Sketch2Series: Agentic Planning for Time-Series Generation without Paired Text Series Data

Text-to-time-series generation can support simulation, diagnosis, and data augmentation in domains where paired textual descriptions and signals are scarce. We introduce Sketch2Series, a hybrid framework that decouples semantic planning from stochastic signal generation. A validated large language model agent translates a natural-language prompt into an executable structural sketch encoding trends, seasonality, and event locations. A conditional Diffusion Transformer then renders this sketch as a realistic time series, while a dataset-aware augmentation strategy bridges the gap between algorithmically extracted training sketches and agent-generated sketches at inference. Consequently, Sketch2Series functions in a zero-shot text regime; it bypasses the need for paired text-to-time-series training data and leverages only domain knowledge specific to time series. Experiments across structured and seasonal datasets show that the method combines accurate prompt alignment with realistic signal morphology, outperforming supervised and zero-shot baselines on strongly constrained generation tasks while remaining competitive on less constrained real-world data.