Zero-shot forecasting of optical network telemetry using large language models

Accurate forecasting of optical network telemetry is critical for enabling predictive maintenance and proactive network management. However, conventional forecasting approaches often rely on task-specific architectures, extensive labeled datasets, and manual feature engineering—factors that limit their adaptability and scalability across heterogeneous deployment scenarios. In this work, we introduce a zero-shot forecasting framework that leverages pretrained large language models (LLMs) to perform prompt-based inference directly on telemetry data without requiring fine tuning, retraining, or supervision. Time series signals are tokenized at the digit level and reformulated as structured language sequences, enabling LLMs to perform forecasting, anomaly detection, and missing data imputation using natural language prompts. We evaluate our approach on real-world measurements from an aerial WDM system—comprising PM and SOP data under live traffic and varying environmental conditions—as well as on other out-of-distribution datasets. LLM demonstrates good forecasting accuracy and robustness to noise, missing inputs, and domain shifts, achieving performance that is competitive with classical and deep learning baselines across a range of metrics and conditions. Crucially, the model generalizes across multivariate and multimodal telemetry streams without architectural modification, enabling accurate forecasting from exogenous signals such as environmental variables. These results position instruction-tuned LLMs as lightweight, interpretable, and scalable building blocks for next-generation telemetry analytics—offering a unified, prompt-driven alternative to fragmented and model-centric forecasting pipelines.