Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting

arXiv:2610.10222 · cs.LG · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting".

Jane: The gist The study formulates post-generation sequence assembly as an explicit design stage in forecasting pipelines based on differentially private synthetic time series.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we're looking at this paper, "Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting." The core idea here is that differentially private generators often produce these fixed-length synthetic windows, but the forecasting models need long continuous sequences to train on. How you stitch those windows together after generation totally changes what the forecaster actually sees.

Jane: Right, so it's not just about making the generator private during training; it's about how we assemble the output afterwards because that assembly process directly affects what kind of data a model gets to look at.

Lu: The paper systematically varies overlap rates and window-weighting schemes to see how these choices impact boundary continuity, statistical and temporal fidelity, and then Train-on-Synthetic-Test-on-Real TSTR forecasting utility across four public datasets and five different forecasting models.

Meng: I'm curious about the actual cost here. We're talking about running all those combinations for every model on every dataset just to see what works at scale, right?

Lalam: The paper finds that there is a clear forecaster-dependent assembly principle, meaning the best way to assemble the data really depends on which forecasting model you are using.

Tom: That's interesting because it means we can't just use one universal assembly rule for everyone; we have to tailor the assembly process to the specific downstream model.

Jane: And what that means in practice is that if you want good performance, you need to consider three things: the forecaster itself, the overlap rate you choose, and the weighting scheme you apply when combining those synthetic windows.

Lu: The study finds that increased overlap generally improves boundary continuity, but it doesn't always mean better forecasting error in terms of individual fidelity or continuity diagnostics.

Tom: That’s a bit counterintuitive; we might expect better stitching to always be good, but the paper suggests that just having a smooth boundary isn't enough on its own to guarantee a better forecast.

Jane: It points out that improvements in those continuity or individual fidelity measurements don't consistently reduce forecasting error, which tells us that those diagnostics aren't the only things we should be looking at when picking an assembly configuration.

Meng: From a deployment standpoint, if we spend time optimizing for a specific continuity metric and end up with no real improvement in the final TSTR utility, that’s wasted effort at scale.

Lalam: The paper does provide some concrete guidance here, showing that recurrent models more often favor intermediate-to-high overlap settings when assembling these sequences.

Tom: Okay, so we have a preference based on the model type; recurrent models like LSTM or RNN seem to prefer higher overlap rates compared to things like ARIMA or Prophet.

Jane: That’s a key distinction because it suggests that the way a model inherently learns temporal dependencies influences how much blending of those synthetic windows it can handle well.

Lu: The paper also notes that ARIMA and Prophet tend to favor zero or low overlap, which makes sense if they are very sensitive to changes in the temporal structure introduced by aggressive blending.

Meng: It’s important to remember that LightGBM shows a stronger dependence on the specific dataset when it comes to assembly settings, which suggests its behavior is more tied to the particular features it learns from that data.

Tom: So, we've got these model-specific preferences—recurrent models want high overlap, ARIMA wants low overlap—which gives us a starting point for configuration selection.

Jane: The paper also uses matched Train-on-Real-Test-on-Real references to show how the TSTR utility compares against training directly on real data, which contextualizes these assembly choices against what we know works best with real data.

Lu: These matched TRTR results help quantify the utility of each assembly setting relative to training on real data, helping us pinpoint which configuration is actually leading to better performance for a specific task.

Meng: So, the paper suggests that downstream TSTR utility is the actual basis for selecting the best assembly setting, rather than just focusing on continuity metrics alone.

Tom: Exactly; we need that final utility score to guide our choice of how we combine those windows after they're generated. This whole study, "Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting," positions sequence assembly as a forecaster-aware model selection problem.

Jane: And what this means for us is that when we use these differentially private synthetic time-series in forecasting systems, we need to be explicit about the assembly strategy you choose based on your chosen forecasting model.

Lu: The paper also validated these findings with a second generator architecture, DP-ConvVAE, and it reproduced the forecaster-dependent assembly response, showing this isn't just an artifact of one specific generator type.

Meng: It sounds like we should start testing configurations based on the forecaster first and then use those preferences to narrow down the search space before doing expensive full evaluations.

Tom: That’s a practical application; it gives us actionable guidance for prioritizing candidate configurations for downstream evaluation, which is what this paper ultimately provides.

Jane: So, in short, we can use these forecaster-dependent patterns to prioritize a smaller set of assembly configurations before we spend all our time validating them on the final target task.

Conclusion: Tom: So, we've been looking at this paper about "Evaluating Sequence Assembly Strategies for Differentially Private Synthetic Time-Series Forecasting." It’s pretty clear that how you stitch those synthetic windows together after they’re made really matters for what a forecasting model actually learns.

Jane: That's right. The authors are testing different overlap rates and weighting schemes on these differentially private time series to see which assembly works best for different models.

Lu: They found something specific about the forecaster, meaning the best assembly method isn't one-size-fits-all; it depends on whether you’re using a recurrent model or an ARIMA model.

Meng: So, what does this mean for deploying these synthetic series in production? We can’t just pick any random overlap setting; we have to tailor it to the AI we plan to use downstream.

Lalam: I'm confident in this finding because they used a complete grid of five different forecasters and matched it against real training data references, showing a clear pattern across four public datasets.

Tom: It’s like having a map for assembly based on which car you plan to drive—a recurrent model prefers higher overlap, while ARIMA favors zero or low overlap.

Jane: That’s the practical implication: you need to consider the forecaster when designing your data pipeline because that dictates how those synthetic windows are put together.

Lu: The study showed that while increasing overlap helps with boundary continuity, it doesn't always translate directly into better forecasting error in terms of individual accuracy metrics.

Meng: That’s a crucial caveat for us; we can optimize the stitch to be smooth, but if it doesn't improve the actual prediction quality against real data, then that optimization is pointless.

Tom: Exactly. The paper suggests that downstream utility—the actual forecast performance on real data—is the true measure we should focus on when picking an assembly strategy.

Jane: So what’s next? We need to look at how these specific model preferences hold up when we test them against a different type of generator architecture, like this DP-ConvVAE they used for comparison.

Guoxiong Longa, Huizhen Huanga, Qikun Caia, Tao Huanga, *Chen Houa

School of Computer and Big Data, Minjiang University

cs.LG

Submitted: 2026-10-07

Updated: 2026-10-07

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 92/100

The gist: The gist The study formulates post-generation sequence assembly as an explicit design stage in forecasting pipelines based on differentially private synthetic time series.

Key concepts

Sequence Assembly: This is the process of stitching together multiple fixed-length synthetic time series into a single long continuous sequence that a forecasting model can use for training. The research systematically varied how much these pieces overlap and how they are weighted to see which assembly method yields better results for different models.
Boundary Continuity: This measures the smoothness or abruptness of the transitions between different synthetic windows when they are stitched together. A good assembly aims for high continuity, meaning there should be no noticeable jumps or sudden changes in the data signal at where one window ends and another begins.
Train-on-Synthetic-Test-on-Real (TSTR) Utility: This is a practical metric used to evaluate how useful the assembled synthetic data is. A forecaster is trained exclusively on this assembled synthetic sequence, and then its performance is tested on real, unseen data. This directly measures the utility of the assembly strategy for actual forecasting tasks.

Terminology

Differentially Private — 1× in this paper · explained in 3 episodes

Summary

The gist The study formulates post-generation sequence assembly as an explicit design stage in forecasting pipelines based on differentially private synthetic time series.

How it works

The research systematically varies overlap rates and window-weighting schemes while holding the trained generator fixed to evaluate the resulting sequences in terms of boundary continuity, statistical and temporal fidelity, and Train-on-Synthetic-Test-on-Real (TSTR) forecasting utility The assembly operation can alter marginal distributions, temporal dependence, and the temporal patterns available to a forecasting model.
The evaluation examines three complementary dimensions of sequence assembly: boundary continuity, statistical and temporal fidelity, and downstream forecasting utility.
Boundary continuity is measured using the Jjump diagnostic, which targets discontinuities at assembly junctions. Proposition 1 shows that gradual weight changes limit the contribution of cross-window mismatch, while within-window variation remains part of the assembled signal.
Temporal fidelity is assessed by computing ∆ACF24h = ACFl24h(Xˆ) − ACFl24h(X), and downstream utility is measured using the TSTR protocol where a forecaster is trained on the assembled synthetic data and evaluated on held-out real observations.
The study identifies forecaster-dependent assembly regularities, revealing that recurrent models more often favor intermediate-to-high overlap, ARIMA and Prophet favor zero or low overlap, and LightGBM is more dataset dependent.

Key Findings Across Datasets

Increased overlap generally improves boundary continuity, but improvements in continuity or individual fidelity diagnostics do not consistently reduce forecasting error.
Complete five-forecaster assembly grids reveal a clear forecaster-dependent pattern, showing that recurrent forecasters more often favored intermediate-to-high overlap, ARIMA and Prophet favored zero or low overlap, and LightGBM showed stronger dataset dependence.
Matched Train-on-Real-Test-on-Real (TRTR) references contextualize utility relative to real training, showing that the lowest MAE configurations in the grid identify candidate settings for the target forecasting task.

Operational Validation and Implications

The identified forecaster-dependent assembly principles are validated in an operational electricity-load forecasting application, where the same broad forecaster-dependent assembly pattern persists.
The results establish that sequence assembly is a forecaster-aware model-selection problem, providing practical guidance for prioritizing candidate configurations for downstream evaluation.
In practice, the findings suggest that Downstream TSTR utility provides the basis for configuration selection, while continuity and fidelity diagnostics offer complementary interpretation and matched TRTR provides the corresponding real-training reference.

Cross-Generator Check

A comparison using a second generator architecture (DP-ConvVAE) confirms the findings, showing that lower Jjump across the three assembly settings is accompanied by lower LSTM MAE and higher ARIMA MAE, reproducing the forecaster-dependent contrast.

The experiments use the public datasets ETTh1, ETTm1, Weather, and Appliances to identify forecaster-dependent sequence-assembly regularities through joint evaluation of boundary continuity, statistical and temporal fidelity, and downstream TSTR utility. The results establish sequence assembly as a forecaster-aware model-selection problem and provide actionable guidance for deploying differentially private synthetic time series in forecasting systems. Across ETTh1, ETTm1, Weather, and Appliances, increased overlap generally reduced junction discontinuity, but improvements in continuity or individual fidelity diagnostics did not consistently reduce forecasting error. Complete five-forecaster assembly grids revealed a clear forecaster-dependent pattern: recurrent forecasters more often favored intermediate-to-high overlap, ARIMA and Prophet favored zero or low overlap, and LightGBM showed stronger dataset dependence. Matched TRTR references placed the resulting TSTR variation on the corresponding real-training error scale. The public-dataset regularities were further examined in an operational electricity-load forecasting task, where the same broad forecaster-dependent assembly tendencies persisted. LSTM and RNN achieve their lowest MAE at 75% and 85.7% overlap, respectively, consistent with the higher-overlap preference of recurrent forecasters. ARIMA and Prophet attain their minima at low or zero overlap, consistent with their public-dataset tendency, whereas LightGBM again shows task-specific behavior. These results provide domain-specific support for the forecaster-dependent assembly regularities and illustrate how the public-dataset findings can guide candidate prioritization for downstream evaluation. The complete five-forecaster grids expose structure that is hidden by any single forecasting model. Recurrent forecasters generally favor intermediateto-high overlap, where averaging and tapering can moderate local junction effects while preserving useful temporal patterns. ARIMA and Prophet favor zero or low overlap more often, indicating greater sensitivity to the temporal changes introduced by aggressive blending. LightGBM shows stronger dataset dependence across assembly settings. These patterns provide a practical ordering of assembly configurations before target-task validation rather than a single globally optimal assembly rule. Downstream TSTR utility provides the basis for configuration selection, while continuity and fidelity diagnostics offer complementary interpretation and matched TRTR provides the corresponding real-training reference. The DP-ConvVAE comparison further reproduced a forecaster-dependent assembly response under a second generator architecture. Together, these results establish sequence assembly as a forecaster-aware model-selection problem and show how multi-dataset regularities can inform domain-specific deployment of differentially private synthetic time series. The study formulated post-generation sequence assembly as an explicit design stage in forecasting pipelines based on differentially private synthetic time series. Across ETTh1, ETTm1, Weather, and Appliances, increased overlap generally reduced junction discontinuity, but improvements in continuity or individual fidelity diagnostics did not consistently reduce forecasting error. Complete five-forecaster assembly grids revealed a clear forecaster-dependent pattern: recurrent forecasters more often favored intermediate-to-high overlap, ARIMA and Prophet favored zero or low overlap, and LightGBM showed stronger dataset dependence. Matched TRTR references placed the resulting TSTR variation on the corresponding real-training error scale. The public-dataset regularities were further examined in an operational electricity-load forecasting task, where the same broad model-specific tendencies persist. In practice, these regularities can be used to prioritize a smaller set of assembly configurations for target-task evaluation with the intended forecaster. Downstream TSTR utility provides the basis for configuration selection, while continuity and fidelity diagnostics offer complementary interpretation and matched TRTR provides the corresponding real-training reference. The DP-ConvVAE comparison further reproduced a forecaster-dependent assembly response under a second generator architecture. Together, these results establish sequence assembly as a forecaster-aware model-selection problem and show how multi-dataset regularities can inform domain-specific deployment of differentially private synthetic time series. Complete five-forecaster assembly grids revealed a clear forecaster-dependent pattern: recurrent forecasters more often favored intermediate-to-high overlap, ARIMA and Prophet favored zero or low overlap, and LightGBM showed stronger dataset dependence. Matched TRTR references placed the resulting TSTR variation on the corresponding real-training error scale. The public-dataset regularities were further examined in an operational electricity-load forecasting task, where the same broad model-specific tendencies persist. LSTM and RNN achieve their lowest MAE at 75% and 85.7% overlap, respectively, consistent with the higher-overlap preference of recurrent forecasters. ARIMA and Prophet attain their minima at low or zero overlap, consistent with their public-dataset tendency, whereas LightGBM again shows task-specific behavior. These results provide domain-specific support for the forecaster-dependent assembly regularities and illustrate how the public-dataset findings can guide candidate prioritization for downstream evaluation. The study formulated post-generation sequence assembly as an explicit design stage in forecasting pipelines based on differentially private synthetic time series. Across ETTh1, ETTm1, Weather, and Appliances, increased overlap generally reduced junction discontinuity, but improvements in continuity or individual fidelity diagnostics did not consistently reduce forecasting error. Complete five-forecaster assembly grids revealed a clear forecaster-dependent pattern: recurrent forecasters more often favored intermediate-to-high overlap, ARIMA and Prophet favored zero or low overlap, and LightGBM showed stronger dataset dependence. Matched TRTR references placed the resulting TSTR variation on the corresponding real-training error scale. The public-dataset regularities were further examined in an operational electricity-load forecasting task, where the same broad model-specific tendencies persist. LSTM and RNN achieve their lowest MAE at 75% and 85.

Improvements for AI systems

Bold header: Forecaster-Aware Assembly Selection

This improves AI systems by shifting sequence assembly from a generic preprocessing choice to a forecaster-aware model-selection problem (Page 5). The improved system can prioritize specific overlap rates and windowing schemes based on the downstream forecaster, as evidenced by the finding that downstream TSTR utility is jointly shaped by the forecaster, overlap rate, and window-weighting schemes, leading to distinct assembly preferences across forecasting models (Page 1).

Bold header: Model-Specific Assembly Prioritization

The system can implement configuration selection based on model architecture; for example, recurrent models should generally favor intermediate-to-high overlap, while ARIMA and Prophet should utilize zero or low overlap (Page 5). This allows the AI to select assembly candidates that align with the inherent temporal structure required by the specific forecasting model being deployed.

Bold header: Utility-Driven Assembly Validation

The system can use Matched Train-on-Real-Test-on-Real (TRTR) references to determine configuration suitability, as downstream utility is evaluated primarily using the Train-on-Synthetic Test-on Real (TSTR) protocol (Page 3). This ensures that assembly choices are selected based on maximizing TSTR performance relative to real data training, rather than solely on local diagnostics like boundary continuity.

Abstract

Differentially private time-series generators commonly produce fixed-length synthetic windows, whereas downstream forecasting models often require long continuous training sequences. How these windows are assembled after generation can therefore alter the effective synthetic data presented to a forecaster, even when the trained generator remains unchanged. We study this post-generation sequence assembly process by systematically varying overlap rates and window-weighting schemes and evaluating the resulting sequences in terms of boundary continuity, statistical and temporal fidelity, and Train-on-Synthetic-Test-on-Real (TSTR) forecasting utility. Across four types of public datasets (ETTh1, ETTm1, Weather, and Appliances) and five forecasting models, the results reveal a clear forecaster-dependent assembly principle: downstream TSTR utility is jointly shaped by the forecaster, overlap rate, and window-weighting scheme, leading to distinct assembly preferences across forecasting models. Increased overlap generally improves boundary continuity, but improvements in continuity or individual fidelity diagnostics do not consistently reduce forecasting error, indicating that these diagnostics alone are insufficient for selecting assembly configurations. Complete five-forecaster assembly grids, together with matched Train-on-Real-Test-on-Real (TRTR) references, further characterize these regularities and quantify assembly-dependent utility relative to real-data training. We then validate the identified principles through additional analyses of robustness and generator variability.