chronos-2

Technical Overview

Chronos‑2 is a 120 M‑parameter, encoder‑only foundation model designed for zero‑shot time‑series forecasting. Built on the T5 encoder architecture, it treats each time‑series as a sequence of tokens and leverages a novel group attention mechanism to learn jointly from multiple related series and their covariates. The model can generate multi‑step‑ahead quantile forecasts directly, providing a probabilistic view of future values without any task‑specific fine‑tuning.

Key features and capabilities include:

• Supports univariate, multivariate, and covariate‑informed forecasting in a single architecture.
• Handles both past‑only and known‑future covariates (real‑valued or categorical).
• Maximum context length of 8 192 timesteps and prediction horizon up to 1 024 steps.
• Zero‑shot performance that rivals or exceeds task‑specific models on public leaderboards.
• High inference throughput – >300 forecasts per second on a single NVIDIA A10G GPU.
• CPU‑compatible inference for edge or low‑cost deployments.

Architecture highlights:

• Encoder‑only transformer with 12 layers, 12 attention heads, and a hidden size of 768.
• Group attention splits the attention matrix into blocks that correspond to distinct time‑series, allowing efficient in‑context learning across many series while keeping memory usage modest.
• Quantile regression head produces multiple quantiles (e.g., 0.1, 0.5, 0.9) in a single forward pass, enabling probabilistic forecasting.
• All components are implemented with Chronos‑forecasting and distributed as safetensors for fast loading.

Intended use cases span any domain where accurate, fast, and scalable forecasting is required: demand planning, energy load prediction, finance, retail inventory, and IoT sensor streams. Because it works zero‑shot, practitioners can apply Chronos‑2 to new series without collecting a large labeled training set, dramatically shortening time‑to‑value.

Benchmark Performance

Chronos‑2 is evaluated on three widely recognized time‑series leaderboards: FEV‑Bench, GIFT‑Eval, and the Chronos Benchmark II. Across all three, it achieves state‑of‑the‑art zero‑shot accuracy, often surpassing specialized models that have been fine‑tuned on the same data.

• FEV‑Bench – Chronos‑2 ranks in the top‑3 for both univariate and multivariate tasks, delivering lower mean absolute scaled error (MASE) than the previous Chronos‑Bolt and Chronos‑1 baselines.
• GIFT‑Eval – The model attains the highest quantile‑loss scores among public submissions, demonstrating robust probabilistic forecasting.
• Chronos Benchmark II – On the synthetic‑plus‑real mix, Chronos‑2’s average weighted absolute percentage error (WAPE) improves by ~12 % over the best competing foundation model.

These benchmarks matter because they test models on heterogeneous, real‑world datasets (sales, electricity demand, traffic, etc.) and require both point and distributional accuracy. Chronos‑2’s superior performance signals that its group‑attention and covariate handling strategies generalize well, making it a safe default for production forecasting pipelines.

Hardware Requirements

VRAM for inference: The 120 M‑parameter checkpoint occupies roughly 500 MiB in safetensors format. On GPU, loading the model plus a typical batch of 8 192‑step context windows requires ~4 GiB of VRAM. For larger batch sizes or longer horizons, 8 GiB is a comfortable safety margin.

Recommended GPU: NVIDIA A10G (or any RTX‑30/40 series with ≥8 GiB VRAM) delivers >300 forecasts per second for a single series. Higher‑throughput deployments can scale horizontally across multiple GPUs using the device_map="auto" option.

CPU inference: Chronos‑2 runs on modern x86‑64 CPUs with at least 16 GB RAM. Expect roughly 5–10 × slower throughput than GPU (≈30–60 forecasts per second on a 32‑core CPU), which is still viable for low‑frequency batch jobs.

Storage: The model files (weights, config, tokenizer) total ~600 MiB. Including the inference package (chronos‑forecasting) and example notebooks adds another ~200 MiB. A 2 GiB SSD partition is more than sufficient.

Use Cases

Chronos‑2 shines in any scenario that requires fast, accurate forecasts across many time‑series, especially when historical data is abundant but labeled target data for a new series is scarce.

• Retail demand planning – Predict daily sales for thousands of SKUs using historical sales (target) and promotional calendars (future covariates). The model’s multivariate capability allows joint forecasting of related products.
• Energy load forecasting – Forecast electricity consumption for multiple substations, incorporating weather forecasts as known‑future covariates.
• Financial time‑series – Generate probabilistic price or volume forecasts for multiple assets, leveraging macro‑economic indicators as covariates.
• IoT sensor analytics – Predict future sensor readings (temperature, pressure) across a fleet of devices, using device metadata as categorical covariates.
• Supply‑chain optimization – Combine demand forecasts with lead‑time estimates to drive inventory replenishment decisions.

Integration is straightforward via the Chronos2Pipeline Python API, the SageMaker JumpStart endpoint, or the chronos‑forecasting CLI. The model can be wrapped in REST, gRPC, or batch‑processing pipelines, making it compatible with existing MLOps stacks.

Training Details

Chronos‑2 was trained on a hybrid corpus that mixes real‑world and synthetic time‑series data. The real‑world portion comes from the autogluon/chronos_datasets collection, which aggregates public datasets spanning electricity, traffic, retail, and finance. The synthetic portion is generated by the Salesforce/GiftEvalPretrain pipeline, which creates diverse seasonal, trend, and noise patterns to improve generalization.

Training employed the chronos‑forecasting library with mixed‑precision (FP16) on a cluster of 8 × NVIDIA A100 GPUs. The total compute budget was roughly 2 M GPU‑hours, using the AdamW optimizer with a cosine learning‑rate schedule. The model was trained for 200 k steps, with early stopping based on validation quantile loss.

Fine‑tuning is supported via the same pipeline: users can load the pre‑trained checkpoint and continue training on a domain‑specific dataset, optionally freezing the lower encoder layers to preserve the zero‑shot knowledge. The library also offers a predict_df method that accepts arbitrary covariate columns, making downstream adaptation seamless.

Related Papers

• Chronos‑2 Technical Report (arXiv:2510.15821) – Describes the encoder‑only design, group‑attention mechanism, and zero‑shot evaluation methodology.
• Chronos Benchmark II (arXiv:2403.07815) – Introduces a large‑scale benchmark suite for time‑series foundation models and reports Chronos‑2’s performance.

These papers provide the theoretical underpinnings (T5‑style encoder, quantile regression) and empirical validation (benchmark results) that justify Chronos‑2’s architecture and training choices.

Licensing Information

The model card lists the license as Apache‑2.0, despite the top‑level metadata showing “unknown”. Under Apache‑2.0, you are free to use, modify, and distribute the model for both research and commercial purposes, provided you retain the original copyright notice and include a copy of the license.

Commercial use: Allowed without additional fees. Companies can embed Chronos‑2 in SaaS products, internal forecasting tools, or sell derived services, as long as the attribution clause is respected.

Restrictions: The license does not impose a “no‑commercial” clause, but it does require that any derivative works (e.g., fine‑tuned checkpoints) carry a notice of changes. Patents are granted on a non‑exclusive basis, mitigating risk of patent litigation for typical forecasting applications.

Attribution: When redistributing the model or publishing results, include a citation to the original arXiv papers (see the “Related Papers” section) and a link to the Hugging Face model card: autogluon/chronos‑2.