TLDR: Chronos-2 is a new pretrained time series model that excels at various forecasting tasks—univariate, multivariate, and those with external factors (covariates)—without needing specific training for each. Its core innovation is a “group attention” mechanism that allows it to learn from context and share information across related time series, leading to state-of-the-art performance on major benchmarks, especially for complex, real-world scenarios involving multiple variables and influencing factors. The model’s training on synthetic data and its in-context learning capabilities make it a versatile tool for diverse forecasting challenges.
Time series forecasting, the art of predicting future values based on historical data, has seen a significant shift with the rise of pretrained models, often called foundation models. These models simplify the forecasting process by allowing a single model to be trained once on vast amounts of data and then applied to various forecasting problems without needing to be retrained for each specific task.
However, a persistent challenge has been the limitation of most existing pretrained models to univariate data, meaning they only consider the past observations of a single time series. Real-world scenarios are far more complex, often requiring predictions for multiple co-evolving time series (multivariate forecasting) or forecasts that depend on external factors (covariate-informed forecasting). Imagine predicting cloud infrastructure metrics like CPU usage and memory consumption, which evolve together, or retail demand influenced by promotions and weather. The absence of these capabilities has hindered the widespread adoption of pretrained models in practical applications.
Introducing Chronos-2: A Universal Forecasting Model
A new research paper, “Technical Report of Chronos-2: Chronos-2: From Univariate to Universal Forecasting”, introduces Chronos-2, a groundbreaking pretrained model designed to overcome these limitations. Developed by a team of researchers including Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, and many others from Amazon Web Services and various universities, Chronos-2 is capable of handling univariate, multivariate, and covariate-informed forecasting tasks in a zero-shot manner. This means it can make accurate predictions for previously unseen data without any task-specific training.
The core innovation behind Chronos-2’s capabilities is its unique “group attention” mechanism. This mechanism facilitates in-context learning (ICL) by efficiently sharing information across multiple time series within a defined group. A group can be flexible, representing related series, different variables of a multivariate series, or even targets and their associated external factors (covariates). This allows the model to infer complex interactions from the available context, a crucial feature given the diverse nature of real-world forecasting problems.
How Chronos-2 Achieves Universal Capabilities
Chronos-2’s general capabilities are achieved through an innovative training approach. Instead of relying solely on scarce high-quality multivariate data, the model is trained on synthetic datasets. These synthetic datasets are generated by imposing diverse multivariate structures on univariate series, effectively teaching the model to understand and predict complex relationships between different data streams.
The model’s pipeline involves several steps: input time series (targets and covariates) are first normalized, then augmented with time index and mask meta-features. These sequences are then broken into non-overlapping “patches” and converted into high-dimensional embeddings. A transformer stack, featuring alternating time and group attention layers, processes these embeddings. The time attention layer focuses on temporal patterns within a single series, while the group attention layer aggregates information across all series within a group at each patch index. Finally, a quantile head produces multi-step quantile forecasts, offering a richer representation of prediction uncertainty.
Performance and Practical Advantages
Chronos-2 has demonstrated state-of-the-art performance across three major forecasting benchmarks: fev-bench, GIFT-Eval, and Chronos Benchmark II. On fev-bench, which specifically emphasizes multivariate and covariate-informed forecasting, Chronos-2’s universal ICL capabilities lead to substantial improvements over existing models. It consistently outperforms baselines by a wide margin on tasks involving covariates.
Case studies in critical domains like energy and retail further highlight its practical advantages. For instance, in energy price forecasting, Chronos-2 effectively uses covariates like load and renewable energy generation forecasts to produce significantly more accurate predictions. Similarly, for retail sales forecasting, it leverages promotion and holiday information to capture true sales dynamics, leading to much more precise forecasts compared to models that ignore these external factors.
The research also includes ablation studies, showing that even a smaller version of Chronos-2 (28M parameters) delivers strong performance, making it suitable for resource-constrained environments. Furthermore, models trained exclusively on synthetic data perform only slightly below those trained with real data, underscoring the importance and potential of synthetic data in advancing pretrained time series models. The benefits of long-context post-training are also highlighted, particularly for high-frequency datasets with long seasonal patterns.
Also Read:
- OmniVinci: A Unified AI Model for Vision, Audio, and Language Understanding
- Evaluating Language Models on Real-World Uncertainty with OPENESTIMATE
Looking Ahead
Chronos-2 establishes itself as a general-purpose forecasting model that can be used “as is” in real-world forecasting pipelines. Its flexible group attention mechanism opens doors for future research, such as retrieval-augmented forecasting, where time series could be grouped using sparse metadata or dense embeddings to improve performance in data-scarce scenarios. The development of Chronos-2 marks a significant step towards truly universal time series forecasting, capable of addressing the complexities of real-world data with unprecedented accuracy and flexibility.
