TLDR: OneCast is a new framework for cross-domain time series forecasting that tackles data heterogeneity by explicitly decomposing time series into seasonal and trend components. It uses a lightweight projection module for seasonal patterns and a semantic-aware tokenizer with a dual-decoder strategy for trend components, which are then predicted using a discrete diffusion mechanism. This modular approach enhances generalization, interpretability, and efficiency across diverse datasets, outperforming existing methods.
Time series forecasting, a critical technique in many web applications like traffic scheduling and e-commerce, has seen rapid advancements. However, a significant hurdle remains: effectively predicting across diverse and heterogeneous time series data, known as cross-domain forecasting. Traditional methods often struggle with domain-specific trend shifts and inconsistent periodic patterns because they treat time series as undifferentiated sequences, failing to explicitly separate their inherent structural components.
A new framework called OneCast addresses this challenge by introducing a structured and modular approach to time series forecasting. It operates on the principle that by explicitly decomposing time series into their fundamental seasonal and trend components, and then modeling each through specialized generative pathways, better generalization across different data domains can be achieved.
Deconstructing Time Series: Seasonal and Trend Components
OneCast’s core innovation lies in its decomposition strategy. It recognizes that seasonal patterns, such as daily or weekly cycles, often exhibit transferable regularities across domains. For these seasonal components, OneCast employs a lightweight projection module. This module reconstructs periodic patterns using interpretable basis functions, essentially mapping inputs onto predefined periodic functions (like sine and cosine waves) to capture recurring behaviors efficiently and clearly.
In parallel, the trend component, which often shows significant disparity and domain-specific dynamics (e.g., steady growth, decline, or irregular fluctuations), is handled differently. OneCast encodes these segment-level trend patterns into discrete tokens using a semantic-aware tokenizer. This tokenizer is trained with an innovative dual-decoder strategy, which helps it create pattern-level representations that are less sensitive to variations in peak values and data distribution shifts across domains.
Predicting the Future: Diffusion and Dual-Decoder Strategies
Once the trend components are tokenized, their future evolution is inferred through a masked discrete diffusion mechanism. This diffusion-based token predictor iteratively generates future tokens, capturing long-range dependencies and enabling the decoding of high-confidence tokens. Unlike traditional autoregressive models that can suffer from accumulating errors, this approach refines multiple tokens simultaneously, leading to more stable and accurate long-term predictions.
The training of OneCast is a two-stage process. First, the seasonal predictor and the trend tokenizer are jointly optimized. This stage focuses on learning stable token representations and ensuring the tokenizer can effectively compress sequence information while also aiding in the reconstruction of future trends, even when future statistics are unknown. The dual-decoder strategy is crucial here, allowing the tokenizer to leverage historical statistics for more precise future token decoding. In the second stage, the diffusion-based token predictor is independently trained on these stabilized representations to learn how to iteratively restore masked tokens.
Also Read:
- EMTSF: A New Hybrid AI Model Excels in Time Series Forecasting by Combining Diverse Expert Systems
- Xihe: A New Era for Zero-Shot Time Series Forecasting with Hierarchical Attention
Performance and Efficiency Across Diverse Data
Extensive experiments conducted across nine diverse real-world datasets, including traffic, electricity, weather, and economic data, demonstrate OneCast’s strong competitive performance. It often outperforms state-of-the-art baselines in cross-domain forecasting tasks. The framework shows particular strengths in handling data heterogeneity, mitigating domain confusion by creating distinct yet uniformly scaled codebook vector distributions, and achieving low information loss during its discretization process.
Furthermore, OneCast exhibits remarkable efficiency in token consumption during series encoding. By using a cross-channel encoding scheme, its token count remains independent of the number of features or channels in the data. This is a significant advantage in high-dimensional scenarios, drastically reducing resource consumption compared to other tokenization methods and supporting efficient deployment in practical applications.
In essence, OneCast offers a fresh perspective on cross-domain time series forecasting by combining structured decomposition with modular generation. Its ability to disentangle and model seasonal and trend components separately, coupled with its innovative diffusion-based prediction and efficient tokenization, makes it a robust and interpretable solution for predicting across varied and complex time series data. You can find more details about this research in the full paper available at arXiv:2510.24028.
