chronos-2

Technical Overview

Chronos‑2 (model ID amazon/chronos-2) is an encoder‑only, 120 M‑parameter foundation model for time‑series forecasting. Built on a T5‑style encoder, it treats a forecasting problem as a sequence‑to‑sequence task: the historical target values and any covariates are tokenised and fed to the encoder, and the model autoregressively predicts quantile‑based values for the next prediction_length steps. Unlike earlier Chronos releases, Chronos‑2 can handle univariate, multivariate and covariate‑informed forecasting all within a single architecture, making it a truly “universal” forecaster.

Key capabilities include:

• Zero‑shot forecasting – no task‑specific fine‑tuning is required; the model learns to extrapolate from the context provided at inference time.
• Multi‑step quantile forecasts – produces probabilistic predictions (e.g., 0.1, 0.5, 0.9 quantiles) for up to 1 024 steps ahead.
• Group attention – a specialised attention mask that lets the model learn relationships across multiple series and covariates in a single forward pass, improving efficiency for large panels.
• Long context support – up to 8 192 timesteps of history can be ingested, far exceeding the 512‑step limit of the original Chronos.
• Covariate flexibility – accepts past and future real‑valued or categorical covariates, enabling demand‑driven or calendar‑aware forecasts.

The architecture mirrors the T5 encoder: a stack of self‑attention layers followed by a feed‑forward network, but with two important modifications. First, the attention mask is “grouped” so that each series and its covariates are processed together while still allowing cross‑series interaction. Second, the output head is a quantile regression head that directly predicts the desired quantile values instead of a single point estimate.

Intended use cases span any domain where accurate, fast, and probabilistic time‑series forecasts are needed: retail demand planning, electricity load forecasting, inventory optimisation, financial price prediction, and even sensor‑data anomaly detection. Because the model runs efficiently on both GPUs and CPUs, it can be deployed on edge devices for low‑latency inference or scaled to massive batch jobs on cloud GPUs.

Benchmark Performance

Chronos‑2 is evaluated on three public leaderboards that are widely regarded as the gold standard for zero‑shot time‑series forecasting:

• FEV‑Bench – a comprehensive suite of real‑world series covering retail, finance, and IoT domains.
• GIFT‑Eval – focuses on multi‑horizon, multivariate forecasting with covariates.
• Chronos Benchmark II – the authors’ own benchmark that tests context length, prediction length, and probabilistic accuracy.

Across all three, Chronos‑2 consistently ranks in the top‑1 position for Mean Absolute Scaled Error (MASE) and Quantile‑Weighted Pinball Loss. In the latest GIFT‑Eval leaderboard (as of Dec 2025) it outperforms the previous Chronos‑Bolt by a margin of 7 % in average pinball loss while handling up to 1 024‑step horizons. On the FEV‑Bench, the model delivers > 300 forecasts per second on a single NVIDIA A10G GPU, a speed advantage of roughly 2× over the 120 M‑parameter Chronos‑Bolt baseline.

These benchmarks matter because they stress different aspects of a forecaster: zero‑shot generalisation (no fine‑tuning), long‑range horizon accuracy, and probabilistic calibration. Chronos‑2’s superior scores indicate that it can be trusted to provide reliable forecasts out‑of‑the‑box across a wide variety of domains, a claim that is reinforced by its strong performance on the Chronos Benchmark II “universal” test set.

Hardware Requirements

Chronos‑2’s 120 M‑parameter encoder is lightweight enough to run on modest hardware, yet it benefits from GPU acceleration for high‑throughput scenarios.

• VRAM for inference – roughly 2 GB of GPU memory is sufficient for a single series with a full 8 192‑step context. Batch inference (e.g., 32 series simultaneously) comfortably fits on 8 GB‑class GPUs.
• Recommended GPU – NVIDIA A10G, A100, or RTX 4090 are ideal; they provide the > 300 forecasts‑per‑second throughput reported in the README. Older GPUs (e.g., T4) work but at reduced speed.
• CPU inference – the model can be run on CPUs (Intel Xeon E5‑2690 v4 or newer) with a typical latency of ~ 30 ms per forecast for a single series. This makes it feasible for batch jobs on inexpensive cloud instances.
• Storage – the model files (safetensors + config) total ~ 500 MB. Adding the inference package (chronos‑forecasting) brings the footprint to < 1 GB.
• Performance characteristics – on a single A10G GPU, Chronos‑2 can generate up to 1 024‑step forecasts for 300 series per second, with a quantile‑loss comparable to state‑of‑the‑art specialised models.

Use Cases

Chronos‑2 shines in any scenario that requires fast, accurate, and probabilistic forecasts without the overhead of task‑specific model training.

• Retail & E‑commerce demand forecasting – predict daily sales for thousands of SKUs using historical sales and promotional calendars as covariates.
• Energy load & price prediction – ingest electricity price history and weather forecasts (future covariates) to produce multi‑day price envelopes.
• Supply‑chain inventory optimisation – combine past inventory levels, inbound shipment schedules, and future demand signals for just‑in‑time replenishment.
• Financial time‑series analysis – generate probabilistic forecasts for asset prices or volatility indices, useful for risk‑adjusted portfolio construction.
• IoT sensor monitoring – forecast temperature, pressure, or vibration signals to detect anomalies before they cause failures.

Integration is straightforward via the chronos‑forecasting Python package, the Hugging Face transformers pipeline, or a SageMaker endpoint for production‑grade latency and scaling. The model’s ability to accept future covariates means it can be combined with external regressors (e.g., weather APIs) for truly data‑driven planning.

Training Details

Chronos‑2 was trained on a hybrid dataset that blends large‑scale synthetic series with real‑world collections:

• Synthetic data – generated using stochastic processes (ARIMA, Prophet‑like seasonality, and random walk) to cover a wide variety of dynamics.
• Real‑world data – the autogluon/chronos_datasets and Salesforce/GiftEvalPretrain corpora, which together contain millions of series from retail, finance, and IoT domains.

Training employed a masked‑language‑model style objective where the model learns to reconstruct future target values and quantiles given a context window of up to 8 192 timesteps. The group‑attention mask enables the model to simultaneously learn intra‑series patterns and inter‑series relationships, a key factor behind its zero‑shot capability.

Compute requirements were modest for a foundation model of this size: the authors report training on a cluster of 8 × NVIDIA A100 40 GB GPUs for roughly 48 hours, using mixed‑precision (FP16) to keep memory usage low. The final checkpoint is released as a safetensors file for safe loading.

Fine‑tuning is supported through the chronos‑forecasting library. Users can load the base checkpoint with Chronos2Pipeline.from_pretrained and continue training on a domain‑specific dataset, optionally adding custom covariate embeddings. The library handles the quantile loss and context‑window management automatically.

Related Papers

Chronos‑2 builds on two recent arXiv pre‑prints:

• Chronos Benchmark II – A Comprehensive Evaluation of Time‑Series Foundation Models (2024) – introduces the benchmark suite used to validate Chronos‑2’s zero‑shot performance.
• Chronos‑2: A Universal Encoder‑Only Model for Multivariate Forecasting with Covariates (2025) – describes the model architecture, group‑attention mechanism, and training methodology.

These works provide the theoretical foundation for the encoder‑only design, the use of quantile regression heads, and the extensive synthetic‑plus‑real data mixture that enables strong out‑of‑the‑box generalisation.

Licensing Information

The README lists the model’s license as Apache‑2.0. This permissive licence grants you the right to use, modify, distribute, and commercialise the model and any derivative works, provided you retain the original copyright notice and include a copy of the licence.

Key points for commercial users:

• You may embed Chronos‑2 in SaaS products, on‑premise analytics platforms, or edge devices without paying royalties.
• Any modifications (e.g., fine‑tuned weights) must also be distributed under Apache‑2.0 if you choose to share them publicly.
• Attribution is required: a clear statement such as “Chronos‑2 © Amazon, licensed under Apache‑2.0” must accompany the model.
• No trademark or endorsement rights are granted; you cannot imply that Amazon endorses your commercial offering unless you obtain separate permission.

If you encounter a discrepancy between the “unknown” license tag on the Hugging Face hub and the Apache‑2.0 notice in the README, the latter is the authoritative source. Always keep a copy of the licence file (found in the repository) in your distribution package.