Zero-shot Forecasting by Simulation Alone
Boris N. Oreshkin, Mayank Jauhari, Ravi Kiran Selvam, Malcolm Wolff, Wenhao Pan, Shankar Ramasubramanian, Kin G. Olivares, Tatiana Konstantinova, Andres Potapczynski, Mengfei Cao, Dmitry Efimov, Michael W. Mahoney, Andrew G. Wilson
TL;DR
This paper introduces SarSim0, a fast and effective SARIMA-based simulation pipeline for zero-shot time-series forecasting, enabling neural models to generalize well across diverse benchmarks without prior training data.
Contribution
The authors develop SarSim0, a novel simulation method that generates diverse, stable, and realistic time series for zero-shot forecasting, outperforming existing statistical and neural baselines.
Findings
SarSim0 enables training on ~1 billion simulated series.
Neural models trained on SarSim0 surpass statistical forecasters.
Models trained on simulations outperform the auto-generated processes.
Abstract
Zero-shot time-series forecasting holds great promise, but is still in its infancy, hindered by limited and biased data corpora, leakage-prone evaluation, and privacy and licensing constraints. Motivated by these challenges, we propose the first practical univariate time series simulation pipeline which is simultaneously fast enough for on-the-fly data generation and enables notable zero-shot forecasting performance on M-Series and GiftEval benchmarks that capture trend/seasonality/intermittency patterns, typical of industrial forecasting applications across a variety of domains. Our simulator, which we call SarSim0 (SARIMA Simulator for Zero-Shot Forecasting), is based off of a seasonal autoregressive integrated moving average (SARIMA) model as its core data source. Due to instability in the autoregressive component, naive SARIMA simulation often leads to unusable paths. Instead, we…
Peer Reviews
Decision·ICLR 2026 Poster
- The work is well-motivated. The authors point out how training with real series can be limited by licensing barriers, data scale, domain etc. And that training with synthetic data offers unique levers that can be controlled.
- There is very limited analysis conducted with the experiments. First, GIFT-Eval allows for easy analysis of performance stratified by domain/frequency/term length/ variate type etc. These are more important to understand the limitations and strengths of the model, more than the aggregate score on GIFT-Eval. - To further improve the evaluation, I would suggest adding one of the foundation model baselines (e.g. Chronos) trained from scratch on KernelSynth/ForecastPFN/SarSim0. Then it would all
* **Core Technical Problem:** The primary strength of this work is its novel and principled solution to the instability problem in autoregressive model simulation. By shifting from sampling coefficients to directly sampling the poles of the transfer function within the unit circle (Section 4.1), the authors guarantee the generation of stable, non-divergent time series. This is a technically sound and elegant idea that makes the powerful but notoriously fragile SARIMA framework a viable engine
In my opinion, the paper's reliance on the SARIMA framework introduces inherent limitations, and the scope of its claims could be tempered by a more nuanced discussion of the benchmarks and the simulator's own complexity. 1. **Inherent Linearity of the SARIMA Core:** The SARIMA model, which forms the backbone of the simulator, is fundamentally a linear process model. While the paper adds complexity via superposition and non-Gaussian noise, it cannot natively generate time series with core non-
The SARIMA-2 approach is an appropriate generating model that describes many practical time series. As noted by the authors, many demand or utility time series benchmarks consist of a slow process (like inflation in a financial time series) modulating a fast process (like seasonal demand). Within this generating model class, the authors make appropriate choices to enforces stability in the AR dynamics by sampling poles inside the unit circle. This likely helps preserve diversity in the generated
**Novelty.** This is not the first paper to train a zero-shot forecast model purely on simulation data and report strong results. ForecastPFN and TabPFN first used this approach, while Mamba4Cast uses an SSM on PFN synthetic data to achieve strong results on the original GluonTS datasets. While the idea itself is not the first of its kind, the arguments in favor of this paper could be (1) the particular choice of synthetic data generation, and (2) the empirical results. I do not currently feel
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsForecasting Techniques and Applications · Stock Market Forecasting Methods · Gaussian Processes and Bayesian Inference