acceptodds
Under review as a conference paper at ICLR 2027

COVPW-RPC: CERTIFIED RELATION-PRESERVING CORRELATION CONTROL FOR PAIRED TIME-SERIES GENERATION

Abstract

Realistic individual time series can form an unrealistic pair. Correcting their comovement may also destroy a relation that the sample already captures. We propose CovPW-RPC, a post-sampling controller for paired diffusion generation. It preserves the declared relation and scales only centered common increments. An exact, monotone response maps this scale to log-return correlation. Its inverse selects the smallest feasible edit toward a target interval, subject to action and energy limits. Targets can be training-set constants or predictions from observed context; the action adapts to each generated path. No generator retraining is required. Identical-draw tests reduce rolling-correlation distribution error on all 15 financial pairs and all ten weather-station pairs. A prospective test on 12 assetdisjoint financial pairs reduces this error by 38.52%. Simple training targets retain over 98% of that gain, supporting the value of the path-dependent inverse. A retrospective fixed-scale comparison further isolates this benefit. The unchanged controller also improves a second diffusion construction. On GRU samples, the frozen policy fails, while a distribution-targeted extension improves development and retrospective temporal results. Exact relation preservation coexists with empirical variance, tail and cointegration tradeoffs. The contribution is a tractable control mechanism: specify what a generated path must retain, expose what it can change, and measure the quality of the resulting edit.

Primary area:
generative models
Keywords:
time-series generation, diffusion models, controllable generation, post-sampling control, relation preservation, correlation calibration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XZrQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to improve the correlation between sample paths of related time series, while preserving a "protected relation". This is done by decoupling the "common coordinate" and "spread" (the ratio between the two series), and editing the correlation while keeping the spread fixed.

Critical strengths

The main text and appendices are very detailed

Critical weaknesses

The writing is very poor and difficult for a reader to understand.

The contribution seems to be a fairly simple trick and the gains are fairly marginal.

Questions for authors

None

Reviewer_ccotOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CovPW-RPC, a post-sampling controller for paired time-series generation. It improves the fidelity of cross-series correlation while exactly preserving a declared log-spread relation, without retraining the generator. Scaling the common-mode log increments with one scalar per path gives an exact, monotone log-return correlation response, which the controller inverts to find the smallest budgeted edit toward a target. Experiments show reduced rolling-correlation distribution error on financial and weather pairs, including a 38.52% reduction on prospective CSI300 pairs. The frozen policy does not transfer to a GRU generator, and the edit trades off variance, tail, and cointegration fidelity.

Critical strengths

S1: The paper proposes a novel post-sampling correlation control method. It exactly preserves the declared log-spread relation and has an exact, monotone correlation response within its scalar family.

S2: The controller is lightweight: it needs no retraining or access to the diffusion sampling loop, and adds about 1% runtime overhead.

S3: On identical diffusion draws, the controller improves correlation fidelity across financial and weather pairs, including a prospective test on asset-disjoint CSI300 pairs.

Critical weaknesses

W1: The evaluation and theory focus on zero-lag correlation. The appendix ACF and cross-lag results are mixed, with some temporal dependencies deteriorating after the edit. Can the authors evaluate multiple temporal scales and characterize the controller’s effect on autocorrelation and lagged cross-correlations?

W2: The method is limited to K = 2 and evaluated only on financial and weather pairs. This restricts its use for multi-asset or complex spatio-temporal data. Discussing how the framework extends to K > 2 and testing on more diverse datasets would strengthen the paper.

W3: The guarantees concern log-return correlation, but the metric is the distributional distance of rolling percentage-return correlations. The provided bound does not imply that this distance decreases, so a theoretical link to distribution-level fidelity is missing.

W4: The gains depend heavily on the target policy. The frozen policy worsens correlation fidelity on 11/12 GRU pairs, and simple training-constant targets recover over 98% of the gain (Table 4), so the benefit of the proposed context-based policy is unclear.

Questions for authors

Please see details in the Weaknesses section.

Reviewer_3phkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CovPW-RPC, a post-sampling correlation for controlling correlation in paired time-series generation. It decomposes generated sequences into common and spread components, then adjusts the common increments using a scaling factor \eta while preserving the original spread. The authors derive an exact relationship between the scaling factor and log-return correlation, allowing sample-adaptive edits under predefined constraints. Experiments on financial and weather data demonstrate improvements in correlation distribution error. However, the practical value and actual benefits of the proposed method remain unclear. The paper is also difficult to follow, and its presentation needs substantial improvement.

Critical strengths

S1. The paper presents a mathematically well-defined post-sampling control mechanism. The exact correlation response provides an efficient way to determine bounded edits without the need to access the underlying generator. S2. The experiments include multiple results to provide useful evidence for correlation correction and disclose several limitations of the method.

Critical weaknesses

W1. The practical value of the proposed method is not clearly demonstrated. First, the method strictly preserves the generated log-spread, but the paper does not sufficiently explain why this particular relation should be preserved in practical time-series generation tasks. The generated spread itself may be inaccurate, and preserving it does not guarantee a more realistic generated pair. Second, although correlation is an important property, improving correlation alone does not support better overall generation quality. Specifically, the paper reports additional distributional metrics and downstream evaluations, but the results show mixed quality trade-offs and limited practical benefits. Third, the method is restricted to paired sequences with K=2. However, real-world multivariate generation often involves more complex dependencies among multiple variables. The considered setting is too narrow. W2. The experimental evidence provides limited support for the proposed method. First, the original RPC policy worsens Corr on 11/12 GRU pairs. Although RPC is presented as a generator-independent post-processing method, the paper does not clearly explain why its effectiveness varies so significantly across generators, raising serious concerns about its generalizability. Second, the comparisons with existing constrained generation methods are insufficient to demonstrate the superiority of the proposed method. Specifically, RPC does not demonstrate significant improvements over the PDM adaptation in Sec. 5.2. Moreover, other relevant methods are not directly evaluated.

Questions for authors

Q1. Why is it necessary to preserve the generated log-spread? Can the authors provide a practical application showing the benefit of this constraint? Q2. Does improving Corr lead to better overall generation quality? How do the authors explain the deterioration in Coint and some marginal-distribution metrics? Q3. Why does the original RPC policy worsen Corr on 11/12 GRU pairs, despite being presented as generator-independent? How can we determine when the method is effective? Q4. What is the main advantage of the sample-adaptive inverse over simpler post-processing methods? Can the authors provide a comparison under similar actual edit budgets and further clarify the advantages over PDM?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.