When does zero-shot forecasting with computed periodicity actually work?
It works when the time series has clear, repeating patterns and little noise. A 2025 study found that LLMs (large language models) perform well on series with obvious patterns and trends, and that they can recognize underlying periods in the data [3]. Another 2024 study showed that a framework explicitly extracting periodic features in the frequency domain achieved superior performance in zero-shot scenarios, even beating specialized models [2]. So if your data is clean and periodic, computed periodicity can be a powerful guide.
What breaks it? Noise, trends, and complex frequencies
The biggest weakness is sensitivity to noise. A 2025 study found that LLM-based zero-shot forecasters often struggle to achieve high accuracy because they are highly sensitive to noise, underperforming even simple domain-specific models [5]. Similarly, a 2024 evaluation showed that predictive accuracy diminishes notably when the signal contains both periodic and trend components, or when it has complex frequency components [1]. This means that if your data has a trend on top of the periodicity, or multiple overlapping cycles, the computed periodicity alone is not enough to keep forecasts accurate.
What does this mean for real-world forecasting tasks?
In practice, you should not rely on zero-shot forecasting with computed periodicity for volatile or non-stationary data. A 2025 study on financial time series found that zero-shot forecasting is critical during abrupt market regime shifts, but that standard meta-learning approaches perform suboptimally in such turbulent conditions [4]. The same study showed that explicitly designing tasks to capture both local patterns and cross-series invariants improved generalization in zero-shot scenarios [4]. So, while the approach is fragile, it can be made more robust with careful design—but not by simply computing periodicity and hoping for the best.
About These Sources
This answer is built on 5 studies (3 peer-reviewed, 2 preprints) — published from 2024 to 2025, 5 from 2024 or later — selected as the most relevant from 6 studies that passed quality screening, drawn from 42 papers retrieved from a database of over 500 million.
Sources used in this answer
An Evaluation of Standard Statistical Models and LLMs on Time Series Forecasting
In an evaluation of the LLMTIME model, predictive accuracy diminished notably on time series with both periodic and trend components, and on signals with complex frequency components.
LLM-TPF: Multiscale Temporal Periodicity-Semantic Fusion LLMs for Time Series Forecasting
The LLM-TPF framework, which extracts periodic features in the frequency domain, achieved superior performance compared to state-of-the-art specialized models and exhibited exceptional generalization in zero-shot scenarios.
Time Series Forecasting with LLMs: Understanding and Enhancing Model Capabilities
LLMs perform well on time series with clear patterns and trends but face challenges with datasets lacking periodicity, and incorporating external knowledge improves predictive performance.
Adapting to the Unknown: Robust Meta-Learning for Zero-Shot Financial Time Series Forecasting
In financial time series from high-volatility periods and multiple international markets, a meta-learning approach using GMM-based task construction outperformed existing methods and showed stronger generalization in zero-shot scenarios.
Revisiting LLMs as Zero-Shot Time Series Forecasters: Small Noise Can Break Large Models
LLM-based zero-shot forecasters often struggle to achieve high accuracy due to sensitivity to noise, underperforming even simple domain-specific models, and improving robustness remains a significant challenge.