97 ± 3.360.0099 ± 0.00140.00040 ± 0.00040
Fixed Score Fusion95.68 ± 2.9496.77 ± 2.640.0093 ± 0.00020.00022 ± 0.00001
Soft Forecast Aggregation95.78 ± 2.6096.88 ± 2.270.0094 ± 0.00020.00022 ± 0.00001
Predictive Insufficiency Only97.79 ± 2.2997.61 ± 2.160.0094 ± 0.00030.00023 ± 0.00002
Structural Novelty Only95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00030.00022 ± 0.00002
w/o Online Evidence Persistence95.29 ± 2.5196.67 ± 2.230.0092 ± 0.00020.00022 ± 0.00001
w/o UPDATE95.53 ± 2.4596.84 ± 2.250.0092 ± 0.00020.00021 ± 0.00001
w/o CREATE95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
w/o Established-Regime Stability Checks at Admission95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
w/o CREATE Validity Requirement95.78 ± 2.6096.88 ± 2.270.0093 ± 0.00020.00022 ± 0.00001
RegimeUnfold97.85 ± 2.2297.68 ± 2.080.0093 ± 0.00020.00022 ± 0.00001

Table 4: Online regime inference and forecasting results on Fixed-K3.

ModelAccuracy (%)F1 (%)MAEMSE
Predictive Branch Only95.09 ± 0.9895.15 ± 0.360.0097 ± 0.00020.00040 ± 0.00004
Temporal Branch Only91.11 ± 1.1190.27 ± 0.930.0109 ± 0.00120.00067 ± 0.00038
Fixed Score Fusion94.52 ± 1.0794.49 ± 0.480.0097 ± 0.00020.00040 ± 0.00004
Soft Forecast Aggregation95.23 ± 1.1295.32 ± 0.480.0098 ± 0.00020.00041 ± 0.00003
Predictive Insufficiency Only96.27 ± 0.1495.83 ± 0.170.0094 ± 0.00010.00026 ± 0.00002
Structural Novelty Only95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o Online Evidence Persistence93.94 ± 1.4394.76 ± 0.580.0098 ± 0.00010.00046 ± 0.00002
w/o UPDATE94.71 ± 1.3795.16 ± 0.470.0097 ± 0.00020.00041 ± 0.00003
w/o CREATE95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o Established-Regime Stability Checks at Admission95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
w/o CREATE Validity Requirement95.23 ± 1.1295.32 ± 0.480.0097 ± 0.00020.00041 ± 0.00003
RegimeUnfold96.31 ± 0.2395.88 ± 0.280.0093 ± 0.00010.00026 ± 0.00001

Table 5: Online regime inference and forecasting results on Progressive.

ModelAccuracy (%)F1 (%)MAEMSE
Predictive Branch Only85.76 ± 4.3785.14 ± 5.300.0147 ± 0.00130.00096 ± 0.00011
Temporal Branch Only81.17 ± 5.6179.66 ± 10.420.0208 ± 0.00260.00208 ± 0.00014
Fixed Score Fusion87.58 ± 7.9285.36 ± 13.490.0143 ± 0.00270.00088 ± 0.00021
Soft Forecast Aggregation86.13 ± 4.0186.23 ± 5.490.0146 ± 0.00100.00089 ± 0.00008
Predictive Insufficiency Only75.89 ± 7.4665.60 ± 11.580.0162 ± 0.00310.00108 ± 0.00033
Structural Novelty Only86.92 ± 4.9386.11 ± 6.480.0153 ± 0.00210.00104 ± 0.00024
w/o Online Evidence Persistence90.08 ± 3.9091.76 ± 4.680.0137 ± 0.00060.00087 ± 0.00003
w/o UPDATE86.90 ± 5.0786.87 ± 6.260.0145 ± 0.00130.00094 ± 0.00011
w/o CREATE81.01 ± 1.6581.74 ± 2.720.0151 ± 0.00030.00095 ± 0.00004
w/o Established-Regime Stability Checks at Admission66.25 ± 16.7955.10 ± 25.170.0260 ± 0.01100.00243 ± 0.00157
w/o CREATE Validity Requirement86.43 ± 3.5986.32 ± 5.360.0148 ± 0.00110.00096 ± 0.00008
RegimeUnfold91.49 ± 0.6191.19 ± 0.770.0134 ± 0.00000.00081 ± 0.00001

The full RegimeUnfold achieves the highest regime discovery accuracy and F1 scores on both offline datasets, while maintaining low dynamics fitting errors. In particular, removing temporal regularization in offline regime decoding reduces accuracy from 99.09% to 50.76% on Fixed-K3. For online evolution, the full model achieves the highest regime inference accuracy across all three datasets and maintains strong forecasting performance. On Progressive, disabling CREATE reduces accuracy from 91.49% to 81.01%, while removing established-regime stability checks causes a larger reduction to 66.25%. These results provide direct evidence of the contributions of the main components to reliable regime discovery and continual regime-space adaptation.

Importantly, all five synthetic datasets use the same hyperparameter configuration, without dataset-specific tuning, despite differences in their regime numbers, transition patterns, and regime-space evolution. For the four real-world datasets reported in the original manuscript, only a small subset of dataset-dependent parameters was adjusted, while the remaining settings were kept unchanged, as shown in Table 6. The sensitivity analysis in our response to Question 3 further indicates that model performance is more sensitive to certain key parameters, whereas many other settings, particularly structural selection thresholds, have relatively limited effects within the tested ranges.

Table 6: Dataset-dependent hyperparameter settings of RegimeUnfold.

HyperparameterSyntheticSMDExchangeTrafficGOLD
Input history length hh12284
Backbone hidden dimension888162
Regime-specific predictor hidden dimension844168
Residual penalty λres\lambda_{\mathrm{res}}0.0050.010.0050.0050.005
Switching-penalty scale cswc_{\mathrm{sw}}0.71.40.70.71.4
Duration scale ℓdur\ell_{\mathrm{dur}}8208820

The complete component-wise ablation studies and hyperparameter sensitivity analyses are provided in Appendix G of the revised manuscript, with the latter discussed separately in our response to Question 3. The full hyperparameter settings across synthetic and real-world datasets are reported in Appendix F.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely appreciate the time, expertise, and thoughtful feedback that the reviewers (Ay5y, NbVW, EK1o) devoted to evaluating our work. We are encouraged by their recognition that discovering and continually adapting an initially unknown regime space addresses an important and practical problem in HMM-based forecasting (Ay5y, NbVW, EK1o), and that the proposed RegimeUnfold framework is meaningful and its effectiveness is demonstrated by the experimental results (Ay5y, EK1o).

We value the reviewers’ careful scrutiny and have strengthened the paper with additional theoretical analyses, experiments, and clarifications. All revisions are highlighted in red in the revised manuscript.

1. Response to Common Concerns

Model Complexity and Component Necessity (All Reviewers) Reviewers raised concerns about the complexity of RegimeUnfold and the insufficient validation of individual components, particularly the necessity and effectiveness of the proposed regime-space evolution mechanisms.

  • We expanded the component-wise ablation studies to examine the contributions of individual modules in offline regime discovery and online regime evolution, including hard regime decoding, UPDATE, CREATE, and stability-aware control (Appendix G.5).
  • We added comparisons with Growing Experts strategies to further validate the advantages of the proposed regime-space evolution mechanisms (Appendix G.6).

Theoretical Foundations and Reliability of Regime-Space Evolution (NbVW, EK1o) Reviewers questioned the theoretical foundations of regime-space learning, particularly whether the learned regimes can converge to the underlying regime structure and whether online adaptation can reliably accommodate changing dynamics while preserving stability.

  • We added theoretical analyses showing that, under appropriate conditions, the learned regime assignments and regime number can converge to their true underlying counterparts (Appendix C.1).
  • We added theoretical analyses showing that, under appropriate conditions, persistent structural-novelty evidence can reduce false activations, while validation-based online adaptation can improve forecasting performance while controlling predictive degradation on established regimes (Appendix C.2).

Experimental Validation and Robustness (All Reviewers) Reviewers emphasized the need for more comprehensive and fair baseline comparisons, stronger evaluation under complex predictive dynamics, and clearer reporting of parameter sensitivity and experimental settings.

  • We added four representative baselines, including two deep switching state-space models (RED-SDS and DS$^3$M) and two online forecasting methods (FSNet and OneNet), and expanded the main tables to include Adaptive, Non-adaptive, and Large-window comparisons (Section 5 and Appendix G).
  • We introduced an additional 64-dimensional nonlinear Progressive dataset and evaluated three multi-step forecasting strategies to further assess the applicability of RegimeUnfold (Appendices G.3 and G.8).
  • We conducted extensive hyperparameter sensitivity analyses and clarified the parameter settings and evaluation protocols (Appendices F and G).

2. Summary of Revisions

Prompted by these valuable reviews, we have significantly strengthened the manuscript:

  • Strengthened Theoretical Foundations: Added theoretical analyses of regime differentiation, convergence of learned regimes to the underlying regime structure, and the reliability and stability of online regime evolution under appropriate conditions (Appendix C).
  • Expanded Evaluation: Added four representative baselines (RED-SDS, DS$^3$M, FSNet, and OneNet), a 64-dimensional nonlinear Progressive dataset, and three multi-step forecasting strategies, while providing more comprehensive Adaptive, Non-adaptive, and Large-window comparisons (Section 5 and Appendix G).
  • Comprehensive Component Validation: Expanded the offline and online component-wise ablation studies to validate the contributions of individual mechanisms, including hard regime decoding, UPDATE, CREATE, and stability-aware control. We also added comparisons with Growing Experts strategies (Appendix G).
  • Hyperparameter Sensitivity and Computational Analysis: Provided detailed hyperparameter settings and sensitivity analyses, together with empirical measurements of training time and online runtime (Appendices F–H).

These revisions substantially improve the theoretical rigor, empirical validation, clarity, and reproducibility of RegimeUnfold.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.