TL;DR
NetBurst is a novel event-centric forecasting framework designed for bursty, intermittent network telemetry time series, significantly improving accuracy and clustering compared to existing methods by leveraging quantile-based codebooks and dual autoregressors.
Contribution
The paper introduces NetBurst, a new approach that reformulates forecasting as event prediction, effectively handling heavy-tailed, bursty data regimes inspired by Mandelbrot's statistical insights.
Findings
Reduces MASE by up to 605x on service-level data
Produces embeddings that cluster 5x more clearly than baselines
Effectively captures burstiness and heavy-tailed behaviors
Abstract
Forecasting on widely used benchmark time series data (e.g., ETT, Electricity, Taxi, and Exchange Rate, etc.) has favored smooth, seasonal series, but network telemetry time series -- traffic measurements at service, IP, or subnet granularity -- are instead highly bursty and intermittent, with heavy-tailed bursts and highly variable inactive periods. These properties place the latter in the statistical regimes made famous and popularized more than 20 years ago by B.~Mandelbrot. Yet forecasting such time series with modern-day AI architectures remains underexplored. We introduce NetBurst, an event-centric framework that reformulates forecasting as predicting when bursts occur and how large they are, using quantile-based codebooks and dual autoregressors. Across large-scale sets of production network telemetry time series and compared to strong baselines, such as Chronos, NetBurst reduces…
Peer Reviews
Decision·Submitted to ICLR 2026
- I like the fact that the authors delved deep into a specific type of time series and collated datasets specific to this domain of time series. We need more such deep-dive analyses instead of generic time series benchmarks (which are also very useful) - The authors motivate their work very well. The preliminary analyses that they present in Fig 1 with Fano factors, autocorrelation and local example is interesting. - The presented model is simple and builds on prior work (Chronos) with targeted
- Many parts of the paper are unclear, all of which cloud my evaluation of the paper. I mention all these in the questions section. The only reason for my score of 6 is the presentation which is very underwhelming.
* **Novelty:** The core strength of this paper is its fundamental rethinking of the forecasting problem for this specific data regime. The shift from a sequence-value regression paradigm to an event-prediction paradigm is a powerful and well-justified conceptual leap. By explicitly separating "when" a burst occurs (IBG) from "how large" it is (BI), the model directly addresses the entanglement of timing and magnitude that destabilizes conventional forecasters. This event-centric view is a sign
The paper has significant issues, and the authors did not sufficiently discuss its potential weaknesses. 1. **Narrow Applicability and Specialized Domain:** The NETBURST framework is explicitly and brilliantly designed for one specific, albeit important, statistical regime: bursty, intermittent, heavy-tailed time series. The paper does not provide any experiments or discussion on how the model would perform on the smooth, seasonal benchmarks (like ETT, Electricity) where conventional models ex
Converting rare outlier or burst values into events is a very natural approach. The resulting model is elegant and sound. Most of the paper is clear and easy to understand. The choices of architecture and data processing are reasonable. Ablation studies answer most of the the questions I would have. The results are good.
It is disappointing that this model does not handle "dense" time series with bursts but is converting a time series to a sparse representation of events and modeling them as a temporal point process. Ideally, I would like to see a conventional time series model which is enriched with outliers or bursts viewed as events. The model would predict normally most of the time but also predict the bursts on top of that. Unless I'm missing something, this model is not doing that. Formatting is bad. Figu
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
