[Paper Review] Zero-shot Forecasting by Simulation Alone

Boris N. Oreshkin, Mayank Jauhari+11 more|arXiv (Cornell University)|Jan 2, 2026

Forecasting Techniques and Applications0 citations

TL;DR

The paper introduces SarSim0, a fast SARIMA-based time series simulator that enables zero-shot forecasting by pretraining neural forecasters on purely synthetic data, achieving strong generalization on M-Series and GiftEval benchmarks. It shows that synthetic data can rival real-data pretraining and even outperform certain baselines in zero-shot settings.

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 follow a three-step procedure: (1) we sample well-behaved trajectories from its characteristic polynomial stability region; (2) we introduce a superposition scheme that combines multiple paths into rich multi-seasonality traces; and (3) we add rate-based heavy-tailed noise models to capture burstiness and intermittency alongside seasonalities and trends. SarSim0 is orders of magnitude faster than kernel-based generators, and it enables training on circa 1B unique purely simulated series, generated on the fly; after which well-established neural network backbones exhibit strong zero-shot generalization, surpassing strong statistical forecasters and recent foundation baselines, while operating under strict zero-shot protocol. Notably, on GiftEval we observe a "student-beats-teacher" effect: models trained on our simulations exceed the forecasting accuracy of the AutoARIMA generating processes.

Motivation & Objective

• Motivate zero-shot forecasting in industrial settings where real data is scarce, biased, or leakage-prone.
• Develop a fast, leakage-free synthetic data generator to pretrain forecasting models at scale.
• Ground the simulator in stable SARIMA dynamics with extensions for multi-seasonality and heavy-tailed noise.
• Demonstrate that models pretrained on synthetic data generalize to heterogeneous benchmarks without target data fine-tuning.

Proposed method

• Model base: use SARIMA as the core data-generating process for synthetic series.
• Stabilize simulation by sampling poles within the unit circle and deriving coefficients from pole representations.
• Introduce SARIMA-2 to capture bi-seasonality via base and envelope processes with additive or multiplicative interaction.
• Attach a Noiser module to inject heavy-tailed, level-dependent disturbances (Poisson, generalized Gamma, and lognormal).
• Vectorize generation across multiple trajectories to enable on-the-fly synthesis of billions of series.
• Train foundation-model backbones (e.g., NBEATS, PatchTST, Chronos-Small T5) exclusively on SarSim0-generated data and evaluate zero-shot performance on benchmarks.

Figure 1: SarSim0 simulator pipeline. Top: two base components are generated by SARIMA with AR (and seasonal) roots sampled via the characteristic polynomial inside the stability region, yielding well-behaved paths at seasonalities $s\!=\!24$ and $s\!=\!7$ . Middle: a SARIMA-2 superposition/modulati

Experimental results

Research questions

• RQ1 Can a SARIMA-based synthetic data generator produce realistic time-series patterns suitable for training forecasters?
• RQ2 Does pretraining on purely simulated data enable strong zero-shot generalization across diverse real-world benchmarks?
• RQ3 How do different architectural inductive biases (e.g., NBEATS, PatchTST, Chronos) perform when pretrained on synthetic data?
• RQ4 What is the contribution of each simulator component (SARIMA, SARIMA-2, Noisers) to zero-shot forecasting performance?

Key findings

• SarSim0-trained models achieve strong zero-shot generalization across heterogeneous benchmarks, outperforming some real-data pretraining baselines.
• Models pretrained on SarSim0 often close the gap with large real-data pretrained models and can even outperform certain synthetic baselines like KernelSynth and ForecastPFN.
• Architectures with diverse inductive biases (dense, attention-based, patching) trained on the same synthetic data achieve competitive performance, indicating robustness to model choice.
• On GiftEval, models pretrained on SarSim0 exhibit a student-beats-teacher effect, outperforming AutoARIMA generated processes.
• Ablation studies show SARIMA-2 and the Noisers contribute meaningfully to generalization, with SARIMA-2 being particularly important for accuracy across backbones.

Figure 2: Sampling of SARIMA poles by SarSim0 . The SARIMA order-10 AR process poles are shown along with the unit circle on the left. The resulting generated processes with these poles are shown on the right. The top pane shows poles sampled according to the proposed procedure, resulting in a reali