OneCast: Structured Decomposition and Modular Generation for Cross-Domain Time Series Forecasting

TL;DR

OneCast addresses cross-domain time series forecasting by explicitly disentangling seasonality and trends. The seasonal component is reconstructed via a lightweight basis-function projection, while the trend component is encoded into discrete tokens through a semantic tokenizer and refined with a diffusion-based predictor. A two-stage training scheme (jointly training the seasonal predictor and tokenizer, then training the diffusion predictor) enables stable cross-domain transfer and long-range forecasting with reduced error accumulation. Across nine real-world datasets, OneCast achieves state-of-the-art performance in many settings, demonstrates strong cross-domain generalization, and showcases substantial token-efficiency and interpretability advantages for practical deployment.

Abstract

Cross-domain time series forecasting is a valuable task in various web applications. Despite its rapid advancement, achieving effective generalization across heterogeneous time series data remains a significant challenge. Existing methods have made progress by extending single-domain models, yet often fall short when facing domain-specific trend shifts and inconsistent periodic patterns. We argue that a key limitation lies in treating temporal series as undifferentiated sequence, without explicitly decoupling their inherent structural components. To address this, we propose OneCast, a structured and modular forecasting framework that decomposes time series into seasonal and trend components, each modeled through tailored generative pathways. Specifically, the seasonal component is captured by a lightweight projection module that reconstructs periodic patterns via interpretable basis functions. In parallel, the trend component is encoded into discrete tokens at segment level via a semantic-aware tokenizer, and subsequently inferred through a masked discrete diffusion mechanism. The outputs from both branches are combined to produce a final forecast that captures seasonal patterns while tracking domain-specific trends. Extensive experiments across eight domains demonstrate that OneCast mostly outperforms state-of-the-art baselines.

Figures (15)

• Figure 1: Time series data from different domains exhibit significant heterogeneity, such as variations in different, numeric ranges, sampling rates and key patterns.
• Figure 2: The inference pipeline of OneCast. Right: detailed prediction process of diffusion-based token predictor.
• Figure 3: The complete training pipeline of OneCast. Left: joint optimization of seasonal predictor and semantic trend tokenizer, where $X_{h}$ and $X_f$ represents the history and future windows, respectively. Right: training process of confidence-aware diffusion-based token predictor.
• Figure 4: The forecasting performance of OneCast between cross-domain and in-domain training, with different prediction length of 48 and 192.
• Figure 5: Comparison of dual- and single-decoder training strategies. "Dual-h" and "Dual-f" are reconstruct errors of historical and future windows under dual-decoder strategy; "Single-h" and "Single-f" denote those under the single one.